The Six Sigma Handbook
Revised and Expanded
A Complete Guide
for Green Belts, Black Belts,
and Managers at All Levels
Part I Six Sigma Implementation and Management 1
Chapter 1 Building the Six Sigma Infrastructure 3
What is Six Sigma? 3
Why Six Sigma? 4
The Six Sigma philosophy 6
The change imperative 11
Change agents and their effects on organizations 13
Implementing Six Sigma 20
Timetable 22
Infrastructure 25
Six Sigma deployment and management 31
Six Sigma communication plan 31
Six Sigma organizational roles and responsibilities 35
Selecting the ‘‘Belts’’ 38
Integrating Six Sigma and related initiatives 49
Deployment to the supply chain 51
Change agent compensation and retention 54
Chapter 2 Six Sigma Goals and Metrics 56
Attributes of good metrics 56
Six Sigma versus traditional three sigma performance 58
The balanced scorecard 61
Measuring causes and effects 62
Information systems 64
For more information about this title, click here.
Copyright 2003 by The McGraw-Hill Companies, Inc. Click Here for Terms of Use.
Customer perspective 65
Internal process perspective 67
Innovation and learning perspective 69
Financial perspective 70
Strategy deployment plan 71
Information systems requirements 74
Integrating Six Sigma with other information systems
technologies 74
OLAP, data mining, and Six Sigma 79
Dashboard design 79
Dashboards for scale data 81
Dashboards for ordinal data 84
Dashboards for nominal data 87
Setting organizational key requirements 89
Benchmarking 91
Chapter 3 Creating Customer-Driven Organizations 97
Elements of customer-driven organizations 97
Becoming a customer- and market-driven enterprise 98
Elements of the transformed organization 98
Surveys and focus groups 102
Strategies for communicating with customers and employees 102
Surveys 103
Focus groups 113
Other customer information systems 114
Calculating the value of retention of customers 116
Complaint handling 118
Kano model of customer expectations 119
Customer expectations, priorities, needs, and ‘‘voice’’ 119
Garden variety Six Sigma only addresses half of the Kano
customer satisfaction model 120
Quality function deployment (QFD) 121
Data collection and review of customer expectations, needs,
requirements, and specifications 123
The Six Sigma process enterprise 125
Examples of processes 126
The source of conflict 128
A resolution to the conflict 129
Process excellence 130
Using QFD to link Six Sigma projects to strategies 132
The strategy deployment matrix 133
Deploying differentiators to operations 136
iv Contents
Deploying operations plans to projects 138
Linking customer demands to budgets 140
Structured decision-making 140
Category importance weights 145
Subcategory importance weights 146
Global importance weights 147
Chapter 4 Training for Six Sigma 150
Training needs analysis 150
The strategic training plan 152
Training needs of various groups 153
Post-training evaluation and reinforcement 162
Chapter 5 Six Sigma Teams 167
Six Sigma teams 167
Process improvement teams 168
Work groups 169
Quality circles 169
Other self-managed teams 170
Team dynamics management, including con?ict resolution 171
Stages in group development 172
Common problems 173
Member roles and responsibilities 173
Facilitation techniques 178
When to use an outside facilitator 178
Selecting a facilitator 178
Principles of team leadership and facilitation 179
Facilitating the group task process 181
Facilitating the group maintenance process 182
Team performance evaluation 182
Team recognition and reward 184
Chapter 6 Selecting and Tracking Six Sigma Projects 187
Choosing the right projects 188
Customer value projects 188
Shareholder value projects 189
Other Six Sigma projects 189
Analyzing project candidates 189
Benefit-cost analysis 189
A system for assessing Six Sigma projects 190
Other methods of identifying promising projects 198
Throughput-based project selection 201
Multi-tasking and project scheduling 205
Summary and preliminary project selection 208
Contents v
Tracking Six Sigma project results 208
Financial results validation 211
Financial analysis 212
Lessons learned capture and replication 233
Part II Six Sigma Tools and Techniques 235
Chapter 7 Introduction to DMAIC and Other Improvement
Models 237
DMAIC, DMADV and learning models 237
Design for Six Sigma project framework 239
Learning models 241
PDCA 243
Dynamic models of learning and adaptation 245
The Define Phase
Chapter 8 Problem Solving Tools 252
Process mapping 252
Cycle time reduction through cross-functional process
mapping 253
Flow charts 254
Check sheets 255
Process check sheets 256
Defect check sheets 257
Stratified defect check sheets 257
Defect location check sheets 258
Cause and effect diagram check sheets 259
Pareto analysis 259
How to perform a Pareto analysis 259
Example of Pareto analysis 260
Cause and e?ect diagrams 261
7M tools 264
Affinity diagrams 264
Tree diagrams 265
Process decision program charts 265
Matrix diagrams 268
Interrelationship digraphs 268
Prioritization matrices 269
Activity network diagram 273
Other continuous improvement tools 273
vi Contents
The Measure Phase
Chapter 9 Basic Principles of Measurement 277
Scales of measurement 277
Reliability and validity of data 280
Definitions 280
Overview of statistical methods 283
Enumerative versus analytic statistical methods 283
Enumerative statistical methods 287
Assumptions and robustness of tests 290
Distributions 291
Probability distributions for Six Sigma 293
Statistical inference 310
Hypothesis testing/Type I and Type II errors 315
Principles of statistical process control 318
Terms and concepts 318
Objectives and benefits 319
Common and special causes of variation 321
Chapter 10 Measurement Systems Analysis 325
R&R studies for continuous data 325
Discrimination, stability, bias, repeatability,
reproducibility, and linearity 325
Gage R&R analysis using Minitab 337
Output 338
Linearity 341
Attribute measurement error analysis 346
Operational definitions 348
Example of attribute inspection error analysis 350
Respectability and pairwise reproducibility 352
Minitab attribute gage R&R example 356
The Analyze Phase
Chapter 11 Knowledge Discovery 361
Knowledge discovery tools 361
Run charts 361
Descriptive statistics 368
Histograms 371
Exploratory data analysis 381
Establishing the process baseline 385
Describing the process baseline 387
Contents vii
Process for creating a SIPOC diagram 389
SIPOC example 390
Chapter 12 Statistical Process Control Techniques 393
Statistical process control (SPC) 393
Types of control charts 393
average and range, average and sigma, control charts for
individual measurements, control charts for proportion
defective, control chart for count of defectives, control
charts for average occurrences-per-unit, control charts
for counts of occurrences-per unit
Short-run SPC 430
control chart selection, rational subgroup sampling,
control charts interpretation
EWMA 453
EWMA charts 453
SPC and automatic process control 465
Minitab example of EWMA
Chapter 13 Process Capability Analysis 467
Process capability analysis (PCA) 467
How to perform a process capability study 467
Statistical analysis of process capability data 471
Process capability indexes 472
Interpreting capability indexes 473
Example of capability analysis using normally distributed
variables data 475
Estimating process yield 484
Rolled throughput yield and sigma level 484
Normalized yield and sigma level 487
Chapter 14 Statistical Analysis of Cause and E?ect 490
Testing common assumptions 490
Continuous versus discrete data 490
Independence assumption 492
Normality assumption 493
Equal variance assumption 496
Regression and correlation analysis 496
Scatter plots 496
Correlation and regression 502
Analysis of categorical data 514
Chi-square, tables 514
viii Contents
Logistic regression 516
binary logistic regression, ordinal logistic regression,
and nominal logistic regression
Non-parametric methods 528
Guidelines on when to use non-parametric tests 533
Minitab’s nonparametric tests
The Improve Phase
Chapter 15 Managing Six Sigma Projects 534
Useful project management tools and techniques 535
Project planning 536
Project charter 538
Work breakdown structures 541
Feedback loops 543
Performance measures 544
Gantt charts 544
Typical DMAIC project tasks and responsibilities 545
PERT-CPM-type project management systems 545
Resources 552
Resource conflicts 552
Cost considerations in project scheduling 552
Relevant stakeholders 556
Budgeting 558
Project management implementation 560
Management support and organizational roadblocks 560
Short-term (tactical) plans 565
Cross-functional collaboration 566
Continuous review and enhancement of quality process 567
Documentation and procedures 568
Chapter 16 Risk Assessment 571
Reliability and safety analysis 571
Reliability analysis 571
Risk assessment tools 590
Fault free analysis 591
Safety analysis 591
Failure mode and e?ect analysis (FMEA) 596
FMEA process 597
Statistical tolerancing 600
Assumptions of formula 605
Tolerance intervals 606
Contents ix
x Contents
Chapter 17 Design of Experiments (DOE) 607
Terminology 608
Definitions 608
Power and sample size 610
Example 610
Design characteristics 610
Types of design 611
One-factor 614
Examples of applying common DOE methods using software 616
Two-way ANOVA with no replicates 617
Two-way ANOVA with replicates 618
Full and fractional factorial 621
Empirical model building and sequential learning 624
Phase 0: Getting your bearings 626
Phase I: The screening experiment 627
Phase II: Steepest ascent (descent) 631
Phase III: The factorial experiment 633
Phase IV: The composite design 636
Phase V: Robust product and process design 640
Data mining, arti?cial neural networks and virtual process
mapping 644
Example 646
The Control Phase
Chapter 18 Maintaining Control After the Project 649
Business process control planning 649
How will we maintain the gains made? 649
Tools and techniques useful for control planning 651
Using SPC for ongoing control 652
Process control planning for short and small runs 655
Strategies for short and small runs 655
Preparing the short run process control plain (PCP) 656
Process audit 658
Selecting process control elements 658
The single part process 660
Other elements of the process control plan 661
PRE-Control 661
Setting up PRE-Control 662
Using PRE-Control 663
Contents xi
Beyond DMAIC
Chapter 19 Design for Six Sigma (DFSS) 665
Preliminary steps 665
De?ne 667
Identify CTQs 667
Beyond customer requirementsLidentifying ‘‘delighters’’ 667
UsingAHPto determine the relative importance of theCTQs 668
Measure 670
Measurement plan 671
Analyze 671
Using customer demands to make design decisions 674
Using weighted CTQs in decision-making 678
Pugh concept selection method 681
Design 682
Predicting CTQ performance 682
Process simulation 685
Virtual DOE using simulation software 699
Design phase cross-references 703
Verify 703
Pilot run 704
Transition to full-scale operations 704
Verify phase cross-references 704
Chapter 20 Lean Manufacturing and Six Sigma 705
Introduction to Lean and muda 705
What is value to the customer? 706
Example: Weld dents 706
The value definition 707
Kinds of waste 708
What is the value stream? 708
Value stream mapping 710
How do we make value ?ow? 711
Example of Takt time calculation 712
Spaghetti charts 712
How do we make value ?ow at the pull of the customer? 713
Tools to help improve flow 714
5S; constraint management; level loading; pull
systems; flexible process; lot size reduction
How can we continue towards perfection? 716
Becoming Lean: A tactical perspective 720
Six Sigma and Lean 721
Appendix 724
Table 19Glossary of basic statistical terms 724
Table 29Area under the standard normal curve 730
Table 39Critical values of the t-distribution 733
Table 49Chi-square distribution 735
Table 59F distribution (a ? 1%? 738
Table 69F distribution (a ? 5%) 740
Table 79Poisson probability sums 742
Table 89Tolerance interval factors 746
Table 99Durbin-Watson test bounds 750
Table 109y factors for computing AOQL 754
Table 119Control chart constants 755
Table 129Control chart equations 757
Table 139Table of d
2 values 759
Table 149Power functions for ANOVA 761
Table 159Factors for short run control charts for
individuals, X-bar, and R charts 770
Table 169Signi?cant number of consecutive highest or
lowest values from one stream of a multiple-stream
process 772
Table 179Sample customer survey 773
Table 189Process s levels and equivalent PPM quality levels 777
Table 199Black Belt e?ectiveness certi?cation 778
Table 209Green Belt e?ectiveness certi?cation 791
Table 219AHP using Microsoft ExcelTM 804
References 806
Index 814
xii Contents

First, a basic question: just what are organizations anyway? Why do they
exist? Some experts believe that the reason organizations exist is because of the
high cost of executing transactions in the marketplace. Within an organization
we can reallocate resources without the need to negotiate contracts, formally
transfer ownership of assets, and so on. No need for lawyers, the managers do
things on their own authority. The question is: how should they do this? In the
free market prices tell us how to allocate resources, but prices don’t exist inside
of an organization. We must come up with some alternative.
Transaction costs aside, organizations exist to serve constituencies.
Businesses have shareholders or private owners. The equivalent for non-profits
are contributors. Organizations also serve ‘‘customer’’ constituencies. In other
words, they produce things that other people want. Businesses must produce
things that people are willing and able to buy for their own benefit. Non-profits
must produce things that contributors are willing and able to buy for the benefit
of others. Both types of organizations must do one thing: create value. The output
must be of greater value than the inputs needed to produce it. If the output
serves the constituencies well, the organization is effective. If it creates added
value with a minimum of resources, it is efficient. (It is a common misconception
that non-profits don’t need to be efficient. But the only difference between
a for-profit and a not-for-profit is that the ‘‘surplus’’ created by adding value is
used for different purposes. A not-for-profit that produces negative value (i.e.,
spends more for its output than contributors are willing to pay) will not survive
any more than a business posting continuous losses.) Boards of directors evaluate
the effectiveness and efficiency of management and have the authority and
duty to direct and replace inefficient or ineffective managers.
Copyright 2003 by The McGraw-Hill Companies, Inc. Click Here for Terms of Use.
Six Sigma’s role in all of this is to help management produce the maximum
value while using minimum resources. It does this by rationalizing management.
By this I mean that it applies scientific principles to processes and products.
By using the Six Sigma DMAIC approach processes or products are
improved in the sense that they are more effective, more efficient, or both. If
no process or product exists, or if existing processes or products are deemed
beyond repair, then design for Six Sigma (DFSS) methods are used to create
effective and efficient processes or products. Properly applied, Six Sigma minimizes
the negative impact of politics on the organization. Of course, in any
undertaking involving human beings, politics can never be completely eliminated.
Even in the best of Six Sigma organizations there will still be the occasional
Six Sigma project where data-based findings are ignored because they
conflict with the preconceived notions of a powerful figure in the organization.
But this will be the exception rather than the rule.
It should be obvious by now that I don’t view Six Sigma either as a panacea or
as a mere tool. The companies that have successfully implemented Six Sigma
are well-known, including GE, Allied Signal, Intuit, Boeing Satellite Systems,
American Express and many others. But the picture isn’t entirely rosy, failures
also exist, most notably Motorola, the company that invented Six Sigma.
Running a successful business is an extremely complicated undertaking and it
involves much more than Six Sigma. Any organization that obsesses on Six
Sigma to the exclusion of such things as radical innovation, solid financial management,
a keen eye for changing external factors, integrity in accounting, etc.
can expect to find itself in trouble some day. Markets are akin to jungles, and
much danger lurks. Six Sigma can help an organization do some things better,
but there are places where Six Sigma doesn’t apply. I seriously doubt that Six
Sigma would’ve helped Albert Einstein discover relativity or Mozart compose
a better opera. Recognizing the limits of Six Sigma while exploiting its strengths
is the job of senior leadership.
If you are working in a traditional organization, deploying Six Sigma will
rock your world. If you are a traditional manager, you will be knocked so far
out of your comfort zone that you will literally lose sleep trying to figure out
what’s happening. Your most cherished assumptions will be challenged by
your boss, the accepted way of doing things will no longer do. A new full-time,
temporary position will be created which has a single mission: change the orgaxiv
Define, Measure, Analyze, Improve, Control.
Whether Six Sigma has anything to do with Motorola’s recent problems is hotly debated. But it is undeniable that Motorola
relied heavily on Six sigma and that it has had difficulties in recent years. Still, Motorola is a fine company with a long and
splendid history, and I expect to see it back on top in the not too distant future.
nization. People with the word ‘‘belt’’ in their job title will suddenly appear,
speaking an odd new language of statistics and project management. What
used to be your exclusive turf will be identified as parts of turf-spanning processes;
your budget authority may be usurped by new ‘‘Process Owners.’’ The
new change agents will prowl the hallowed halls of your department, continuously
stirring things up as they poke here and peek there, uncovering inefficiency
and waste in places where you never dreamed improvement was
possible. Your data will be scrutinized and once indispensable reports will be
discontinued, leaving you feeling as if you’ve lost the star you use to naviage.
New reports, mostly graphical, will appear with peculiar lines on them labeled
‘‘control limits’’ and ‘‘process mean.’’ You will need to learn the meaning of
such terms to survive in the new organization; in some organizations you
won’t be eligible for advancement until you are a trained ‘‘belt.’’ In others, you
won’t even be allowed to stay.
When done properly, the result of deploying Six Sigma is an organization
that does a better job of serving owners and customers. Employees who adapt
to the new culture are better paid and happier. The work environment is exciting
and dynamic and change becomes a way of life. Decisions are based on reason
and rationality, rather than on mysterious back-room politics.
However, when done half-heartedly, Six Sigma (or any other improvement
initiative) is a colossal waste of money and time. The message is clear: do it
right, or don’t do it at all.
It has been nearly two decades since Six Sigma began and the popularity of
the approach continues to grow. As more and more firms adopt Six Sigma as
their organizational philosophy, they also adapt it to their own unique circumstances.
Thus, Six Sigma has evolved. This is especially true in the way Six
Sigma is used to operationalize the organization’s strategy. Inspired leaders,
such as Jack Welch and Larry Bossidy, have incorporated Six Sigma into the fabric
of their businesses and achieved results beyond the predictions of the most
enthusiastic Six Sigma advocate. Six Sigma has also been expanded from merely
improving existing processes to the design of new products and processes that
start life at quality and performance levels near or above Six Sigma. Six Sigma
has also been integrated with that other big productivity movement, Lean
Manufacturing. In this second edition I attempt to capture these new developments
and show how the new Six Sigma works.
Preface xv
^ ^ ^
The goal of this book remains the same as for the first edition, namely, to provide
you with the comprehensive guidance and direction necessary to realize
Six Sigma’s promise, while avoiding traps and pitfalls commonly encountered.
In this book you will find a complete overview of the management and organization
of Six Sigma, the philosophy which underlies Six Sigma, and those problem
solving techniques and statistical tools most often used in Six Sigma. It is not
intended to be an ASQ certification study guide, although it includes coverage
of most of the topics included in the ASQ body of knowledge. Rather it is
intended as a guide for champions, leaders, ‘‘belts,’’ team members and others
interested in using the Six Sigma approach to make their organizations more
efficient, more effective, or both. In short, it is a user’s manual, not a classroom
Compared to the first edition, you will find less discussion of theory. I love
theory, but Six Sigma is quite hard-nosed in its bottom-line emphasis and I
know that serious practitioners are more interested in how to use the tools and
techniques to obtain results than in the theory underlying a particular tool.
(Of course, theory is provided to the extent necessary to understand the proper
use and limitations of a given tool.) Minitab and other software are used extensively
to illustrate how to apply statistical techniques in a variety of situations
encountered during Six Sigma projects. I believe that one of the major differences
between Six Sigma and previous initiatives, such as TQM, is the integration
of powerful computer-based tools into the training. Many actual examples
are used, making this book something of a practical guide based on the school
of hard knocks.
Several different constituencies can benefit from this book. To serve these
constituents I separate the book into different parts. Part I is aimed at senior
Copyright 2003 by The McGraw-Hill Companies, Inc. Click Here for Terms of Use.
leaders and those managers who are charged with developing strategies and
deploying the Six Sigma systems within the organization. In Part I you will
find a high level presentation of the philosophy behind Six Sigma, but I get
down to the nuts and bolts very quickly. By this I mean identifying how Six
Sigma will change the organization, and answer such questions as what are the
new positions that will be created? What knowledge, skills, abilities and personal
attributes should those filling these positions possess? What personnel
assessment criteria should be used, and how can these criteria be used to evaluate
candidates? Do we need to formally test applicants? What are the specific
responsibilities of people in the organization with respect to Six Sigma?
Unless such issues are carefully considered and addressed, Six Sigma will fail.
There’s no real point to training Black Belts, Green Belts, and other parts of
the Six Sigma infrastructure if the supporting superstructure isn’t in place.
Part I also addresses the issue of linking Six Sigma to the enterprise’s strategic
goals and objectives. Six Sigma is not Management By Objectives, but MBO
didn’t fail because it was an entirely bad idea. What was missing from MBO
was an understanding that results are process-driven and the development of a
resource pool and the building of an infrastructure that was dedicated to driving
the change necessary to accomplish the objectives. With Six Sigma one doesn’t
achieve objectives by directly manipulating results, but by changing the way
things are done. The driving force behind this change are the ‘‘belts,’’ who are
highly trained full- and part-time change agents. These people lead and support
projects, and it is the projects that drive change. But not just any projects will
do. Projects must be derived from the needs of the enterprise and its customers.
This is accomplished via a rigorous flow-down process that starts at the top of
the organization. In addition to describing the mechanisms that accomplish
this linkage, Part I describes the importance of rewards and incentives to success.
In short, Six Sigma becomes the way senior leaders reach their goals.
Part II presents the tools and techniques of Six Sigma. Six Sigma provides
an improvement framework known as Define-Measure-Analyze-Improve-
Control (DMAIC), and I have elected to organize the technical material within
the DMAIC framework. It is important to note that this isn’t always the best
way to first learn these techniques. Indeed, as a consultant I find that the Black
Belt trainee often needs to use tools from the improve or control phase while
she is still working in the define or measure phase of her project. Also,
DMAIC is often used to establish ‘‘tollgates’’ at the end of each phase to help
with project tracking, but there is usually considerable back-and-forth movement
between the phases as the project progresses and one often finds that a
‘‘closed gate’’ must be kept at least partially ajar. Still,DMAICserves the important
purpose of providing a context for a given tool and a structure for the
change process.
Introduction xvii
The presentation of DMAIC is followed by design for Six Sigma (DFSS)
principles and practices. The DFSS methodology focuses on the Define-
Measure-Analyze-Design-Verify (DMADV) approach, which builds on the
reader’s understanding of DMAIC. DFSS is used primarily when there is no
process in existence, or when the existing process is to be completely redesigned.
Finally, a chapter on Lean Manufacturing provides the reader with an overview
of this important topic and discusses its relationship to Six Sigma.
DMAIC overview
. The De?ne phase of the book covers process mapping and ?owcharting,
project charter development, problem solving tools, and the so-called 7M
. Measure covers the principles of measurement, continuous and discrete
data, scales of measurement, an overview of the principles of variation,
and repeatability-and-reproducibility (RR) studies for continuous and
discrete data.
. Analyze covers establishing a process base line, how to determine process
improvement goals, knowledge discovery, including descriptive and
exploratory data analysis and data mining tools, the basic principles of statistical
process control (SPC), specialized control charts, process capability
analysis, correlation and regression analysis, analysis of categorical
data, and non-parametric statistical methods.
. Improve covers project management, risk assessment, process simulation,
design of experiments (DOE), robust design concepts (including
Taguchi principles), and process optimization.
. Control covers process control planning, using SPC for operational
control, and PRE-control.
DFSS covers the DMADV framework for process design, statistical tolerancing,
reliability and safety, using simulation software to analyze variation and
risk, and performing ‘‘virtual DOE’’ using simulation software and artificial
neural networks.
Lean covers the basic principles of Lean, Lean tools and techniques, and a
framework for deployment. It also discusses the considerable overlap between
Lean and Six Sigma and how to integrate the two related approaches to achieve
process excellence.
xviii Introduction
^ ^ ^
Six Sigma Implementation
and Management
Copyright 2003 by The McGraw-Hill Companies, Inc. Click Here for Terms of Use.
This page intentionally left blank.
^ ^ ^
Building the Six Sigma
This section provides a 10,000 foot overview of Six Sigma. Subsequent sections
elaborate and provide additional information on tools and techniques.
Six Sigma is a rigorous, focused and highly effective implementation of proven
quality principles and techniques. Incorporating elements from the work
of many quality pioneers, Six Sigma aims for virtually error free business performance.
Sigma, s, is a letter in the Greek alphabet used by statisticians to measure
the variability in any process. A company’s performance is measured by
the sigma level of their business processes. Traditionally companies accepted
three or four sigma performance levels as the norm, despite the fact that these
processes created between 6,200 and 67,000 problems per million opportunities!
The Six Sigma standard of 3.4 problems per million opportunities* is a response
to the increasing expectations of customers and the increased complexity of
modern products and processes.
If you’re looking for new techniques, don’t bother. Six Sigma’s magic isn’t in
statistical or high-tech razzle-dazzle. Six Sigma relies on tried and true methods
that have been around for decades. In fact, Six Sigma discards a great deal of
*Statisticians note: the area under the normal curve beyond Six Sigma is 2 parts-per-billion. In calculating failure rates for Six
Sigma purposes we assume that performance experienced by customers over the life of the product or process will be much
worse than internal short-term estimates predict. To compensate, a ‘‘shift’’ of 1.5 sigma from themean is added before calculating
estimated long-term failures. Thus, youwill find 3.4 parts-per-million as the area beyond 4.5 sigma on the normal curve.
Copyright 2003 by The McGraw-Hill Companies, Inc. Click Here for Terms of Use.
the complexity that characterized Total Quality Management (TQM). By one
expert’s count, there were over 400 TQM tools and techniques. Six Sigma
takes a handful of proven methods and trains a small cadre of in-house technical
leaders, known as Six Sigma Black Belts, to a high level of proficiency in the
application of these techniques. To be sure, some of the methods Black Belts
use are highly advanced, including up-to-date computer technology. But the
tools are applied within a simple performance improvement model known as
Define-Measure-Analyze-Improve-Control, or DMAIC. DMAIC is described
briefly as follows:
D De?nethe goals of the improvement activity.
M Measurethe existing system.
A Analyzethe system to identify ways to eliminate the gap
between the current performance of the system or
process and the desired goal.
I Improvethe system.
C Controlthe new system.
Why Six Sigma?
When a Japanese firm took over a Motorola factory that manufactured
Quasar television sets in the United States in the 1970s, they promptly set
about making drastic changes in the way the factory operated. Under Japanese
management, the factory was soon producing TV sets with 1/20th as many
defects as they had produced under Motorola’s management. They did this
using the same workforce, technology, and designs, and did it while lowering
costs, making it clear that the problem was Motorola’s management. It took a
while but, eventually, even Motorola’s own executives finally admitted ‘‘Our
quality stinks’’ (Main, 1994).
It took until nearly the mid-1980s before Motorola figured out what to do
about it. Bob Galvin, Motorola’s CEO at the time, started the company on
the quality path known as Six Sigma and became a business icon largely as a
result of what he accomplished in quality at Motorola. Using Six Sigma
Motorola became known as a quality leader and a profit leader. After
Motorola won the Malcolm Baldrige National Quality Award in 1988 the
secret of their success became public knowledge and the Six Sigma revolution
was on. Today it’s hotter than ever. Even though Motorola has been struggling
the past few years, companies such as GE and AlliedSignal have taken up the
Six Sigma banner and used it to lead themselves to new levels of customer service
and productivity.
It would be a mistake to think that Six Sigma is about quality in the traditional
sense. Quality, defined traditionally as conformance to internal requirements,
has little to do with Six Sigma. Six Sigma is about helping the
organization make more money by improving customer value and efficiency.
To link this objective of Six Sigma with quality requires a new definition of
quality. For Six Sigma purposes I define quality as the value added by a productive
endeavor. Quality comes in two flavors: potential quality and actual
quality. Potential quality is the known maximum possible value added per
unit of input. Actual quality is the current value added per unit of input. The
difference between potential and actual quality is waste. Six Sigma focuses
on improving quality (i.e., reducing waste) by helping organizations produce
products and services better, faster and cheaper. There is a direct correspondence
between quality levels and ‘‘sigma levels’’ of performance. For example,
a process operating at Six Sigma will fail to meet requirements about 3 times
per million transactions. The typical company operates at roughly four
sigma, which means they produce roughly 6,210 failures per million transactions.
Six Sigma focuses on customer requirements, defect prevention, cycle
time reduction, and cost savings. Thus, the benefits from Six Sigma go straight
to the bottom line. Unlike mindless cost-cutting programs which also reduce
value and quality, Six Sigma identifies and eliminates costs which provide no
value to customers, waste costs.
For non-Six Sigma companies, these costs are often extremely high.
Companies operating at three or four sigma typically spend between 25 and 40
percent of their revenues fixing problems. This is known as the cost of quality,
or more accurately the cost of poor quality. Companies operating at Six Sigma
typically spend less than 5 percent of their revenues fixing problems (Figure
1.1). COPQ values shown in Figure 1.1 are at the lower end of the range of
results reported in various studies. The dollar cost of this gap can be huge.
General Electric estimated that the gap between three or four sigma and Six
Sigma was costing them between $8 billion and $12 billion per year.
One reason why costs are directly related to sigma levels is very simple: sigma
levels are a measure of error rates, and it costs money to correct errors. Figure
1.2 shows the relationship between errors and sigma levels. Note that the error
rate drops exponentially as the sigma level goes up, and that this correlates
well to the empirical cost data shown in Figure 1.1. Also note that the errors
are shown as errors per million opportunities, not as percentages. This is
another convention introduced by Six Sigma. In the past we could tolerate percentage
error rates (errors per hundred opportunities), today we cannot.
What Is Six Sigma? 5
The Six Sigma philosophy
Six Sigma is the application of the scientific method to the design and operation
of management systems and business processes which enable employees
to deliver the greatest value to customers and owners. The scientific method
works as follows:
1. Observe some important aspect of the marketplace or your business.
2. Develop a tentative explanation, or hypothesis, consistent with your
3. Based on your hypothesis, make predictions.
4. Test your predictions by conducting experiments or making further
careful observations. Record your observations. Modify your hypothesis
based on the new facts. If variation exists, use statistical tools to help
you separate signal from noise.
5. Repeat steps 3 and 4 until there are no discrepancies between the hypothesis
and the results from experiments or observations.
At this point you have a viable theory explaining an important relationship in
your market or business. The theory is your crystal ball, which you can use to
predict the future. As you can imagine, a crystal ball is a very useful thing to
have around. Furthermore, it often happens that your theory will explain things
other than the thing you initially studied. Isaac Newton’s theory of gravity
may have begun with the observation that apples fell towards the earth, but
Newton’s laws of motion explained a great deal about the way planets moved
about the sun. By applying the scientific method over a period of years you will
develop a deep understanding of what makes your customer and your business
Figure 1.1. Cost of poor quality versus sigma level.
In Six Sigma organizations this approach is applied across the board. The
result is that political influence is minimized and a ‘‘show me the data’’ attitude
prevails. Not that corporate politics are eliminated, they can never be where
human beings interact. But politics aremuch less an influence in Six Sigma organizations
than in traditional organizations. People are often quite surprised at
the results of this seemingly simple shift in attitude. The essence of these results
is stated quite succinctly by ‘‘Pyzdek’s Law’’:
Most of what you know is wrong!
Like all such ‘‘laws,’’ this is an overstatement. However, you’ll be stunned by
how often people are unable to provide data supporting their positions on
basic issues when challenged to do so. For example, the manager of a technical
support call center was challenged by the CEO to show that customers cared
deeply about hold time. When he looked into it the manager found that customers
cared more about the time it took to reach a technician and whether or
not their issue was resolved. The call center’s information system was measuring
hold time not only as the time until the technician first answered the
phone, but also the time the customer was on hold while the technician
researched the answer to the call. The customer cared much less about this
What Is Six Sigma? 7
Figure 1.2. Error rate versus sigma level.
‘‘hold time’’ because it helped with the resolution of the issue. This fundamental
change in focus made a great deal of difference in the way the call center operated.
The Six Sigma philosophy focuses the attention of everyone on the stakeholders
for whom the enterprise exists. It is a cause-and-effect mentality. Well-designed
management systems and business processes operated by happy employees
cause customers and owners to be satisfied or delighted. Of course, none of
this is new. Most leaders of traditional organizations honestly believe that this
is what they already do. What distinguishes the traditional approach from Six
Sigma is the degree of rigor.
Six Sigma organizations are not academic institutions. They compete in the
fast-paced world of business and they don’t have the luxury of taking years to
study all aspects of a problem before deciding on a course of action. A valuable
skill for the leader of a Six Sigma enterprise, or for the sponsor of a Six Sigma
project, is to decide when enough information has been obtained to warrant taking
a particular course of action and moving on. Six Sigma leadership is very
hard-nosed when it comes to spending the shareholder’s dollars and project
research tends to be tightly focused on delivering information useful for management
decision-making. Once a level of confidence is achieved, management
must direct the Black Belt to move the project from the Analyze phase to the
What we know
We all know that there was a surge in births nine months after the November
1965 New York City power failure, right? After all, the New YorkT|mes said so
in a story that ran August 8, 1966. If that’s not prestigious enough for you, consider
that the source quoted in the T|mes article was the city’s Mt. Sinai
Hospital, one of the best.
What the data show
The newspaper compared the births on August 8, 1965 with those on August
8, 1966. This one-day comparison did indeed show an increase year-over-year.
However, J. Richard Udry, director of the Carolina Population Center at the
University of North Carolina, studied birthrates at several New York City hospitals
between July 27 and August 14, 1966. His ?nding: the birthrate nine
months after the blackout was slightly below the ?ve-year average.
Improve phase, or from the Improve phase to the Control phase. Projects are
closed and resources moved to new projects as quickly as possible.
Six Sigma organizations are not infallible, they make their share of mistakes
and miss some opportunities they might have found had they taken time to
explore more possibilities. Still, they make fewer mistakes than their traditional
counterparts and scholarly research has shown that they perform significantly
better in the long run.
While working with an aerospace client, I was helping an executive set up a
system for identifying potential Six Sigma projects in his area. I asked ‘‘What
are your most important metrics? What do you focus on?’’ ‘‘That’s easy,’’ he
responded. ‘‘We just completed our monthly ops review so I can show you.’’
He then called his secretary and asked that she bring the ops review copies.
Soon the secretary came in lugging three large, loose-leaf binders filled with
copies of PowerPoint slides. This executive and his staff spend one very long
day each month reviewing all of these metrics, hoping to glean some direction
to help them plan for the future. This is not focusing, it’s torture!
Sadly, this is not an isolated case. Over the years I’ve worked with thousands
of people in hundreds of companies and this measurement nightmare is commonplace,
even typical. The human mind isn’t designed to make sense of such
vast amounts of data. Crows can track three or four people, beyond that they
lose count.* Like crows, we can only hold a limited number of facts in our
minds at one time. We are simply overwhelmed when we try to retain too
much information. One study of information overload found the following
(Waddington, 1996):
. Two-thirds of managers report tension with work colleagues, and loss of
job satisfaction because of stress associated with information overload.
. One-third of managers su?er from ill health, as a direct consequence of
stress associated with information overload. This ?gure increases to 43%
among senior managers.
. Almost two-thirds (62%) of managers testify that their personal relationships
su?er as a direct result of information overload.
. 43% of managers think important decisions are delayed, and the ability to
make decisions is a?ected as a result of having too much information.
. 44% believe the cost of collating information exceeds its value to business.
What Is Six Sigma? 9
*See JoeWortham, ‘‘Corvus brachyhynchos,’’
Clearly, more information isn’t always better.
When pressed, nearly every executive or manager will admit that there are
a half-dozen or so measurements that really matter. The rest are either derivatives
or window dressing. When asked what really interested him, my client
immediately turned to a single slide in the middle of one of the binders.
There were two ‘‘Biggies’’ that he focused on. The second-level drill down
involved a half-dozen major drivers. Tracking this number of metrics is well
within the abilities of humans, if not crows! With this tighter focus the executive
could put together a system for selecting good Six Sigma projects and
team members.
Six Sigma activities focus on the few things that matter most to three key constituencies:
customers, shareholders, and employees. The primary focus is on
customers, but shareholder interests are not far behind. The requirements of
these two groups are determined using scientific methods, of course. But the
science of identifying what people want is not fully mature, so the data are supplemented
with a great deal of personal contact at all levels of the organization.
Employee requirements are also aggressively sought. Well-treated employees
stay longer and do a better job.
Focus comes from two perspectives: down from the top-level goals and up
from problems and opportunities. The opportunities meet the goals at the Six
Sigma project. Six Sigma projects link the activities of the enterprise to its
improvement goals. The linkage is so tight that in a well-run enterprise people
working on Six Sigma projects can tell you which enterprise objectives will be
impacted by their project, and senior leaders are able to measure the impact of
Six Sigma on the enterprise in clear and meaningful terms. The costs and benefits
of Six Sigma are monitored using enterprise-wide tracking systems that can
slice and dice the data in many different ways. At any point in time an executive
can determine if Six Sigma is pulling its weight. In many TQM programs of the
past people were unable to point to specific bottom-line benefits, so interest gradually
waned and the programs were shelved when times got tough. Six Sigma
organizations know precisely what they’re getting for their investment.
Six Sigma also has an indirect benefit on an enterprise, and one that is seldom
measured. That benefit is its impact on the day-to-day way of doing
things. Six Sigma doesn’t operate in a vacuum. When people observe Six
Sigma getting dramatic results, they naturally modify the way they approach
their work. Seat-of-the-pants management doesn’t sit well (pardon the pun!)
in Six Sigma organizations that have reached ‘‘critical mass.’’ Critical mass
occurs when the organization’s culture has changed as a result of Six Sigma
being successfully deployed in a large segment of the organization. The initial
clash of cultures has worked itself out and those opposed to the Six Sigma
way have either left, converted, or learned to keep quiet.
There is also a ‘‘dark side’’ to Six Sigma that needs to be discussed. There are
parts of the enterprise that don’t lend themselves to scientific rigor. For example,
successful R&D involves a good deal of original creative thinking. The
‘‘R’’ (research) part of R&D may actually suffer from too much rigor and the
Six Sigma focus on defects. Cutting edge research is necessarily trial and error
and requires a high tolerance for failure. The chaos of exploring new ideas is
not something to be managed out of the system, it is to be expected and encouraged.
To the extent that it involves process design and product testing, Six
Sigma may be able to make a contribution to the ‘‘D’’ (development) part of
R&D. The point is to selectively apply Six Sigma to those areas where it will
provide a benefit.
A second aspect of Six Sigma’s dark side is that some companies obsess on it
to the exclusion of other important aspects of the business. Business is a complex
undertaking and leading a business enterprise requires creativity, innovation,
and intuition. While it’s all well and good to be ‘‘data driven,’’ leaders
need to heed their inner voice as well. Keep in mind that some of the most
important things are unmeasured and immeasurable. Challenge counterintuitive
data and subject it to a gut check. It may be that the counterintuitive result
represents a startling breakthrough in knowledge, but it may simply be wrong.
Here’s an example. A software client had a technical support call center to
help their customers solve problems with the software. Customer surveys were
collected and the statistician made an amazing discovery; hold time didn’t matter!
The data showed that customer satisfaction was the same for customers
served immediately and for those on hold for an hour or more. Discussions
began along the lines of how many fewer staff would be required due to this
new information. Impressive savings were forecast.
Fortunately, the support center manager hadn’t left his skepticism at the
front door. He asked for additional data, which showed that the abandon rate
increased steadily as people were kept on hold. The surveys were given only to
those people who had waited for service. These people didn’t mind waiting.
Those who hung up the phone before being served apparently did. In fact,
when a representative sample was obtained, excessive hold time was the number
one complaint.
The change imperative
Six Sigma is not a completely new way to manage an enterprise, but it is a very
different way. In essence, Six Sigma forces change to occur in a systematic way.
What Is Six Sigma? 11
In traditional organizations the job of management is to design systems to create
and deliver value to customers and shareholders. This is, of course, a neverending
task. Competitors constantly innovate in an attempt to steal your customers.
Customers continuously change their minds about what they want.
Capital markets offer investors new ways to earn a return on their investment.
The result is an imperative to constantly change management systems.
Despite the change imperative, most enterprises resist change until there are
obvious signs that current systems are failing one or more stakeholder groups.
Perhaps declining market share makes it clear that your products or services are
not as competitive as they once were. Or maybe your customers are still loyal,
but customer complaints have reached epidemic proportions. Or your share
price may be trending ominously downward. Traditional organizations watch
for such signs and react to them. Change occurs, as it must, but it does so in an
atmosphere of crisis and confusion. Substantial lossmayresult before the needed
redesign is complete. People may lose their jobs or even their careers. Many organizations
that employ these reactionary tactics don’t survive the shock.
The Six Sigma enterprise proactively embraces change by explicitly incorporating
change into their management systems. Full- and part-time change
agent positions are created and a complete infrastructure is created. As contradictory
as it sounds, the infrastructure is designed to make change part of the
routine. New techniques are used to monitor changing customer, shareholder,
and employee inputs, and to rapidly integrate the new information by changing
business processes. The approach employs sophisticated computer modeling,
mathematics, and statistical analysis to minimize unneeded tampering by separating
signal from noise. These analytical techniques are applied to stakeholder
inputs and to enterprise and process metrics at all levels.
As a consequence of deploying Six Sigma, people require a great deal of training.
Communication systems are among the first things that need to be changed
so people know what to make of the new way of doing things. Think about it;
when Six Sigma is deployed the old reports are no longer used. Six Sigma
requires that internal data be presented only if there is a direct linkage to a stakeholder.
The phrase ‘‘How do you know?’’ is heard repeatedly.
. ‘‘Nice report on on-time deliveries, Joan, but show me why you think this
is important to the customer. If it is, I want to see a chart covering the
last 52 weeks, and don’t forget the control limits.’’
. ‘‘This budget variance report is worthless! I want to see performance
across time, with control limits.’’
. ‘‘Have these employee survey results been validated? What is the reliability
of the questions? What are the main drivers of employee satisfaction?
How do you know?’’
. ‘‘How do your dashboards relate to the top-level dashboards?’’
Add to this the need to do more than simply operate the system you work
with. Six Sigma demands that you constantly look for ways to improve your systems.
This often means that systems are eliminated entirely. In the face of this
insecurity, employees watch like a hawk for signs of leadership inconsistency.
Trust is essential. Leaders who don’t communicate a clear and consistent
message and walk the talk will be faced with stiff resistance to Six Sigma.
The need for a well-designed approach to making the transition from a traditional
organization to a Six Sigma organization is clear. This is the subject of
Part I of this book. It is the foundation building phase. If it isn’t done properly,
then the DMAIC approach and all of the tools and techniques presented later
in the book will be of little use.
Change agents and their effects on organizations
Experts agree: change is difficult, disruptive, expensive, and a major cause of
error. Given these problems, it’s reasonable to ask: Why change? Here are the
most common reasons organizations choose to face the difficulties involved
with change:
. LeadershipLSome organizations choose to maintain product or service
leadership as a matter of policy. Change is a routine.
. CompetitionLWhen competitors improve their products or services
such that their o?ering provides greater value than yours, you are forced
to change. Refusal to do so will result in the loss of customers and revenues
and can even lead to complete failure.
. Technological advancesLE?ectively and quickly integrating new technology
into an organization can improve quality and e?ciency and provide
a competitive advantage. Of course, doing so involves changing
management systems.
. Training requirementsLMany companies adopt training programs
without realizing that many such programs implicitly involve change.
For example, a company that provides employees with SPC training
should be prepared to implement a process control system. Failure to do
so leads to morale problems and wastes training dollars.
. Rules and regulationsLChange can be forced on an organization from
internal regulators via policy changes and changes in operating procedures.
Government and other external regulators and rule-makers (e.g.,
ISO for manufacturing, JCAHO for hospitals) can also mandate change.
. Customer demandsLCustomers, large and small, have the annoying
habit of refusing to be bound by your policies. The nice customers will
What Is Six Sigma? 13
demand that you change your policy and procedures. The really nasty
customers don’t say anything at all, they simply go somewhere else to do
Johnson (1993b, p. 233) gives the following summary of change management:
1. Change will meet resistance for many di?erent reasons.
2. Change is a balance between the stable environment and the need to
implement TQM [Six Sigma]. Change can be painful while it provides
many improvements.
3. There are four times change can most readily be made by the leader:
when the leader is new on the job, receives new training, has new technology,
or when outside pressures demand change.
4. Leaders must learn to implement change they deem necessary, change
suggested from above their level, and change demanded from above
their level.
5. There are all kinds of reaction to change. Some individuals will resist,
some will accept, and others will have mixed reactions.
6. There is a standard process that supports the implementation of change.
Some of the key requirements for change are leadership, empathy, and
solid communications.
7. It is important that each leader become a change leader. This requires
self-analysis and the will to change those things requiring change.
Change requires new behaviors from everyone involved. However, four specific
roles commonly appear during most successful change processes
(Hutton, 1994, pp. 2^4):
. O?cial change agent. An o?cially designated person who has primary
responsibility for helping management plan and manage the change process
[Sometimes called ‘‘Champions.’’]
. Sponsors. Sponsors are senior leaders with the formal authority to legitimize
the change. The sponsor makes the change a goal for the organization
and ensures that resources are assigned to accomplish it. No
major change is possible without committed and suitably placed sponsors.
. Advocate. An advocate for change is someone who sees a need for change
and sets out to initiate the process by convincing suitable sponsors. This
is a selling role. Advocates often provide the sponsor with guidance and
advice. Advocates may or may not hold powerful positions in the organization.
. Informal change agent. Persons other than the o?cial change agent who
voluntarily help plan and manage the change process. While the contri-
bution of these people is extremely important, it is generally not su?cient
to cause truly signi?cant, organization-wide change.
The position of these roles within a typical organizational hierarchy is illustrated
graphically in Figure 1.3.
There are three goals of change:
1. Change the way people in the organization think.Helping people change
the way they think is a primary activity of the change agent. All change
begins with the individual, at a personal level. Unless the individual is willing
to change his behavior, no real change is possible. Changing behavior
requires a change in thinking. In an organization where people are
expected to use their minds, people’s actions are guided by their thoughts
and conclusions. The change agent’s job starts here.
2. Change the norms. Norms consist of standards, models, or patterns
which guide behavior in a group. All organizations have norms or expec-
What Is Six Sigma? 15
Figure 1.3. Cascading of sponsorship.
From Hutton, D.W. (1994). The Change Agent’s Handbook: A Survival Guide for Quality
Improvement Champions. Copyright#1994 by David W. Hutton.
Reprinted with permission.
tations of their members. Change cannot occur until the organization’s
norms change.
3. Change the organization’s systems or processes. This is the ‘‘meat’’ of the
change. Ultimately, all work is a process and quality improvement
requires change at the process and system level. However, this cannot
occur on a sustained basis until individuals change their behavior and
organizational norms are changed.
Mechanisms used by change agents
The change agents help accomplish the above goals in a variety of ways.
Education and training are important means of changing individual perceptions
and behaviors. In this discussion, a distinction is made between training and
education. Training refers to instruction and practice designed to teach a person
how to perform some task. Training focuses on concretes that need to be done.
Education refers to instruction in how to think. Education focuses on integrating
abstract concepts into one’s knowledge of the world. An educated person
will view the world differently after being educated than they did before. This
is an essential part of the process of change.
Change agents help organize an assessment of the organization to identify its
strengths and weaknesses. Change is usually undertaken to either reduce areas
of weakness, or exploit areas of strength. The assessment is part of the education
process. Knowing one’s specific strengths and weaknesses is useful in mapping
the process for change.
Change agents play an important role in quality improvement (remember,
‘‘improvement’’ implies change). As shown in Figure 1.3, change agents are in
strategic positions throughout the organization. This makes it possible for
them to assist in the coordination of the development and implementation of
quality improvement plans. Quality improvement of any significance nearly
always involves multiple departments and levels in the organization.
In the final analysis, all we humans really have to spend is our time. Change
agents see to it that senior management spends sufficient time on the transformation
process. Senior managers’ time is in great demand from a large number
of people inside and outside of the organization. It is all too easy to
schedule a weekly meeting to discuss ‘‘Six Sigma’’ for an hour, then think you’ve
done your part. In fact, transforming an organization, large or small, requires a
prodigious commitment of the time of senior leadership. At times the executive
will not understand what he or she is contributing by, say, attending team meetings.
The change agent must constantly assure the leader that time spent on
transformation activities is time well spent.
One way of maximizing the value of an executive’s time investment is for the
executive to understand the tremendous power of certain symbolic events.
Some events generate stories that capture the essence of management’s commitment
(or lack of it) to the change being undertaken. People repeat stories
and remember them far better than proclamations and statements. For example,
there’s a story told by employees of a large U.S. automotive firm that goes
as follows:
In the early 1980s the company was just starting their quality improvement
e?ort. At a meeting between upper management and a famous quality
consultant, someone casually mentioned that quality levels were
seasonalLquality was worse in the summer months. The consultant
asked why this should be so. Were di?erent designs used? Were the
machines di?erent? How about the suppliers of raw materials? The
answer to each of these questions was ‘‘No.’’ An investigation revealed
that the problem was vacations. When one worker went on vacation,
someone else did her job, but not quite as well. And that ‘‘someone’’ also
vacated a job, which was done by a replacement, etc. It turned out that
the one person going on vacation lead to six people doing jobs they did
not do routinely. The solution was to have a vacation shutdown of two
weeks. This greatly reduced the number of people on new jobs and
brought summer quality levels up to the quality levels experienced the
rest of the year.
This worked ?ne for a couple of years since there was a recession in the
auto industry and there was plenty of excess capacity. However, one summer
the senior executives were asked by the ?nance department to reconsider
their shutdown policy. Demand had picked up and the company
could sell every car it could produce. The accountants pointed out that
the shutdown would cost $100 million per day in lost sales.
The vice president of the truck division asked if anything had been done
to address the cause of the quality slippage in the summer. No, nothing
had been done. The president asked the sta? ‘‘If we go back to the old policy,
would quality levels fall like they did before?’’ Yes, he was told, they
would. ‘‘Then we stay with our current policy and shut down the plants
for vacations,’’ the President announced.
The President was challenged by the vice president of ?nance. ‘‘I know
we’re committed to quality, but are you sure you want to lose $1.4 billion
in sales just to demonstrate our commitment?’’ The President replied,
‘‘Frank, I’m not doing this to ‘demonstrate’ anything. We almost lost our
company a few years back because our quality levels didn’t match our
overseas competition. Looking at this as a $1.4 billion loss is just the kind
What Is Six Sigma? 17
of short-term thinking that got us in trouble back then. I’m making this
decision to save money.’’
This story had tremendous impact on the managers who heard it, and it
spread like wildfire throughout the organization. It demonstrated many
things simultaneously: senior leadership’s commitment to quality, political
parity between operations and finance, how seemingly harmless policies
can have devastating effects, an illustration of how short-term thinking had
damaged the company in the past, and how long-term thinking worked in a
specific instance, etc. It is a story worth 100 speeches and mission statements.
Leadership support activities
The change agent provides technical guidance to the leadership team. This
guidance takes the form of presenting management with alternative strategies
for pursuing the transformation, education on methods that can be used to
implement the strategies, and selection of key personnel for key transformation
Change agents help to monitor the status of quality teams and quality projects
relating to the transformation (see Chapter 15 for a complete discussion
of project management). In addition to being a vehicle for local quality improvement,
projects can be used as one of the mechanisms for actually implementing
the transformation. If used in this way, it is important that projects be properly
chartered to align the project activities with the goals of the transformation.
All teams, chartered or not, must avoid projects and methods that conflict
with the goals of the transformation. Project team membership must be carefully
planned to assure that both task and group maintenance roles are properly
filled. Project charters must clearly identify the scope of the project to prevent
the confusion between teams that results from overlapping charters.
Change agents also serve as coaches to senior leaders. Culture involves innumerable
subtle characteristics and behaviors that become unconsciously
‘‘absorbed’’ into one’s being. At times, it is nearly impossible for the individual
executive to see how his or her behavior or relationships are interpreted by
others. The change agent must quietly advise leadership on these issues.
The press of day-to-day business, combined with the inherent difficulties of
change, make it easy to let time slip by without significant progress. Keeping
operations going is a full-time job, and current problems present themselves
with an urgency that meeting a future goal can’t match. Without the constant
reminders from change agents that goals aren’t being met, the leadership can
simply forget about the transformation. It is the change agent’s job to become
the ‘‘conscience’’ of the leadership and to challenge them when progress falls
short of goals.
Change networks
Change agents should work to develop an internal support network. The network
provides resources to support the change process by disseminating education
and guidance. The network’s tasks will eventually be subsumed by normal
operations, but in the early stages of the transformation it is vital that the network
exist since the control of resources is determined by the existing infrastructure
and may be difficult to divert to the change process. Usually, the
members of the network are formal and informal change agents in various
areas of the organization.
Once the network has been established, it is the change agent’s job to assure
that the activities in the network are consistent with and in support of the organization’s
vision. For example, if a hospital has a vision where physicians obtain
real-time clinical information and decision support at the patient’s bedside,
then a financially based and centralized information system is inconsistent
with that vision. The change agent, especially the formal change agent, provides
leadership and moral support to networkmembers, who may otherwise feel isolated
and vulnerable. Change agents ensure that members of the network
receive the education and guidance they need. Finally, the change agent acts as
a conduit and a stimulant to maintain regular communication in the network.
This may take the form of setting up an email list, scheduling lunches for
network members, etc.
Transforming sta? functions
Table 1.1 illustrates the contrast between the way that staff functions used
to operate under the traditional system of management, and the way they can
operate more effectively.
There are several ways in which change agents can assist staff functions in
transforming their roles:
. Collaborate with sta? functions.
. Encourage sta? functions to take a proactive approach to change.
. Make support functions partners in the support network.
. Encourage sta? members to become role models.
. Help sta? functions develop transition plans that are aligned and integrated
with the overall transformation plan.
. Encourage sta? members to share their concerns.
What Is Six Sigma? 19
After nearly two decades of experience with Six Sigma and TQM, there is
now a solid body of scientific research regarding the experience of thousands
of companies implementing major programs such as Six Sigma. Researchers
Table 1.1. How sta? functions are changing.
From Hutton, D.W. (1994). The Change Agent’s Handbook: A Survival Guide for Quality
Improvement Champions. Page 220. Copyright#1994 by David W. Hutton.
Reprinted with permission.
Role CustomerLfor information,
evidence, and reports from
SupplierLof information,
expertise, and other services
Strategy ControlLby imposition of
policies and procedures, and by
audit and inspection
SupportLby gearing e?orts
to the needs of others
Self-control by client
Goal DepartmentalLachievement of
departmental objectives
Collective achievement of the
organization’s objectives
Style of working
with others
Competitive, adversarial Integrating, collaborative
Focus of
Some aspects of outcomes; for
example, product quality,
?nancial results
Some pieces of the process; for
example, adherence to policy
and procedure
The relationship between
the entire underlying process
and the achievement of all
the desired outcomes
Image Regulator, inspector, policeman Educator, helper, guide
have found that successful deployment of Six Sigma involves focusing on a small
number of high-leverage items. The steps required to successfully implement
Six Sigma are well-documented.
1. Successful performance improvement must begin with senior leadership.
Start by providing senior leadership with training in the philosophy,
principles, and tools they need to prepare their organization for
success. Using their newly acquired knowledge, senior leaders direct
the development of a management infrastructure to support Six Sigma.
Simultaneously, steps are taken to ‘‘soft-wire’’ the organization and to
cultivate an environment where innovation and creativity can ?ourish.
This involves reducing levels of organizational hierarchy, removing procedural
barriers to experimentation and change, and a variety of other
changes designed to make it easier to try new things without fear of
2. Systems are developed for establishing close communication with customers,
employees, and suppliers. This includes developing rigorous methods
of obtaining and evaluating customer, owner, employee, and
supplier input. Base line studies are conducted to determine the starting
point and to identify cultural, policy, and procedural obstacles to
3. Training needs are rigorously assessed. Remedial basic skills education
is provided to assure that adequate levels of literacy and numeracy
are possessed by all employees. Top-to-bottom training is
conducted in systems improvement tools, techniques, and philosophies.
4. A framework for continuous process improvement is developed, along
with a system of indicators for monitoring progress and success. Six
Sigma metrics focus on the organization’s strategic goals, drivers, and
key business processes.
5. Business processes to be improved are chosen by management, and by
people with intimate process knowledge at all levels of the organization.
Six Sigma projects are conducted to improve business performance
linked to measurable ?nancial results. This requires knowledge of the
organization’s constraints.
6. Six Sigma projects are conducted by individual employees and teams
lead by Green Belts and assisted by Black Belts.
Although the approach is simple, it is by no means easy. But the results justify
the effort expended. Research has shown that firms that successfully
implement Six Sigma perform better in virtually every business category,
including return on sales, return on investment, employment growth, and
share price increase.
Implementing Six Sigma 21
Six Sigma’s timeline is usually very aggressive. Typically, companies look for
an improvement rate of approximately 10 every two years, measured in
terms of mistakes or errors using defects per million opportunities (DPMO).*
The subject of DPMO is treated in greater detail elsewhere in this book. For
our purposes here, think of DPMO as the organization’s overall performance
as observed by customers. While calculating this can become very complicated,
for illustration we will look at a very simple example. Assume that you have
the data on key processes in a technical support call center operation shown in
Table 1.2. It is very important to understand that the requirements shown in
this table are derived from customer input. For example, in Table 1.2, the 5 minute
hold time requirement assumes that we have surveyed customers and
found that they are willing to accept hold times of 5 minutes or less. Likewise,
we have data to indicate that support engineers rated higher than 5 are acceptable
to customers. ‘‘Problem resolved’’ means that the customer told us his
problem was resolved.
A Six Sigma program on a typical timetable would seek to reduce the overall
DPMO from approximately 58,000 to about 5,800 in two years time. This
would improve the sigma level from 3.1 to around 4.0 (see Figure 1.2).
Remember, Six Sigma corresponds to aDPMOof 3.4, so there’s still away to go.
*This is about twice the rate of improvement reported by companies using TQM. For example, Baldrige winner Milliken &
Co. implemented a ‘‘ten-four’’ improvement program requiring reductions in key adverse measures by a factor of ten every
four years.
Table 1.2. Process defect rates.
Process Element
Calls Meeting
Requirement DPMO
Hold time
<5 minutes
120,000 110,000 83,333 2.9
SE rating >5 119,000 118,000 8,403 3.9
Problem resolved 125,000 115,000 80,000 2.9
Total 364,000 343,000 57,692 3.1
The time needed to reach Six Sigma performance levels depends on the
organization’s starting point and their level of commitment. Figure 1.4 provides
a rough guideline for determining when you will reach Six Sigma,
assuming an aggressive deployment schedule. The times are only approximate,
your mileage may vary. Keep in mind that even if the enterprise is
operating at, say, 5 or 6 Sigma overall, there may still be processes operating
at poor sigma levels. Never forget that individual customers judge your
organization based on their individual experiences with you. Relationships
are made one person at a time. For our example, the company can expect
it to take about five years from the time they have deployed Six Sigma to
the time they begin to approach Six Sigma performance levels. If they follow
the deployment timeline shown in Figure 1.4 it will be approximately
seven years altogether. This is not to say that they’ll have to wait seven
years to see results. Results will begin to appear within a year of starting
the deployment.
Implementing Six Sigma 23
Figure 1.4. Time to Reach Six Sigma performance levels.
Obtaining these revolutionary rates of improvements will not come without
concerted effort. An aggressive deployment plan must be developed. Figure 1.5
shows a typical set of deployment activities and a timetable designed to reach
maturity within two years. This is not to say that the enterprise will be finished
in two years, nor that all of its processes will be operating at Six Sigma performance
levels. The organization is never finished! Competition, innovation,
changing customer requirements and a host of other factors will assure that
the quest for excellence is ongoing. However, if the enterprise completes the
tasks depicted in Figure 1.5 the systems and infrastructure needed to keep
them at the cutting edge will have been developed.
The deployment timetable shown in Figure 1.5 will produce sufficient savings
to cover its costs during the first year. In the second and subsequent years
the benefits will outpace the costs. The benefit-to-cost ratio will improve as
time goes by. Figure 1.6 shows General Electric’s published data on their Six
Figure 1.5. Typical deployment activities and timeline.
Sigma program. Note that in 1996, the first full year of GE’s program, costs and
benefits were approximately equal. The amount by which benefits exceed costs
is increasing because, while costs level out, benefits continue to increase. These
results are consistent with those reported by academic research for companies
which implemented TQM.
A very powerful feature of Six Sigma is the creation of an infrastructure to
assure that performance improvement activities have the necessary resources.
In this author’s opinion, failure to provide this infrastructure is a major reason
why 80% of all TQM implementations failed in the past. TQM presented general
principles and left it to each organization to decide how to put the principles
into practice. Companies that did an excellent job of operationalizing the
principles of TQM obtained excellent results, comparable to the results
reported by companies implementing Six Sigma. Those that didn’t, failed. Six
Sigma provides a quasi-standardized set of guidelines for deployment. This is
Implementing Six Sigma 25
Figure 1.6. GE’s reported cost of Six Sigma versus bene?ts.
why, I believe, Six Sigma enjoys a much higher success rate than TQM. Of
course, there are still those companies that kludge together half-hearted efforts
and call it Six Sigma. They will fail just as those who deployed half-baked
TQM programs failed.
Six Sigma makes improvement and change the full-time job of a small but critical
percentage of the organization’s personnel. These full-time change agents
are the catalyst that institutionalizes change. Figure 1.7 illustrates the commitment
required by Six Sigma.
Assessing organization culture on quality
Juran and Gryna (1993) define the company quality culture as the opinions,
beliefs, traditions, and practices concerning quality. While sometimes difficult
to quantify, an organization’s culture has a profound effect on the quality
produced by that organization. Without an understanding of the cultural
Figure 1.7. Six Sigma infrastructure.
aspects of quality, significant and lasting improvements in quality levels are
Two of the most common means of assessing organization culture is the
focus group and the written questionnaire. These two techniques are discussed
in greater detail below. The areas addressed generally cover attitudes, perceptions,
and activities within the organization that impact quality. Because of the
sensitive nature of cultural assessment, anonymity is usually necessary. The
author believes that it is necessary for each organization to develop its own set
of questions. The process of getting the questions is an education in itself. One
method for getting the right questions that has produced favorable results in
the past is known as the critical-incident technique. This involves selecting a
small representative sample (n  20) from the group you wish to survey and
asking open-ended questions, such as:
‘‘Which of our organization’s beliefs, traditions and practices have a
bene?cial impact on quality?’’
‘‘Which of our organization’s beliefs, traditions and practices have a detrimental
impact on quality?’’
The questions are asked by interviewers who are unbiased and the respondents
are guaranteed anonymity. Although usually conducted in person or
by phone, written responses are sometimes obtained. The order in which the
questions are asked (beneficial/detrimental) is randomized to avoid bias in
the answer. Interviewers are instructed not to prompt the respondent in any
way. It is important that the responses be recorded verbatim, using the
respondent’s own words. Participants are urged to provide as many responses
as they can; a group of 20 participants will typically produce 80^100
The responses themselves are of great interest and always provide a great deal
of information. In addition, the responses can be grouped into categories and
the categories examined to glean additional insight into the dimensions of the
organization’s quality culture. The responses and categories can be used to
develop valid survey items and to prepare focus-group questions. The followup
activity is why so few people are needed at this stageLstatistical validity is
obtained during the survey stage.
Six Sigma involves changing major business value streams that cut across
organizational barriers. It provides the means by which the organization’s strategic
goals are to be achieved. This effort cannot be lead by anyone other than
the CEO, who is responsible for the performance of the organization as a
Implementing Six Sigma 27
whole. Six Sigma must be implemented from the top down. Lukewarm leadership
endorsement is the number 1 cause of failed Six Sigma attempts.
Conversely, I don’t know of a single case where top leadership fully embraced
Six Sigma (or TQM, for that matter) that hasn’t succeeded. Six Sigma has zero
chance of success when implemented without leadership from the top. This
is because of the Six Sigma focus on cross-functional, even enterprise-wide
processes. Six Sigma is not about local improvements, which are the only
improvements possible when top-level support is lacking.
Six Sigma champions are high-level individuals who understand Six Sigma
and are committed to its success. In larger organizations Six Sigma will be
lead by a full-time, high-level champion, such as an Executive Vice President.
In all organizations, champions also include informal leaders who use Six
Sigma in their day-to-day work and communicate the Six Sigma message at
every opportunity. Sponsors are owners of processes and systems who help
initiate and coordinate Six Sigma improvement activities in their areas of
Candidates for Black Belt status are technically oriented individuals held in
high regard by their peers. They should be actively involved in the process of
organizational change and development. Candidates may come from a wide
range of disciplines and need not be formally trained statisticians or analysts.
However, because they are expected to master a wide variety of technical tools
in a relatively short period of time, Black Belt candidates will probably possess
a background in college-level mathematics, the basic tool of quantitative analysis.
Coursework in statistical methods should be considered a strong plus or
even a prerequisite. As part of their training, Black Belts typically receive 160
hours of classroom instruction, plus one-on-one project coaching from Master
Black Belts or consultants. The precise amount of training varies considerably
from firm to firm. In the financial sector Black Belts generally receive three
weeks of training, while Black Belts in large research facilities may get as much
as six weeks of training.
Successful candidates will be comfortable with computers. At a minimum,
they should be proficient with one or more operating systems, spreadsheets,
database managers, presentation programs, and word processors. As part of
their training they will also be required to become proficient in the use of one
or more advanced statistical analysis software packages and probably simula-
tion software. Six Sigma Black Belts work to extract actionable knowledge from
an organization’s information warehouse. To assure access to the needed information,
Six Sigma activities should be closely integrated with the information
systems of the organization. Obviously, the skills and training of Six Sigma
Black Belts must be enabled by an investment in software and hardware. It
makes no sense to hamstring these experts by saving a few dollars on computers
or software.
Green Belts are Six Sigma project leaders capable of forming and facilitating
Six Sigma teams and managing Six Sigma projects from concept to completion.
Green Belt training consists of five days of classroom training and is conducted
in conjunction with Six Sigma projects. Training covers project management,
quality management tools, quality control tools, problem solving, and descriptive
data analysis. Six Sigma champions should attend Green Belt training.
Usually, Six Sigma Black Belts help Green Belts define their projects prior to
the training, attend training with their Green Belts, and assist them with their
projects after the training.
This is the highest level of technical and organizational proficiency. Master
Black Belts provide technical leadership of the Six Sigma program. Thus,
they must know everything the Black Belts knows, as well as additional skills
vital to the success of the Six Sigma program. The additional skill might be
deep understanding of the mathematical theory on which the statistical methods
are based. Or, perhaps, a gift for project management, coaching skills to
help Black Belts, teaching skills, or program organization at the enterprise
level. Master Black Belts must be able to assist Black Belts in applying the
methods correctly in unusual situations, especially advanced statistical methods.
Whenever possible, statistical training should be conducted only by qualified
Master Black Belts or equivalently skilled consultants. If it becomes
necessary for Black Belts and Green Belts to provide training, they should
only do so under the guidance of Master Black Belts. Otherwise the familiar
‘‘propagation of error’’ phenomenon will occur, i.e., Black Belt trainers pass
on errors to Black Belt trainees who pass them on to Green Belts, who pass
on greater errors to team members. Because of the nature of the Master’s
duties, all Master Black Belts must possess excellent communication and
teaching skills.
Implementing Six Sigma 29
As stated earlier in this chapter, the number of full-time personnel devoted to
Six Sigma is not large as a percentage of the total work force. Mature Six Sigma
programs, such as those of General Electric, Johnson & Johnson, AlliedSignal,
and others average about one percent of their workforce as Black Belts, with
considerable variation in that number. There is usually about one Master
Black Belt for every ten Black Belts, or about one Master Black Belt per 1,000
employees. A Black Belt will typically complete 5 to 7 projects per year, usually
working with teams. Project teams are often lead by Green Belts, who, unlike
Black Belts and Master Black Belts, are not employed full time in the Six
Sigma program. Green Belts usually devote between 5 and 10 percent of their
time to Six Sigma project work.
Black Belts are highly prized employees and are often recruited for key management
positions elsewhere in the company. After Six Sigma has been in place
for three or more years, the number of former Black Belts in management positions
tends to be greater than the number of active Black Belts. These people
take the rigorous, customer-driven, process-focused Six Sigma approach with
themwhenthey movetonewjobs. The ‘‘Six Sigma way’’ soon becomes pervasive.
Estimated savings per project vary from organization to organization.
Reported results average about $150,000 to $243,000. Some industries just starting
their Six Sigma programs average as high as $700,000 savings per project,
although these projects usually take longer. Note that these are not the huge
mega-projects pursued by reengineering. Still, by completing 5 to 7 projects
per year per Black Belt the company will add in excess of $1 million per year
per Black Belt to its bottom line. For a company with 1,000 employees the
numbers would look something like this:
Master Black Belts: 1
Black Belts: 10
Projects: 50 to 70 (5 to 7 per Black Belt)
Estimated saving: $9 million to $14.6 million (i.e., $14,580 savings
per employee)
Do the math for your organization and see what Six Sigma could do for you.
Because Six Sigma savingsLunlike traditional slash and burn cost cuttingL
impact only non-value-added costs, they flow directly to your company’s bottom
line. Traditional, income-statement based cost cutting inevitably hurts
value-adding activities. As a result, the savings seldom measure up to expectations
and revenues often suffer as well. The predicted bottom-line impact is
not actually realized. Firms engaging in these activities hurt their prospects for
future success and delay their recovery.
Six Sigma deployment and management
Six Sigma deployment is the actual creation of an organization that embodies
the Six Sigma philosophy. Doing this involves asking a series of questions,
then answering these questions for organizational units. The deployment process
is outlined in Figure 1.8.
Creating an organization to carry out this process is no easy task. Traditional
organizations are structured to carry out routine tasks, while Six Sigma is all
about non-routine activity. Look at the action words in Figure 1.8: improve,
increase, eliminate, reduce, breakthrough. These are challenging things to do in
any environment, and nearly impossible in an enterprise focused on carrying
out routine assignments. The job of the leadership team is to transform the organization’s
culture so that Six Sigma will flourish. It’s a tough job, but not an
impossible one. Think of it as writing a book. No one sits down and writes a
book as a single unit. Books are organized into smaller sub-units, such as sections,
chapters, pages, and paragraphs. Similarly, deploying Six Sigma involves
sub-units, such as those shown in Figure 1.9.
Although leadership is ultimately responsible for creating the Six Sigma
Deployment Manual, they will not have the time to write it themselves.
Writing the manual is itself a project, and it should be treated as such. A formal
charter should be prepared and responsibility for writing the deployment manual
should be assigned by senior leadership to a project sponsor. The sponsor
should be a senior leader, either the CEO or a member of the CEO’s staff. An
aggressive deadline should be set. The plan itself is the deliverable, and the
requirements should be clearly defined. The CEO and the Executive Six Sigma
Council must formally accept the plan.
All of the subjects in the table of contents in Figure 1.9 are discussed in this
book. Some are covered in several pages, while others take an entire chapter or
more. Although you won’t be able to get all of your answers from a single
book, or from any number of books, the material you will find here should
give you enough information to start the journey. You will encounter enough
challenges along the way that perhaps the word ‘‘adventure’’ would be more
Six Sigma communication plan
Successful implementation of Six Sigma will only happen if the leadership’s
vision and implementation plans are clearly understood and embraced by
employees, shareholders, customers, and suppliers. Because it involves cultural
change, Six Sigma frightens many people. Good communications are an antidote
to fear. Without it rumors run rampant and morale suffers. Although
Implementing Six Sigma 31
1. Deployment goals
1.1. Business level
1.1.1. Increase shareholder value
1.1.2. Increase revenues
1.1.3. Improve market share
1.1.4. Increase profitability and ROI
1.2. Operations level
1.2.1. Eliminate ‘‘hidden factory’’ (i.e., resources used because things
were not done right the first time)
1.2.2. Improve rolled throughput yield and normalized yield
1.2.3. Reduce labor costs
1.2.4. Reduce material costs
1.3. Process level
1.3.1. Improve cycle time
1.3.2. Reduce resource requirements
1.3.3. Improve output volume
1.3.4. Improve process yield (ratio of inputs to outputs)
1.3.5. Reduce defects
1.3.6. Reduce variability
1.3.7. Improve process capability
2. Identify key value streams
2.1. Which processes are critical to business performance?
2.2. How do processes deliver value to customers?
3. Determine metrics and current performance levels
3.1. How will we measure key value streams?
3.2. Are our measurements valid, accurate, and reliable?
3.3. Are the processes stable (i.e., in statistical control)?
3.3.1. If not, why not?
3.3.2. What are the typical cycle times, costs, and quality opportunities
of these processes?
3.3.3. What is the short- and long-term process capability?
3.4. Detailed as-is and should-be process maps for critical processes
3.5. How does current performance relate to benchmark or best-in-class
4. Breakthrough to new performance levels
4.1. Which variables make the most difference?
4.2. What are the best settings for these variables?
4.3. Can the process be redesigned to become more robust?
4.4. Can product be redesigned to become more robust and/or more easily
5. Standardize on new approach
5.1. Write procedures describing how to operate the new process
5.2. Train people in the new approach
5.3. When necessary, use SPC to control process variation
5.4. Modify inventory, cost accounting, and other business systems to
assure that improved process performance is reflected in bids, order
quantities, inventory trigger points, etc.
Figure 1.8. Six Sigma deployment process outline.
change is the byword of Six Sigma, you should try to cause as little unnecessary
disruption as possible. At the same time, the commitment of the enterprise to
Six Sigma must be clearly and unambiguously understood throughout the organization.
This doesn’t happen by accident, it is the result of careful planning
and execution.
Responsibility for the communication process should be determined and
assigned at the outset. The communication process owner will be held
accountable by the Executive Six Sigma Council for the development and
oversight of the communication plan. This may include impact on the process
owner’s compensation and/or bonus, or other financial impact. Of course,
the owner will need to put together a team to assist with the efforts.
Development of the communication plan is a subproject, so the communication
process owner will report to the sponsor of the overall Six Sigma deployment
The communication plan must identify the communication needs for each
stakeholder group in the enterprise. Significant stakeholder groups include,
but are not limited to, the following:
. Key customers
. Shareholders or other owners
. Senior leadership
. Middle management
. Six Sigma change agents
. The general employee population
. Suppliers
Metrics are, of course, a vital means of communication. Six Sigma metrics
are discussed in detail elsewhere in this book. Suffice it to say here that Six
Sigma metrics are based on the idea of a balanced scorecard. A balanced
Implementing Six Sigma 33
Preface: The leadership vision for the organization
I. Six Sigma communication plan
II. Six Sigma organizational roles and responsibilities
III. Six Sigma training
IV. Six Sigma project selection, tracking, and management
V. Six Sigma rewards and recognition
VI. Six Sigma compensation and retention
VII. Deploying Six Sigma to the supply chain
Figure 1.9. Chapters in the Six Sigma Deployment Manual.
scorecard is like the instrument panel in the cockpit of an airplane. It displays
information that provides a complete view of the way the organization
appears to its customers and shareholders, as well as a picture of key internal
processes and the rate of improvement and innovation. Balanced scorecards
also provide the means of assuring that Six Sigma projects are addressing
key business issues.
Communicating the Six Sigma message is a multimedia undertaking. The
modern organization has numerous communications technologies at its disposal.
Keep in mind that communication is a two-way affair; be sure to provide
numerous opportunities for upward and lateral as well as downward
communication. Here are some suggestions to accomplish the communications
. All-hands launch event, with suitable pomp and circumstance
. Mandatory sta? meeting agenda item
. House organs (newsletters, magazines, etc.)
. Web site content on Six Sigma (Internet and Intranet)
. Highly visible links to enterprise Six Sigma web site on home page
. Six Sigma updates in annual report
. Stock analyst updates on publicly announced Six Sigma goals
. Intranet discussion forums
. Two-way email communications
. Surveys
. Suggestion boxes
. V|deotape or DVD presentations
. Closed circuit satellite broadcasts by executives, with questions and
. All-hands discussion forums
. Posters
. Logo shirts, gear bags, keychains, co?ee mugs, and other accessories
. Speeches and presentations
. Memoranda
. Recognition events
. Lobby displays
. Letters
The list goes on. Promoting Six Sigma awareness is, in fact, an internal marketing
campaign. A marketing expert, perhaps from your company’s marketing
organization, should be consulted. If your organization is small, a good book
on marketing can provide guidance (e.g., Levinson et al. (1995)).
For each group, the communication process owner should determine the
1. Who is primarily responsible for communication with this group?
2. What are the communication needs for this group? For example, key
customers may need to know how Six Sigma will bene?t them; employees
may need to understand the process for applying for a change agent
position such as Black Belt.
3. What communication tools, techniques and methods will be used? These
include meetings, newsletters, email, one-on-one communications, web
sites, etc.
4. What will be the frequency of communication? Remember, repetition
will usually be necessary to be certain the message is received and understood.
5. Who is accountable for meeting the communication requirement?
6. How will we measure the degree of success? Who will do this?
The requirements and responsibilities can be organized using tables, such as
Table 1.3.
Six Sigma organizational roles and responsibilities
Six Sigma is the primary enterprise strategy for process improvement. To
make this strategy a success it is necessary not only to implement Six Sigma,
but also to institutionalize it as a way of doing business. It is not enough to
train a few individuals to act as champions for Six Sigma. To the contrary, such
a plan virtually guarantees failure by placing the Six Sigma activities somewhere
other than the mainstream. After all, isn’t process improvement an ongoing
part of the business?
Leadership’s primary role is to create a clear vision for Six Sigma success and
to communicate their vision clearly, consistently, and repeatedly throughout
the organization. In other words, leadership must lead the effort.
The primary responsibility of leadership is to assure that Six Sigma goals,
objectives, and progress are properly aligned with those of the enterprise as a
whole. This is done by modifying the organization in such a way that personnel
naturally pursue Six Sigma as part of their normal routine. This requires the
creation of new positions and departments, and modifying the reward, recogni-
Implementing Six Sigma 35
Table 1.3. Six Sigma communications plan and requirements matrix.
Group Method Frequency Accountability
Senior Leadership
Program strategy,
goals and highlevel
program plan
 Senior sta?
 At least
 Start of
 Six Sigma Director
performance to
program plan
 Senior sta?
meetings  At least
monthly  Six Sigma Director
Middle Management
Program strategy, goals
and management-level
program plan
 Regular ?ow
down of
upper level
sta? meeting
 At least
monthly for
piece every 2
weeks during
rollout, as
 Prior to 1st
wave of Six
 Senior Leadership
for sta? meeting
?ow down
via core team for
Etc. for customers, owners, stock analysts, change agents, bargaining unit, exempt
employees, suppliers, or other stakeholder group.
tion, incentive, and compensation systems for other positions. Leadership must
decide such key issues as:
. How will leadership organize to support Six Sigma? (E.g., Executive Six
Sigma Council, designation of an executive Six Sigma champion, creation
of Director of Six Sigma, where will the new Six Sigma function report?
. At what rate do we wish to make the transition from a traditional to a Six
Sigma enterprise?
. What will be our resource commitment to Six Sigma?
. What new positions will be created? To whom will they report?
. Will Six Sigma be a centralized or a decentralized function?
. What level of ROI validation will we require?
. Howwill Six Sigma be integrated with other process excellence initiatives,
such as Lean?
. Will we create a cross-functional core team to facilitate deployment? Who
will be on the team? To whom will they be accountable?
. How will leadership monitor the success of Six Sigma?
. How will executive support of Six Sigma be assessed?
Although each organization will develop its own unique approach to Six
Sigma, it is helpful to know how successful companies have achieved their
success. Most importantly, successful Six Sigma deployment is always a topdown
affair. I know of no case where Six Sigma has had a major impact on
overall enterprise performance that was not fully embraced and actively
lead by top management. Isolated efforts at division or department levels
are doomed from the outset. Like flower gardens in a desert, they may flourish
and produce a few beautiful results for a time, but sustaining the results
requires immense effort by local heroes in constant conflict with the mainstream
culture, placing themselves at risk. Sooner or later, the desert will
reclaim the garden. Six Sigma shouldn’t require heroic effortLthere are
never enough heroes to go around. Once top management has accepted its
leadership responsibility the organizational transformation process can
A key decision is whether Black Belts will report to a central Six Sigma
organization or to managers located elsewhere in the organization. The
experience of most successful Six Sigma enterprises is that centralized reporting
is best. Internal studies by one company that experimented with both
types of reporting revealed the results shown in Table 1.4. The major reason
for problems with the decentralized approach was disengaging people from
Implementing Six Sigma 37
routine work and firefighting. Six Sigma is devoted to change, and it seems
change tends to take a back seat to current problems. To be sure, the Black
Belt possesses a skill set that can be very useful in putting out fires. Also,
Black Belts tend to be people who excel at what they do. This combination
makes it difficult to resist the urge to pull the Black Belt off of his or her projects
‘‘just for a while.’’ In fact, some organizations have trouble getting the
Black Belt out of their current department and into the central organization.
In one case the CEO intervened personally on behalf of the Black Belts to
break them loose. Such stories are testimony to the difficulties encountered
in making drastic cultural changes.
The transformation process involves new roles and responsibilities on the
part of many individuals in the organization. In addition, new change agent
positions must be created. Table 1.5 lists some typical roles and responsibilities.
Obviously, the impact on budgets, routines, existing systems, etc. is substantial.
Six Sigma is not for the faint-hearted. It isn’t hard to see why it takes a
number of years for Six Sigma to become ‘‘mature.’’ The payoff, however,
makes the effort worthwhile. Half-hearted commitments take nearly as much
effort and produce negligible results, or even negative impacts.
Selecting the ‘‘Belts’’
Past improvement initiatives, such as TQM, shared much in common with
Six Sigma. TQM also had management champions, improvement projects,
sponsors, etc. One of the main differences in the Six Sigma infrastructure is
the creation of more formally defined change agent positions. Some observers
criticize this practice as creating corps of ‘‘elites,’’ especially Black Belts and
Master Black Belts. However, I believe this criticism is invalid. Let’s examine
Table 1.4. Black Belt certi?cation versus reporting arrangement.
Where Black Belt Reported
Black Belts Successfully
Local organization 40%
Centralized Six Sigma
Implementing Six Sigma 39
Table 1.5. Six Sigma roles and responsibilities.
Entity Roles Responsibilities
Six Sigma
Strategic leadership . Ensures Six Sigma goals are linked to
enterprise goals
. Develops new policies as required
. Aligns process excellence e?orts across
the organization
. Suggests high-impact projects
. Approves project selection strategy
Assures progress . Provides resources
. Tracks and controls progress toward
. Reviews improvement teams’ results
(BB, GB, Lean, Supply Chain, other)
. Reviews e?ectiveness of Six Sigma
deployment: systems, processes,
infrastructure, etc.
Cultural transformation . Communicates vision
. Removes formal and informal barriers
. Commissions modi?cation of
compensation, incentive, reward and
recognition systems
Six Sigma
Manages Six Sigma
infrastructure and
. Six Sigma champion for ACME
. Develops Enterprise Six Sigma
. Owns the Six Sigma project selection
and prioritization process for ACME
. Assures Six Sigma strategies and
projects are linked through quality
function deployment to business plans
. Achieves defect reduction and cost takeout
targets through Six Sigma activities
. Member of Executive Six Sigma Council
. Leads and evaluates the performance of
Black Belts andMaster Black Belts
. Communicates Six Sigma progress with
customers, suppliers and the enterprise
Entity Roles Responsibilities
. Champions Six Sigma reward and
recognition, as appropriate
Six Sigma
Certi?es Black Belts
Board representatives
include Master Black
Belts and key Six Sigma
. Works with local units to customize
Black Belt and Green Belt requirements
to ?t business needs
. Develops and implements systems for
certifying Black Belts and Green Belts
. Certi?es Black Belts
Six Sigma
Core Team
Cross-functional Six
Sigma team
Part-time change agent
. Provides input into policies and
procedures for successful
implementation of Six Sigma across
. Facilitates Six Sigma activities such as
training, special recognition events,
Black Belt testing, etc.
Master Black
Enterprise Six Sigma
Permanent full-time
change agent
Certi?ed Black Belt with
additional specialized
skills or experience
especially useful in
deployment of Six Sigma
across the enterprise
.Highly pro?cient in using Six Sigma
methodology to achieve tangible
business results
. Technical expert beyond Black Belt level
on one or more aspects of process
improvement (e.g., advanced statistical
analysis, project management,
communications, program
administration, teaching, project
. Identi?es high-leverage opportunities for
applying the Six Sigma approach across
the enterprise
. Basic Black Belt training
. Green Belt training
. Coach/Mentor Black Belts
. Participates on ACME Six Sigma
Certi?cation Board to certify Black Belts
and Green Belts
Table 1.5. Six Sigma roles and responsibilities (continued)
Implementing Six Sigma 41
Entity Roles Responsibilities
Black Belt Six Sigma technical expert
Temporary, full-time
change agent (will return
to other duties after
completing a two to three
year tour of duty as a
Black Belt)
. Leads business process improvement
projects where Six Sigma approach is
. Successfully completes high-impact
projects that result in tangible bene?ts
to the enterprise
. Demonstrated mastery of Black Belt
body of knowledge
. Demonstrated pro?ciency at achieving
results through the application of the Six
Sigma approach
. Internal Process Improvement
Consultant for functional areas
. Coach/Mentor Green Belts
.Recommends Green Belts for
Green Belt Six Sigma project
Six Sigma project leader
Part-time Six Sigma
change agent. Continues
to perform normal duties
while participating on Six
Sigma project teams
Six Sigma champion in
local area
. Demonstrated mastery of Green Belt
body of knowledge
. Demonstrated pro?ciency at achieving
results through the application of the Six
Sigma approach
. Recommends Six Sigma projects
. Participates on Six Sigma project teams
. Leads Six Sigma teams in local
improvement projects
. Works closely with other continuous
improvement leaders to apply formal
data analysis approaches to projects
. Teaches local teams, shares knowledge
of Six Sigma
. Successful completion of at least one Six
Sigma project every 12 months to
maintain their Green Belt certi?cation
Six Sigma
Primary ACME vehicle
for achieving Six Sigma
. Completes chartered Six Sigma projects
that deliver tangible results
. Identi?es Six Sigma project candidates
Table 1.5. Six Sigma roles and responsibilities (continued)
Entity Roles Responsibilities
Leaders and
Champions for Six Sigma . Ensures ?ow-down and follow-through
on goals and strategies within their
. Plans improvement projects
. Charters or champions chartering
. Identi?es teams or individuals required
to facilitate Six Sigma deployment
. Integrates Six Sigma with performance
appraisal process by identifying
measurable Six Sigma goals/objectives/
. Identi?es, sponsors and directs Six
Sigma projects
. Holds regular project reviews in
accordance with project charters
. Includes Six Sigma requirements in
expense and capital budgets
. Identi?es and removes organizational
and cultural barriers to Six Sigma
. Rewards and recognizes team and
individual accomplishments (formally
and informally)
. Communicates leadership vision
. Monitors and reports Six Sigma progress
. Validates Six Sigma project results
. Nominates highly quali?ed Black Belt
and/or Green Belt candidates
Charter and support Six
Sigma project teams
. Sponsor is ultimately responsible for the
success of sponsored projects
. Actively participates in projects
. Assures adequate resources are provided
for project
Table 1.5. Six Sigma roles and responsibilities (continued)
the commonly proposed alternatives to creating a relatively small group of
highly trained technical experts:
. Train the masses. This is the ‘‘quality circles’’ approach, where people in
the lowest level of the organizational hierarchy are trained in the use of
basic tools and set loose to solve problems without explicit direction
Implementing Six Sigma 43
Entity Roles Responsibilities
. Personal review of progress
. Identi?es and overcomes barriers and
. Evaluates and accepts deliverable
Manages Six Sigma
resources dedicated to a
particular area (e.g., teams
of Black Belts on special
Champions Six Sigma
Black Belt team
. Provides day-to-day direction for Six
Sigma project Black Belt and team
. Provides local administrative support,
facilities, and materials
. Conducts periodic reviews of projects
. Provides input on Black Belt
performance appraisals
. Makes/implements decisions based on
recommendations of Six Sigma Black
Six Sigma
Team Member
Learns and applies Six
Sigma tools to projects
. Actively participates in team tasks
. Communicates well with other team
. Demonstrates basic improvement tool
. Accepts and executes assignments as
determined by team
Table 1.5. Six Sigma roles and responsibilities (continued)
from leadership. When this approach was actually tried in America in the
1970s the results were disappointing. The originators of the quality circles
idea, the Japanese, reported considerably greater success with the
approach. This was no doubt due to the fact that Japanese circles were
integrated into decades old company-wide process improvement activities,
while American ?rms typically implemented circles by themselves.
Indeed, when Six Sigma deployments reach a high level of maturity,
more extensive training is often successful.
. Train the managers. This involves training senior and middle management
in change agent skills. This isn’t a bad idea of itself. However, if the basic
structure of the organization doesn’t change, there is no clear way to
apply the newly acquired skills. Training in and of itself does nothing to
change an organization’s environment. Historically, trained managers
return to pretty much the same job. As time goes by their skills atrophy
and their self-con?dence wanes. If opportunities to apply their knowledge
do arise, they often fail to recognize it or, if they do recognize it, fail to correctly
apply the approach. This is natural for a person trying to do something
di?erent for the ?rst time. The full-time change agents in Six Sigma
learn by doing. By the end of their tenure, they can con?dently apply Six
Sigma methodology to a wide variety of situations.
. Use the experts in other areas. The tools of Six Sigma are not new. In fact,
Industrial Statisticians, ASQ Certi?ed Quality Engineers, Certi?ed
Reliability Engineers, Certi?ed Quality Technicians, Systems Engineers,
Industrial Engineers, Manufacturing Engineers and other specialists
already possess a respectable level of expertise in many Six Sigma tools.
Some have a level of mastery in some areas that exceeds that of Black
Belts. However, being a successful change agent involves a great deal
more than mastery of technical tools. Black Belts, Green Belts, and
Master Black Belts learn tools and techniques in the context of following
the DMAIC approach to drive organizational change. This is very di?erent
than using the same techniques in routine daily work. Quality analysts,
for example, generally work in the quality department as
permanent, full-time employees. They report to a single boss and have
well-de?ned areas of responsibility. Black Belts, in contrast, go out and
seek projects rather than work on anything routine. They report to many
di?erent people, who use di?erent criteria to evaluate the Black Belt’s performance.
They are accountable for delivering measurable, bottom-line
results. Obviously, the type of person who is good at one job may not be
suitable for the other.
. Create permanent change agent positions.Another option to the Black Belt
position is to make the job permanent. After all, why not make maximum
use of the training by keeping the person in the Black Belt job inde?nitely?
Furthermore, as Black Belts gain experience they become more pro?cient
at completing projects. There are, however, arguments against this
approach. Having temporary Black Belts allows more people to go through
the position, thus increasing the number of people in management with
Black Belt experience. Since Black Belts work on projects that impact
many di?erent areas of the enterprise, they have a broad, process-oriented
perspective that is extremely valuable in top management positions. The
continuous in?ux of new blood into Black Belt and Green Belt positions
keeps the thinking fresh and prevents the ‘‘them-versus-us’’ mentality that
often develops within functional units.NewBlack Belts have di?erent networks
of contacts throughout the organization, which leads to projects in
areas that might otherwise be missed. Permanent Black Belts would almost
certainly be more heavily in?uenced by their full-time boss than temporary
Black Belts, thus leading to a more provincial focus.
There are usually many more Black Belt candidates than there are positions.
Thus, although there are minimum requirements, those selected generally
exceed the minimums by a considerable degree. The process for selecting Black
Belts should be clearly defined. This assures consistency and minimizes the
possibility of bias and favoritism.
The next question is, what’s important to the success of a Black Belt? I
worked with a group of consultants and Master Black Belts to answer this question.
We came up with a list of seven success factors, then used Expert Choice
2000 software* to calculate relative importance weights for each category. The
results are shown in Figure 1.10.
The weights are, of course, subjective and only approximate. You may feel
free to modify them if you feel strongly that they’re incorrect. Better yet, you
may want to identify your own set of criteria and weights. The important thing
is to determine the criteria and then develop a method of evaluating candidates
on each criterion. The sum of the candidate’s criterion score times the criterion
weight will give you an overall numerical assessment that can be useful in sorting
out those candidates with high potential from those less likely to succeed
as Black Belts. Of course, the numerical assessment is not the only input into
the selection decision, but it is a very useful one.
Implementing Six Sigma 45
*Expert Choice 2000 is a software package that converts pairwise comparisons into relative weights using the Analytic
Hierarchy Process. The AHP is described elsewhere in this book. See for details.
You may be surprised to see the low weight given to math skills. The rationale
is that Black Belts will receive 200 hours of training, much of it focused
on the practical application of statistical techniques using computer software
and requiring very little actual mathematics. Software automates the analysis,
making math skills less necessary. The mathematical theory underlying a technique
is not discussed beyond the level necessary to help the Black Belt properly
apply the tool. Black Belts who need help with a particular tool have
access to Master Black Belts, other Black Belts, consultants, professors, and
a wealth of other resources. Most statistical techniques used in Six Sigma are
relatively straightforward and often graphical; spotting obvious errors is
usually not too difficult for trained Black Belts. Projects seldom fail due to a
lack of mathematical expertise. In contrast, the Black Belt will often have to
rely on their own abilities to deal with the obstacles to change they will inevitably
encounter. Failure to overcome the obstacle will often spell failure of
the entire project.
Figure 1.11 provides an overview of a process for the selection of Black Belt
Figure 1.10. Black Belt success factors and importance weights.
Implementing Six Sigma 47
Minimum Criteria
Education—Bachelors Degree, minimum.
Work Experience—At least 3years of business, technical, or managerial experience
plus technical application of education and experience as a member or leader of
functional and cross-functional project teams.
Technical Capability—Project management experience is highly desired.
Understanding of basic principles of process management. Basic college algebra
proficiency as demonstrated by exam.
Computer Proficiency—MS Office Software Suite.
Communication—Demonstrate excellent oral and written communication skills.
Team Skills—Ability to conduct meetings, facilitate small groups and successfully
resolve conflicts. Ability to mentor and motivate people.
Final Candidate Selection
To ensure that the Black Belts will be able to address enterprise-wide issues and
processes, the Director of Six Sigma and the Executive Six Sigma Council will determine
the number of Black Belts to be trained in each functional area, division, department,
etc. Black Belt candidates are ranked using a system of points assigned during
the screening process. Rank-ordered lists of Black Belt candidates are prepared for
designated areas and presented to the senior management of the area for final selection.
Area management nominates candidates from their list in numbers sufficient to fill
the spaces allocated by the Director of Six Sigma and the Executive Six Sigma
Commitment to Black Belt Assignment
Selected candidates are required to attend 200 hours of Black Belt training (see
Chapter 4 for the training content). Within one year of completing training, the
Black Belt candidate is required to become certified by passing a written examination
and successfully completing at least two major projects. (See the Appendix for
detailed Black Belt certification process information.) The Black Belt is assigned to
Six Sigma full time as a Black Belt for a minimum period of 2 full years, measured
from the time he or she is certified as a Black Belt.
Reintegration of Black Belts into the Organization
Black Belts are employed in the Black Belt role for two or three years. After that time
they leave the Six Sigma organization and return to other duties. Accomplishing this
transition is the joint responsibility of the Black Belt, the Director of Six Sigma, and the
management of the Black Belt’s former department. Collectively this group comprises
the ‘‘Transition Team’’ for the Black Belt. However, senior leadership must accept
ultimate responsibility for assuring that Black Belts are not ‘‘homeless’’ after completing
their Black Belt tour of duty.
The Director of Six Sigma will inform the Black Belt at least six months prior to the
scheduled return. The Black Belt should maintain contact with their ‘‘home’’ organization
during his tenure in Six Sigma. If it appears that there will be a suitable position
available at approximately the time the Black Belt is scheduled to return, arrangements
should be made to complete or hand-off the Black Belt’s Six Sigma projects in
preparation for his return. If no suitable openings will be available, the Transition
Team needs to develop alternative plans. Alternatives might include extending the
Black Belt’s term of service in Six Sigma, looking for openings in other areas, or
making temporary arrangements.
Figure 1.11. Black Belt candidate selection process and criteria.
Green Belts are change agents who work part time on process improvement.
The bulk of the Green Belt’s time is spent performing their normal
work duties. Although most experts (including me) advocate that the Green
Belt spend 10% to 20% of their time on projects, the time a typical Green
Belt spends on projects in a given year is more like 2% to 5%. A Green Belt
will usually complete one or two major projects per year. Also, unlike Black
Belt projects, Green Belt projects may address processes that are not crossfunctional.
Few Green Belt projects cover enterprise-wide processes.
However, since there are usually more Green Belts than Black Belts by a factor
of 2 to 5, these Green Belt projects have a tremendous impact on
the enterprise. Also, it is common to have a Black Belt coordinating a ‘‘portfolio’’
of ‘‘Green Belt projects’’* that, taken together, cover a cross-functional
Figure 1.12 provides an overview of a process for the selection of Green Belt
Master Black Belts are recruited from the ranks of Black Belts. The process
is usually less formal and less well defined than that for Black Belts or
Green Belts and there is a great deal of variability between companies.
Master Black Belt candidates usually make their interest known to Six
Sigma leadership. Leadership selects candidates based on the needs of the
enterprise and Six Sigma’s role in meeting those needs. For example, in the
early stages of deployment Master Black Belt candidates with excellent organizational
skills and the ability to communicate the leadership’s Six Sigma
vision may be preferred. Intermediate deployments might favor candidates
who excel at project selection and Black Belt coaching. Mature Six Sigma programs
might look for Master Black Belts with training ability and advanced
statistical know-how. Master Black Belts often have advanced technical
degrees and extensive Black Belt experience. Many organizations provide
Master Black Belts with additional training. Certification requirements for
Master Black Belts varies with the organization. Many organizations do not
certify Master Black Belts.
*Personally, I would prefer that the term ‘‘Six Sigma project’’ be used instead of Black Belt project or Green Belt project.
However, I bow to common usage in this book.
Integrating Six Sigma and related initiatives
At any given time most companies have numerous activities underway to
improve their operations. For example, the company might be pursuing one or
more of the following:
. Lean manufacturing
. Lean service
. Continuous improvement
. Kaizen
. Business process reengineering
. Theory of constraints
. Variation reduction
The list can be extended indefinitely. Six Sigma can’t simply be thrown into
the mix without causing tremendous confusion. People will find themselves in
conflict with one another over jurisdiction, resources, and authority.
Leadership must give careful thought as to how the various overlapping activities
can best be organized to optimize their impact on performance. An
‘‘umbrella concept’’ often provides the needed guidance to successfully integrate
the different but related efforts. One concept that I’ve found to be particularly
useful is that of ‘‘Process Excellence’’ (PE).
Implementing Six Sigma 49
Minimum Criteria
Education—High school or equivalent.
Work Experience—At least 3years of business, technical, or managerial experience.
Technical Capability—High school algebra proficiency as demonstrated by a passing
grade in an algebra course.
Computer Proficiency—Word processing, presentation and spreadsheet software.
Team Skills—Willingness to lead meetings, facilitate small groups and successfully
resolve conflicts. Ability to mentor and motivate people.
Final Candidate Selection
Based on the organizational need for Green Belts, as determined by the Director of
Six Sigma and the Executive Six Sigma Council, Green Belt training allotments are
provided to Master Black Belts, Black Belts and/or General Managers. Green Belt
candidacy requires the consent of the candidate’s management.
Each Green Belt candidate selected will be required to complete a 40 hour Green Belt
training course, and to lead at least one successful Six Sigma project every 12
months, or participate on at least two successful Six Sigma projects every 12 months.
Green Belt certification is accomplished as described in the Appendix.
Figure 1.12. Green Belt candidate selection process and criteria.
Organizations are typically designed along functional lines. Functions,
such as engineering, marketing, accounting, manufacturing, and so on are
assigned responsibility for certain tasks. The functions tend to correspond closely
to university degree programs. Persons with higher education in a functional
area specialize in the work assigned to the function. General
management and finance allocate resources to each function based on the
needs of the enterprise.
If the enterprise is to be successful the ‘‘needs of the enterprise’’ must be
based on the needs of its customers. However, customers typically obtain
value not from organizational functions but from products or services that are
created by the cooperative efforts and resources of many different functional
areas. Most customers couldn’t care less about how the enterprise creates the
values they are purchasing.* A similar discussion applies to owners and shareholders.
In fact, there is a substantial body of opinion among management
experts that focusing internally on functional concerns can be detrimental to
the enterprise as a whole. An alternative is to focus on the process or value
stream that creates and delivers value.
A process focus means that stakeholder values are determined and activities
are classified as either relating to the creation of the final value (valueadded
activity) or not (non-value-added activity). Processes are evaluated on
how effectively and efficiently they create value. Effectiveness is defined as
delivering what the customer requires, or exceeding the requirements; it
encompasses quality, price, delivery, timeliness and everything else that goes
into perceived value. Efficiency is defined as being effective using a minimum
of resources; more of an owner’s perspective. Excellent processes are those
that are both effective and efficient.
PE is the set of activities specifically designed to create excellent processes.
PE is change-oriented and cross-functional. It includes Six Sigma, all of the
initiatives listed earlier, and many more as well. By creating a top-level position
for PE, leadership assigns clear responsibility for this important work.
The PE leader, usually a Vice President, leads a Process Excellence
Leadership Team (PELT) which includes functional leaders as well as fulltime
PE personnel such as the Director of Six Sigma. The VP of PE isn’t
responsible for particular processes, but she has the authority to identify key
processes and nominate owners for approval by the CEO or the PELT.
Examples of processes include:
*There are exceptions to this.Many large customers, such as the Department of Defense or automobile or aircraft manufacturers,
take a very active interest in the internal operations of their key suppliers.
. Order ful?llment
. Coordinating improvement activities of Six Sigma, Lean, etc.
. Customer contact with the company
. Handling public relations emergencies
. Getting ideas for improvement projects
. Matching improvement projects with customer needs
. Innovating
. Communicating with the outside world
. Communicating internally
. Identifying talent
. Handling customer problems
. Avoiding legal disputes
In other words, the VP of PE has a ‘‘meta-process’’ responsibility. She is
responsible for the process of identifying and improving processes. PE activities
such as Six Sigma, Lean, etc. provide PE with resources to direct toward the
organization’s goal of developing internal processes that give it a competitive
advantage in securing the best employees, delivering superior customer value,
and earning a premium return for its investors.
Deployment to the supply chain
In the early part of the twentieth century Henry Ford pursued a great vision
by building the Ford River Rouge Complex. By 1927 the Rouge was handling
all production of Ford automobiles. It was truly a marvel. The Rouge was the
largest single manufacturing complex in the United States, with peak employment
of about 120,000. Here Henry Ford achieved self-sufficiency and vertical
integration in automobile production, a continuous work flow from iron ore
and other raw materials to finished automobiles. The complex included dock
facilities, blast furnaces, open-hearth steel mills, foundries, a rolling mill, metal
stamping facilities, an engine plant, a glass manufacturing building, a tire
plant, and its own power house supplying steam and electricity.
On June 2, 1978, the Rouge was listed a National Historic Landmark. From
state-of-the-art wonder to historical curiosity in just fifty years.
A related historical artifact is the idea that a firm can produce quality products
or services by themselves. This may’ve been the case in the heyday of the
Rouge, when the entire ‘‘supply chain’’ was a single, vertically integrated behemoth
entity, but it is certainly no longer true. In today’s world fully 50^80% of
the cost of a manufactured product is in purchased parts and materials. When
the customer forks over her good money for your product, she doesn’t differentiate
between you and your suppliers.
Implementing Six Sigma 51
You say you’re not in manufacturing? The situation is the same for you. Say,
for example, your product is personal finance software. Your customer runs
your software on a computer you didn’t design with an operating system you
have no control over. They’re using your software to access their account at
their financial institution to complete a tax return, which they’ll file electronically
with the IRS. When your customers click the icon to run your product,
they consider all of these intermediaries to be part of the value they are paying
to receive.
The service industry is no different. Let’s say you are a discount brokerage
company. Your customers want to be able to use your service to buy common
stocks, fixed income instruments, derivatives, etc. They also want debit cards,
check writing, bill paying, pension plans, and a variety of other services. Oh,
and don’t forget financial advice, investment portfolio analysis, and annuities.
When your customers put their money into their account at your firm, they
expect you to be responsible for making all of the ‘‘third parties’’ work together
In short, you’ll never reach Six Sigma quality levels with three sigma suppliers.
Your primary mission in the supplier Six Sigma activity is to obtain Six Sigma
supplier quality with minimal costs. In pursuit of this mission you will initiate
a number of Six Sigma projects that involve suppliers. The organization responsible
for supply chain management (SCM) will take the lead in developing the
supplier Six Sigma program. Leadership includes preparing the Supplier Six
Sigma Deployment Plan. The plan should include the following:
. Policies on supplier Six Sigma
. Goals and deliverables of the supplier Six Sigma program
. Supplier communication plan
. Timetable for deployment, including phases (e.g., accelerated deployment
to most critical suppliers)
. Procedures de?ning supplier contact protocols, supplier project charter,
supplier project reporting and tracking, etc.
. Training requirements and timeline
. Methods of assessing supplier Six Sigma e?ectiveness
. Integration of the supplier Six Sigma program and in-house activities
SCM receives guidance from the Executive Six Sigma Council and the Six
Sigma organization. The Six Sigma organization often provides expertise and
other resources to the supplier Six Sigma effort.
SCM should sponsor or co-sponsor supplier Six Sigma projects. In some
cases SCM will lead the projects, often with supplier personnel taking a coleadership
role. In others they will assist Black Belts or Green Belts working
on other projects that involve suppliers. Full SCM sponsorship is usually
required when the project’s primary focus is on the supplier’s product or process.
For example, a Six Sigma project chartered to reduce the number of
late deliveries of a key product. Projects involving suppliers, but not focused
on them, can be co-sponsored by SCM. For example, a project involving the
redesign of an order fulfillment process that requires minor changes to the
supplier’s web ordering form. SCM assistance can take a number of different
forms, e.g.:
. Acting as a liaison between the internal team members and suppliers
. Negotiating funding and budget authority for supplier Six Sigma projects
. Estimating and reporting supplier project savings
. Renegotiating contract terms
. Resolving con?icts
. De?ning responsibility for action items
. Scheduling supplier visits
. De?ning procedures for handling of proprietary supplier information
. Responding to supplier requests for assistance with Six Sigma
In addition to SCM, other elements within your organization play important
supporting roles. Usually Black Belts will come from the Six Sigma organization,
although some larger enterprises assign a team of Black Belts to work on
SCM projects full time. Green Belts often come from organizations sponsoring
supplier-related projects. Team members are assigned from various areas, as
with any Six Sigma project.
Never forget that the supplier’s processes are owned and controlled by the
supplier, not by you. As the customer you certainly have the final say in the
requirements, but ultimate responsibility for the process itself should remain
with the supplier. To do otherwise may have legal ramifications, such as liability
and warranty implications. Besides these issues is the simple human tendency
of caring less when someone else is responsible. Six Sigma teams also need to
be careful about making it clear that onlySCMhas the authority to make official
requests for change. It can be embarrassing if a Black Belt makes a suggestion
that the supplier believes to be a formal requirement to change. SCM may
receive a new bid, price change, complaint letter, etc. from the supplier over
such misunderstandings. Supplier relationships are often quite fragile and
‘‘Handle with care’’ is a good motto for the entire Six Sigma team to follow.
Implementing Six Sigma 53
In addition to accepting responsibility for their processes, suppliers must
often take the lead role in Six Sigma teams operating in supplier facilities.
Supplier leadership must support Six Sigma efforts within their organizations.
Suppliers must agree to commit the resources necessary to successfully complete
projects, including personnel and funding.
Experienced Certified Black Belts and Master Black Belts are in great
demand throughout the manufacturing and services sectors.* Small wonder.
Here are people who have proven that they can effect meaningful change in a
complex environment. Since organizations exist in a competitive world, steps
must be taken to protect the investment in these skilled change agents, or they
will be lured away by other organizations, perhaps even competitors. The most
common (and effective) actions involve compensation and other financial
incentives, such as:
. Bonuses
. Stock options
. Results sharing
. Payment of dues to professional societies
. Pay increases
There are also numerous non-financial and quasi-financial rewards. For
example, Black Belts reentering the workforce after their tour of duty often
enter positions that pay significantly higher than the ones they left when becoming
Black Belts. In fact, in some companies the Black Belt position is viewed as
a step on the fast track to upper management positions. Also, change is ‘‘news’’
and it is only natural that the names of Master Black Belts and Black Belts
involved in major change initiatives receive considerable publicity on company
web sites as well as in newsletters, recognition events, project fairs, etc. Even if
they don’t receive formal recognition, Six Sigma projects often generate a great
deal of internal excitement and discussion. The successful Black Belt usually
finds that his work has earned him a reputation that makes him a hot commodity
when it’s time to end his Black Belt career.
There are, of course, innumerable complexities and details to be decided and
worked out. Usually these issues are worked out by a team of individuals with
members from Human Resources, the Six Sigma Core Team, and other areas
of the organization. The team will address such issues as:
*Although Green Belts are also highly trained change agents, they are not full-time change agents and we will not discuss their
compensation here.
. What pay grade is to be assigned to the Black Belt and Master Black Belt
. Should the pay grade be determined by the pay grade of the candidate’s job
prior to becoming a Black Belt?
. Should the Black Belt pay grade be guaranteed when the Black Belt leaves
the Black Belt position to return to the organization?
. How do we determine eligibility for the various rewards? For example, are
there key events such as acceptance as a Black Belt candidate, completion
of training, completion of ?rst project, successful certi?cation, etc.?
. What about Black Belts who were certi?ed by other organizations or third
. Do we provide bene?ts to Green Belts as well? If so, what and how?
. Who will administer the bene?ts package?
The plan will be of great interest to Black Belt candidates. If not done properly,
the organization will find it difficult to recruit the best people.
Change Agent Compensation and Retention 55
^ ^ ^
Six Sigma Goals and Metrics
The choice of what to measure is crucial to the success of the organization.
Improperly chosen metrics lead to suboptimal behavior and can lead people
away from the organization’s goals instead of towards them. Joiner (1994) suggests
three systemwide measures of performance: overall customer satisfaction,
total cycle time, and first-pass quality. An effective metric for quantifying firstpass
quality is total cost of poor quality (later in this chapter). Once chosen,
the metrics must be communicated to the members of the organization. To
be useful, the employee must be able to influence the metric through his
performance, and it must be clear precisely how the employee’s performance
influences the metric.
Rose (1995) lists the following attributes of good metrics:
. They are customer centered and focused on indicators that provide value
to customers, such as product quality, service dependability, and timeliness
of delivery, or are associated with internal work processes that
address system cost reduction, waste reduction, coordination and team
work, innovation, and customer satisfaction.
. They measure performance across time, which shows trends rather than
. They provide direct information at the level at which they are applied.
No further processing or analysis is required to determine meaning.
. They are linked with the organization’s mission, strategies, and actions.
They contribute to organizational direction and control.
. They are collaboratively developed by teams of people who provide,
collect, process, and use the data.
Copyright 2003 by The McGraw-Hill Companies, Inc. Click Here for Terms of Use.
Rose also presents a performance measurement model consisting of eight
. Step 1: performance categoryLThis category is the fundamental division
of organizational performance that answers the question: What do
we do? Sources for determining performance categories include an organization’s
strategic vision, core competencies, or mission statement. An
organization will probably identify several performance categories.
These categories de?ne the organization at the level at which it is being
. Step 2: performance goalLThe goal statement is an operational de?nition
of the desired state of the performance category. It provides the target
for the performance category and, therefore, should be expressed in explicit,
action-oriented terms. An initial goal statement might be right on the
mark, so complex that it needs further division of the performance category,
or so narrowly drawn that it needs some combination of performance
categories. It might be necessary to go back and forth between the
performance goals in this step and the performance categories in step 1
before a satisfactory result is found for both.
. Step 3: performance indicatorLThis is the most important step in the
model because this is where progress toward the performance goal is disclosed.
Here irrelevant measures are swept aside if they do not respond
to an organizational goal. This is where the critical measuresLthose that
communicate what is important and set the course toward organizational
successLare established. Each goal will have one or more indicators, and
each indicator must include an operational de?nition that prescribes the
indicator’s intent and makes its role in achieving the performance goal
clear. The scope of the indicator might be viewed di?erently at various
levels in the organization.
. Step 4: elements of measureLThese elements are the basic components
that determine how well the organization meets the performance
indicator. They are the measurement data sourcesLwhat is actually
measuredLand are controlled by the organization. Attempting to measure
things that are beyond organizational control is a futile diversion of
resources and energy because the organization is not in a position to
respond to the information collected. This would be best handled in the
next step.
. Step 5: parametersLThese are the external considerations that in?uence
the elements of measure in some way, such as context, constraint, and
boundary. They are not controlled by the organization but are powerful
factors in determining how the elements of measure will be used. If
measurement data analysis indicates that these external considerations
Attributes of Good Metrics 57
present serious roadblocks for organizational progress, a policy change
action could be generated.
. Step 6: means of measurementLThis step makes sense out of the
preceding pieces. A general, how-to action statement is written that
describes how the elements of measure and their associated parameters
will be applied to determine the achievement level in the performance
indicator. This statement can be brief, but clarifying intent is more important
than the length.
. Step 7: notional metricsLIn this step, conceptual descriptions of possible
metrics resulting from the previous steps are put in writing. This step
allows everyone to agree on the concept of how the information compiled
in the previous steps will be applied to measuring organizational performance.
It provides a basis for validating the process and for subsequently
developing speci?c metrics.
. Step 8: speci?c metricsLIn this ?nal step, an operational de?nition and
a functional description of the metrics to be applied are written. The
de?nition and description describe the data, how they are collected,
how they are used, and, most importantly, what the data mean or how
they a?ect organizational performance. A prototype display of real or
imaginary data and a descriptive scenario that shows what actions
might be taken as a result of the measurement are also made. This last
step is the real test of any metric. It must identify what things need to
be done and disclose conditions in su?cient detail to enable subsequent
improvement actions.
Rose presents an application of his model used by the U.S. Army Materiel
Command, which is shown in Figure 2.1.
The traditional quality model of process capability differed from Six Sigma
in two fundamental respects:
1. It was applied only to manufacturing processes, while Six Sigma is
applied to all important business processes.
2. It stipulated that a ‘‘capable’’ process was one that had a process standard
deviation of no more than one-sixth of the total allowable spread,
where Six Sigma requires the process standard deviation be no more
than one-twelfth of the total allowable spread.
These differences are far more profound than one might realize. By addressing
all business processes Six Sigma not only treats manufacturing as part of a
larger system, it removes the narrow, inward focus of the traditional approach.
Customers care about more than just how well a product is manufactured.
Price, service, financing terms, style, availability, frequency of updates and
enhancements, technical support, and a host of other items are also important.
Also, Six Sigma benefits others besides customers. When operations become
more cost-effective and the product design cycle shortens, owners or investors
benefit too. When employees become more productive their pay can be
increased. Six Sigma’s broad scope means that it provides benefits to all stakeholders
in the organization.
The second point also has implications that are not obvious. Six Sigma is,
basically, a process quality goal, where sigma is a statistical measure of variability
in a process. As such it falls into the category of a process capability
technique. The traditional quality paradigm defined a process as capable if
the process’s natural spread, plus and minus three sigma, was less than the
engineering tolerance. Under the assumption of normality, this three sigma
quality level translates to a process yield of 99.73%. A later refinement considered
the process location as well as its spread and tightened the minimum
acceptance criterion so that the process mean was at least four sigma from
Six Sigma Versus Traditional Three Sigma Performance 59
Figure 2.1. Organizational performance metrics.
From ‘‘A performance measurement model,’’ by Kenneth H. Rose. Quality Progress,
February 1995, p. 65. Reprinted by permission.
the nearest engineering requirement. Six Sigma requires that processes operate
such that the nearest engineering requirement is at least Six Sigma from
the process mean.
Six Sigma also applies to attribute data, such as counts of things gone wrong.
This is accomplished by converting the Six Sigma requirement to equivalent
conformance levels, as illustrated in Figure 2.2.
One of Motorola’s most significant contributions was to change the discussion
of quality from one where quality levels were measured in percent (partsper-
hundred), to a discussion of parts-per-million or even parts-per-billion.
Motorola correctly pointed out that modern technology was so complex that
old ideas about ‘‘acceptable quality levels’’ could no longer be tolerated.
Modern business requires near perfect quality levels.
One puzzling aspect of the ‘‘official’’ Six Sigma literature is that it states that
a process operating at Six Sigma will produce 3.4 parts-per-million (PPM) nonconformances.
However, if a special normal distribution table is consulted
(very few go out to Six Sigma) one finds that the expected non-conformances
are 0.002 PPM (2 parts-per-billion, or PPB). The difference occurs because
Motorola presumes that the process mean can drift 1.5 sigma in either direction.
The area of a normal distribution beyond 4.5 sigma from the mean is indeed 3.4
PPM. Since control charts will easily detect any process shift of this magnitude
Figure 2.2. Sigma levels and equivalent conformance rates.
in a single sample, the 3.4 PPM represents a very conservative upper bound on
the non-conformance rate.
In contrast to Six Sigma quality, the old three sigma quality standard of
99.73% translates to 2,700 PPM failures, even if we assume zero drift. For processes
with a series of steps, the overall yield is the product of the yields of the different
steps. For example, if we had a simple two step process where step #1
had a yield of 80% and step #2 had a yield of 90%, then the overall yield would
be 0:8  0:9 ? 0:72 ? 72%. Note that the overall yield from processes involving
a series of steps is always less than the yield of the step with the lowest
yield. If three sigma quality levels (99.97% yield) are obtained from every step
in a ten step process, the quality level at the end of the process will contain
26,674 defects per million! Considering that the complexity of modern processes
is usually far greater than ten steps, it is easy to see that Six Sigma quality
isn’t optional, it’s required if the organization is to remain viable.
The requirement of extremely high quality is not limited to multiple-stage
manufacturing processes. Consider what three sigma quality would mean if
applied to other processes:
. Virtually no modern computer would function.
. 10,800,000 mishandled healthcare claims each year.
. 18,900 lost U.S. savings bonds every month.
. 54,000 checks lost each night by a single large bank.
. 4,050 invoices sent out incorrectly each month by a modest-sized telecommunications
. 540,000 erroneous call detail records each day from a regional telecommunications
. 270,000,000 (270 million) erroneous credit card transactions each year in
the United States.
With numbers like these, it’s easy to see that the modern world demands
extremely high levels of error free performance. Six Sigma arose in response to
this realization.
Given the magnitude of the difference between Six Sigma and the traditional
three sigma performance levels, the decision to pursue Six Sigma performance
obviously requires a radical change in the way things are done. The organization
that makes this commitment will never be the same. Since the expenditure of
time and resources will be huge, it is crucial that Six Sigma projects and activities
are linked to the organization’s top-level goals. It is even more important
that these be the right goals. An organization that uses Six Sigma to pursue the
wrong goals will just get to the wrong place more quickly. The organization’s
The Balanced Scorecard 61
goals must ultimately come from the constituencies it serves: customers, shareholders
or owners, and employees. Focusing too much on the needs of any one
of these groups can be detrimental to all of them in the long run. For example,
companies that look at shareholder performance as their only significant goal
may lose employees and customers. To use the balanced scorecard senior management
must translate these stakeholder-based goals into metrics. These goals
and metrics are then mapped to a strategy for achieving them. Dashboards are
developed to display the metrics for each constituency or stakeholder. Finally,
Six Sigma is used to either close gaps in critical metrics, or to help develop new
processes, products and services consistent with top management’s strategy.
Balanced scorecards help the organization maintain perspective by providing
a concise display of performance metrics in four areas that correspond roughly
to the major stakeholdersLcustomer, financial, internal processes, and learning
and growth (Kaplan and Norton, 1992). The simultaneous measurement from
different perspectives prevents local suboptimization, the common phenomenon
where performance in one part of the organization is improved at the
expense of performance in another part of the organization. This leads to the
well-known loop where this year we focus on quality, driving up costs. Next year
we focus on costs, hurting cycle time. When we look at cycle time people take
short cuts, hurting quality. And so on. This also happens on a larger scale, where
we alternately focus on employees, customers, or shareholders at the expense of
the stakeholders who are not the current focus. Clearly, such ‘‘firefighting’’
doesn’t make anyone happy.Wetruly need the ‘‘balance’’ in balanced scorecards.
Well-designed dashboards include statistical guidance to aid in interpreting
the metrics. These guidelines most commonly take the form of limits, the calculation
of which are discussed in detail elsewhere in this book. Limits are statistically
calculated guidelines that operationally define when intervention is
needed. Generally, when metrics fall within the limits, the process should be
left alone. However, when a metric falls outside of the limits, it indicates that
something important has changed that requires attention. An exception to
these general rules occurs when a deliberate intervention is made to achieve a
goal. In this case the metric is supposed to respond to the intervention by moving
in a positive direction. The limits will tell leadership if the intervention produced
the desired result. If so, the metric will go beyond the proper control
limit indicating improvement. Once the metric stabilizes at the new and
improved level, the limits should be recalculated so they can detect slippage.
Measuring causes and effects
Dashboard metrics are measurements of the results delivered by complex
processes and systems. These results are, in a sense, ‘‘effects’’ caused by things
taking place within the processes. For example, ‘‘cost per unit’’ might be a
metric on a top-level dashboard. This is, in turn, composed of the cost of materials,
overhead costs, labor, etc. Cost of materials is a ‘‘cause’’ of the cost per
unit. Cost of materials can be further decomposed into, say, cost of raw materials,
cost of purchased sub-assemblies, etc. and so on. At some level we reach a
‘‘root cause,’’ or most basic reason behind an effect. Black Belts and Green
Belts learn numerous tools and techniques to help them identify these root
causes. However, the dashboard is the starting point for the quest.
In Six Sigma work, results are known as ‘‘Ys’’ and root causes are known as
‘‘Xs.’’ Six Sigma’s historical roots are technical and its originators generally
came from engineering and scientific backgrounds. In the mathematics taught
to engineers and scientists equations are used that often express a relationship
in the form:
Y ? f X ? ? ?2:1?
This equation simply means that the value identified by the letter Y is determined
as a function of some other value X. The equation Y = 2X means that if
we know what X is, we can find Y if we multiply X by 2. If X is the temperature
of a solution, then Y might be the time it takes the solution to evaporate.
Equations can become more complicated. For example, Y = f(X1, X2) indicates
that the value Y depends on the value of two different X variables. You should
think of the X in Equation 2.1 as including any number of X variables. There
can be many levels of dashboards encountered between the top-level Y, called
the ‘‘Big Y,’’ and the root cause Xs. In Six Sigma work some special notation
has evolved to identify whether a root cause is being encountered, or an intermediate
result. Intermediate results are sometimes called ‘‘Little Ys.’’
In these equations think of Y as the output of a process and the Xs as inputs.
The process itself is symbolized by the f(). The process can be thought of as a
transfer function that converts inputs into outputs in some way. An analogy is
a recipe. Here’s an example:
Corn Crisp Recipe
12 servings
4 cup yellow stone-ground cornmeal
1 cup boiling water
2 teaspoon salt
3 tablespoons melted butter
Preheat the oven to 4008F. Stir the cornmeal and boiling water together in a
large glass measuring cup. Add the salt and melted butter. Mix well and
The Balanced Scorecard 63
pour onto a cookie sheet. Using a spatula, spread the batter out as thin as you
possibly canLthe thinner the crisper. Bake the cornmeal for half an hour or
until crisp and golden brown. Break into 12 roughly equal pieces.
Here the Big Y is the customer’s overall satisfaction with the finished corn
crisp. Little Ys would include flavor ratings, ‘‘crunchiness’’ rating, smell, freshness,
and other customer-derived metrics that drive the Big Y. Xs that drive the
little Ys might include thinness of the chips, the evenness of the salt, the size of
each chip, the color of the chip, and other measurements on the finished product.
Xs could also be determined at each major step, e.g., actual measurement
of the ingredients, the oven temperature, the thoroughness of stirring, how
much the water cools before it is stirred with the cornmeal, actual bake time,
etc. Xs would also include the oven used, the cookware, utensils, etc.
Finally, the way different cooks follow the recipe is the transfer function or
actual process that converts the ingredients into corn crisps. Numerous sources
of variation (more Xs) can probably be identified by observing the cooks in
action. Clearly, even such a simple process can generate some very interesting
discussions. If you haven’t developed dashboards it might be worthwhile to do
so for the corn crisps as a practice exercise.
Figure 2.3 illustrates how dashboard metrics flow down until eventually
linking with Six Sigma projects.
Information systems
Balanced scorecards begin with the highest level metrics. At any given level,
dashboards will display a relatively small number of metrics. While this allows
the user of the dashboard to focus on key items, it also presents a problem
when the metric goes outside a control limit for reasons other than deliberate
management action. When this happens the question is: Why did this metric
change? Information systems (IS) can help answer this question by providing
‘‘drill down’’ capability. Drill down involves disaggregating dashboard metrics
into their component parts. For example, a cost-per-unit metric can be decomposed
by division, plant, department, shift, worker, week, etc. These components
of the higher-level metric are sometimes already on dashboards at lower
levels of the organization, in which case the answer is provided in advance.
However, if the lower-level dashboard metrics can’t explain the situation,
other exploratory drill downs may be required. On-line analytic processing
(OLAP) cubes often ease the demands on the IS caused by drill down requests.
This raises an important point: in Six Sigma organizations the IS must be
accessed by many more people. The attitude of many IS departments is ‘‘The
data systems belong to us. If you want some data, submit a formal request.’’ In
a Six Sigma organization, this attitude is hopelessly outmoded. The demands on
the IS increase dramatically when Six Sigma is deployed. In addition to the creation
of numerous dashboards, and the associated drill downs and problem
investigations, the Black Belts and Green Belts make frequent use of IS in their
projects. Six Sigma ‘‘show me the data’’ emphasis places more demands on the
IS. In planning for Six Sigma success, companies need to assign a high-level
champion to oversee the adaptation of the IS to the new realities of Six Sigma.
A goal is to make access as easy as possible while maintaining data security and
Although it’s important to be timely, most Six Sigma data analyses don’t
require real-time data access. Data that are a day or a few days old will often suffice.
The IS department may want to provide facilities for off-line data analysis
by Six Sigma team members and Belts. A few high-end workstations capable of
handling large data sets or intensive calculations are also very useful at times,
especially for data mining analyses such as clustering, neural networks, or classification
and decision trees.
Customer perspective
Let’s take a closer look at each of the major perspectives on the balanced
scorecard, starting with the customer. The balanced scorecard requires that
The Balanced Scorecard 65
Figure 2.3. Flowdown of strategies to drivers and projects.
management translate their vague corporate mission (‘‘Acme will be #1 in
providing customer value’’) into specific measures of factors that matter to
customers. The customer scorecard answers the question: ‘‘How do our customers
view us?’’
To answer this, you must ask yourself two related questions: What things do
customers consider when evaluating us? How do we know? While the only
true way to answer these questions is to communicate with real customers, it is
well established that customers in general tend to consider four broad categories
of factors when evaluating an organization:
. Quality. How well do you keep your promises by delivering error free
service or defect free product. Did I receive what I ordered? Was it
undamaged? Are your promised delivery times accurate? Do you honor
your warranty or pay your claims without a hassle?
. Timeliness. How fast is your service? How long does it take to have my
order delivered? Do improvements appear in a timely manner?
. Performance and service.How do your products and services help me? Are
they dependable?
. Value. What is the cost of buying and owning your product or service? Is it
worth it?
The first step in the translation is to determine precisely what customers consider
when evaluating your organization. This can be done by communicating
with customers via one-on-one contacts, focus groups, questionnaires, chat
rooms, forums, etc. Management should see the actual, unvarnished words
used by customers to describe what they think about the company, its products,
and its services. Once management is thoroughly familiar with their target customer,
they need to articulate their customer goals in words meaningful to
them. For example, management might say:
. We will cut the time required to introduce a new product from 9 months
to 3 months.
. We will be the best in the industry for on-time delivery.
. We will intimately involve our customers in the design of our next major
These goals must be operationalized by designating metrics to act as surrogates
for the goals. Think of the goals themselves as latent or hidden constructs.
The objective is to identify observable things directly related to the goals that
can be measured. These are indicators that help guide you towards your goals.
Table 2.1 shows examples of how the goals mentioned above might be operationalized.
These goals are key requirements that employees will be asked to achieve. It is
crucial that they not be set arbitrarily. More will be said about this later in this
chapter (see ‘Setting organizational key requirements’).
Internal process perspective
In the Internal Process section of the balanced scorecard we develop metrics
that help answer the question: What internal processes must we excel at?
Internal process excellence is linked to customer perceived value, but the linkage
is indirect and imperfect. It is often possible to hide internal problems
from customers by throwing resources at problems; for example, increased
inspection and testing. Also, customer perceived value is affected by factors
other than internal processes such as price, competitive offerings, etc.
Similarly, internal operations consume resources so they impact the shareholders.
Here again, the linkage is indirect and imperfect. For example, sometimes
it is in the organization’s strategic interest to drive up costs in order to
meet critical short-term customer demands or to head off competitive moves
in the market. Thus, simply watching the shareholder or customer dashboards
won’t always give leadership a good idea of how well internal processes are
performing. A separate dashboard is needed for this purpose.
This section of the scorecard gives operational managers the internal direction
they need to focus on customer needs. Internal metrics should be chosen
to support the leadership’s customer strategy, plus knowledge of what customers
need from internal operations. Process maps should be created that show
the linkage between suppliers, inputs, process activities, outputs and customers
(SIPOC). SIPOC is a flowcharting technique that helps identify those processes
The Balanced Scorecard 67
Table 2.1. Operationalizing goals.
Goal Candidate Metrics
We will cut the time required to
introduce a new product from 9 months
to 3 months
. Average time to introduce a new product
for most recent month or quarter
. Number of new products introduced in
most recent quarter
We will be the best in the industry for
on-time delivery
. Percentage of on-time deliveries
. Best in industry on-time delivery percentage
divided by our on-time delivery percentage
. Percentage of late deliveries
We will intimately involve our
customers in the design of our next
major product
. Number of customers on design team(s)
. Number of customer suggestions
incorporated in new design
that have the greatest impact on customer satisfaction; it is covered elsewhere in
this book.
Companies need to identify and measure their core competencies. These are
areas where the company must excel. It is the source of their competitive advantage.
Goals in these areas must be ambitious and challenging. This where you
‘‘Wow’’ your customer. Other key areas will pursue goals designed to satisfy
customers, perhaps by maintaining competitive performance levels. Table 2.2
shows how core competencies might drive customer value propositions. The
metrics may be similar for the different companies, but the goals will differ significantly.
For example, Company A would place greater emphasis on the time
required to develop and introduce new services. Companies B and C would
not ignore this aspect of their internal operations, but their goals would be less
ambitious in this area than Company A’s. Company A is the industry benchmark
for innovation.
Of course, it is possible that your competitor will try to leapfrog you in your
core competency, becoming the new benchmark and stealing your customers.
Or you may find that your customer base is dwindling and the market for your
particular competency is decreasing. Leadership must stay on the alert for such
developments and be prepared to react quickly. Most companies will fight to
maintain their position of industry leadership as long as there is an adequate
market. Six Sigma can help in this battle because Six Sigma projects are usually
Table 2.2. Customer value proposition versus core competency.
Internal Process Company A Company B Company C
Innovation X
Operations and
Customer value
Product or service
‘‘X’’ indicates the company’s core competency.
of short duration strategically speaking, and Black Belts offer a resource that can
be redeployed quickly to where they are most needed.
Innovation and learning perspective
In the Innovation and Learning Perspective section of the balanced scorecard
we develop metrics that help answer the question: Can we continue to improve
and create value? Success is a moving target. What worked yesterday may fail
miserably tomorrow. Previous sections of the balanced scorecard have identified
the metrics the leadership considers to be most important for success in
the near future. But the organization must be prepared to meet the new and
changing demands that the more distant future will surely bring. Building shareholder
value is especially dependent on the company’s ability to innovate,
improve, and learn. The intrinsic value of a business is the discounted value of
the cash that can be taken out of the business during its remaining life (Buffett,
1996). Intrinsic value is directly related to a company’s ability to create new products
and processes, to improve operating efficiency, to discover and develop
new markets, and to increase revenues and margins. Companies able to do this
well will throw off more cash over the long term than companies that do it
poorly. The cash generated can be withdrawn by the owners, or reinvested in
the business.
Innovation and learning were the areas addressed by the continuous
improvement (CI) initiatives of the past. Devotees of CI will be happy to
learn that it’s alive and well in the Six Sigma world. However, CI projects
were often local in scope, while most Black Belt Six Sigma projects are crossfunctional.
Many so-called Green Belt projects (Six Sigma projects that don’t
have a dedicated Black Belt on the project team) are reminiscent of the CI
projects in the past. Also, CI tended to focus narrowly on work processes,
while Green Belt projects cover a broader range of business processes, products,
and services. A well-designed Six Sigma program will have a mix of
Green Belt and Black Belt projects addressing a range of enterprise and local
process improvement issues.
Dashboards designed to measure performance in the area of Innovation and
Learning often address three major areas: employee competencies, technology,
and corporate culture. These are operationalized in a wide variety of ways.
One metric is the average rate of improvement in the sigma level of an organizational
unit. Six Sigma attempts to reduce mistakes, errors, and defects by a factor
of 10 every two years, which translates to about 17% per month. This
breakthrough rate of improvement is usually not attained instantly and ametric
of the actual rate is a good candidate for including on the Innovation and
Learning dashboard. The rate of improvement is a measure of the overall matur-
The Balanced Scorecard 69
ity of the Six Sigma initiative. Other Innovation and Learning metric candidates
might include such things as:
. Results of employee feedback
. R&D cycle time
. Closure of gaps identi?ed in the training needs audit
Financial perspective
Obsession with financial metrics has been the undoing of many improvement
initiatives. When senior leaders look only at results they miss the fact
that these results come from a complex chain of interacting processes that effectively
and efficiently produce value for customers. Only by providing value
that customers are willing to pay for can an enterprise generate sales, and only
by creating these values at a cost less than their price can it produce profits for
owners. For many companies the consequence of looking only at short-term
financial results has been a long-term decline in business performance. Many
companies have gone out of business altogether.
The result of this unfortunate history is that many critics have advocated the
complete abandonment of the practice of using financial metrics to guide leadership
action. The argument goes something like this: since financial results
are determined by a combination of customer satisfaction and the way the organization
runs its internal operations, if we focus on these factors the financial
performance will follow in due course. This is throwing the baby out with the
bathwater. The flaw in the logic is that it assumes that leaders and managers
know precisely how customer satisfaction and internal operational excellence
lead to financial results. This arrogance is unjustified. Too often we learn in retrospect
that we are focusing on the wrong things and the financial results fail
to materialize. For example, we may busily set about improving the throughput
of a process that already has plenty of excess capacity. All we get from this effort
is more excess capacity. Many Six Sigma improvements don’t result in bottomline
impact because management fails to take the necessary steps such as
reducing excess inventory, downsizing extra personnel, selling off unneeded
equipment, etc. As Toyota’s Taiichi Ohno says:
If , as a result of labor saving, 0.9 of a worker is saved, it means nothing. At
least one person must be saved before a cost reduction results. Therefore,
we must attain worker saving.
Taiichi Ohno
Toyota Production System: Beyond Large-Scale Production
The truth is, it’s very difficult to lay people off and a poor reward for people
who may have participated in creating the improvement. Most managers agree
that this is the worst part of their job. However, simply ignoring the issue isn’t
the best way to deal with it. Plans must be made before starting a project for
adjusting to the consequences of success. If there will be no bottom-line impact
because there are to be no plans to convert the savings into actual reductions
in resource requirements, the project shouldn’t be undertaken in the first
place.Onthe other hand, plans can often be made at the enterprise level for dealing
with the positive results of Six Sigma by such means as hiring moratoriums,
early retirement packages, etc. Better still are plans to increase sales or to grow
the business to absorb the new capacity. This can often be accomplished by
modifying the customer value proposition through more reliable products,
lower prices, faster delivery time, lower cycle times, etc. These enhancements
are made possible as a result of the Six Sigma improvements.
There are other ways to go wrong if financial results are not explicitly monitored.
We may blindly pour resources into improving customer satisfaction as
measured by a faulty or incomplete survey. Or the competition may discover a
new technology that makes ours obsolete. The list of things that can break the
link between internal strategies and financial performance is endless. Financial
performance metrics provide us with the feedback we need to assure that we
haven’t completely missed the boat with our assumptions.
Actual metrics for monitoring financial performance are numerous. The toplevel
dashboard will often include metrics in the areas of improved efficiency
(e.g., cost per unit, asset utilization) or improved effectiveness (e.g., revenue
growth, market share increase, profit per customer).
Unlike traditional measurement systems, which tend to have a control bias,
balanced scorecards are based on strategy. The idea is to realize the leadership
vision using a set of linked strategies. Metrics operationalize these strategies
and create a bond between the activities of the organization and the vision of
the leadership.
Figure 2.4 illustrates these principles for a hypothetical organization. Things
that will actually be measured are shown in rectangles. The dashboard metrics
appear on the left side of the figure. The strategy deployment plan makes it clear
that the metrics are not ends in themselves, they are merely measurements of bigger
items of interest. These unobserved, or latent constructs are shown in ellipses
and are inferred from the metrics. This perspective helps leadership understand
the limitations of metrics, as well as their value. If, for example, all of the metrics
leading to shareholder perceived value are strongly positive, but surveys of the
shareholders (Voice of Shareholder) indicate shareholder dissatisfaction, then
the dashboard metrics are obviously inadequate and need to be revised.
Strategy Deployment Plan 71
Figure 2.4. Strategy deployment plan for a hypothetical organization.
The organization is pursuing a particular strategy and emphasizing certain
dashboard metrics, which are shown in boldface type. Goals for these metrics
will be set very high in an attempt to differentiate this organization from its
competition. Goals for other metrics (key requirements) will be set to achieve
competitiveness. Usually this means to maintain historical levels for these
The organization’s leaders believe their core competencies are in the areas of
technology and customer service. They want their customers to think of them as
The company to go to for the very best products completely customized to
meet extremely demanding needs.
However, note that the organization’s differentiators are:
1. Cost per unit
2. Revenues from new sources
3. [Customer] service relationship
4. Product introductions, [new product] revenues
5. Research deployment time
It appears that item 1 is inconsistent with the leadership vision. Most people
would be confused if asked to achieve benchmark status for items 2^5 as
well as for item 1. The plan indicates that the productivity strategy for this
organization should be reevaluated. Unless the company is losing its market
due to uncompetitive prices, or losing its investors due to low profits, item 1
should probably be a key requirement maintained at historical levels. If costs
are extremely out of line, cost per unit might be the focus of a greater than
normal amount of attention to bring it down to reasonable levels. However,
it should not be shown as a differentiator on the strategic dashboard. The
company has no desire to become a cost leader in the eyes of customers or
Six Sigma plays a vital role in achieving the leadership vision by providing the
resources needed to facilitate change where it is needed. Six Sigma projects are
linked to dashboard metrics through the project selection process discussed
elsewhere in this book. The process involves calculating the expected impact of
the project on a dashboard metric. The metrics used for Six Sigma projects are
typically on a lower-level dashboard, but since the lower-level dashboard
metrics flow down from the top level, the linkage is explicit. The process begins
by identifying the gap between the current state and the goal for each top-level
dashboard metric; Master Black Belts commonly assist with this activity. Six
Sigma projects impacting differentiator dashboard metrics which show large
gaps are prime candidates. This determination is usually done by Master Black
Belts. This information is also very useful in selecting Black Belt candidates.
Candidates with backgrounds in areas where high-impact projects will be
Strategy Deployment Plan 73
pursued may be given preference over equally qualified candidates from elsewhere
in the organization.
Six Sigma technical leaders work to extract actionable knowledge from an
organization’s information warehouse. To assure access to the needed information,
Six Sigma activities should be closely integrated with the information
systems (IS) of the organization. Obviously, the skills and training of Six
Sigma technical leaders must be supplemented by an investment in software
and hardware. It makes little sense to hamstring these experts by saving a few
dollars on computers or software. Six Sigma often requires the analysis of huge
amounts of data using highly sophisticated algorithms. The amount of time
required to perform the analysis can be considerable, even with today’s
advanced processing equipment. Without state-of-the-art tools, the situation is
often hopeless.
Integrating Six Sigma with other information
systems technologies
There are three information systems topics that are closely related to Six
Sigma activities:
. Data warehousing
. On-line Analytic Processing (OLAP)
. Data mining
The first topic relates to what data is retained by the organization, and therefore
available for use in Six Sigma activities. It also impacts on how the data is
stored, which impacts on ease of access for Six Sigma analyses. OLAP enables
the analysis of large databases by persons who may not have the technical background
of a Six Sigma technical leader. Data mining involves retrospective analysis
of data using advanced tools and techniques. Each of these subjects will
be discussed in turn.
Data warehousing has progressed rapidly. Virtually non-existent in 1990,
now every large corporation has at least one data warehouse and some have several.
Hundreds of vendors offer data warehousing solutions, from software to
hardware to complete systems. Few standards exist and there are as many data
warehousing implementations as there are data warehouses. However, the
multitiered approach to data warehousing is a model that appears to be gaining
favor and recent advances in technology and decreases in prices have made this
option more appealing to corporate users.
Multitiered data warehousing architecture focuses onhowthe data are used in
the organization. While access and storage considerations may require summarization
of data into multiple departmental warehouses, it is better for Six
Sigma analysis if the warehouse keeps all of the detail in the data for historical
analysis. The major components of this architecture are (Berry and Linoff, 1997):
. Source systems are where the data come from.
. Data transport and cleansing move data between di?erent data stores.
. The central repository is the main store for the data warehouse.
. The metadata describes what is available and where.
. Data marts provide fast, specialized access for end users and applications.
. Operational feedback integrates decision support back into the operational
. End users are the reason for developing the warehouse in the ?rst place.
Figure 2.5 illustrates the multitiered approach.
Every data warehouse includes at least one of these building blocks. The data
originates in the source systems and flows to the end users through the various
components. The components can be characterized as hardware, software, and
networks. The purpose is to deliver information, which is in turn used to create
new knowledge, which is then acted on to improve business performance. In
other words, the data warehouse is ultimately a component in a decisionsupport
On-line analytic processing, or OLAP, is a collection of tools designed to
provide ordinary users with a means of extracting useful information from
large databases. These databases may or may not reside in a data warehouse.
If they do, then the user obtains the benefit of knowing the data has already
been cleansed, and access is likely to be more efficient. OLAP consists of
client-server tools that have an advanced graphical interface that accesses
data arranged in ‘‘cubes.’’ The cube is ideally suited for queries that allow
users to slice-and-dice the data in any way they see fit. OLAP tools have
very fast response times compared to SQLqueries on standard relational
The basic unit of OLAP is the cube. An OLAP cube consists of subcubes that
summarize data from one or more databases. Each cube is composed of multiple
dimensions which represent different fields in a database. For example, an
OLAP cube might consist of warranty claims arranged by months, products,
and region, as shown in Figure 2.6.
Information Systems Requirements 75
Data mining is the exploration and analysis by automatic or semi-automatic
means of large quantities of data in order to uncover useful patterns. These patterns
are studied in order to develop performance rules, i.e., new and better
ways of doing things. Data mining, as used in Six Sigma, is directed toward
improving customer satisfaction, lowering costs, reducing cycle times, and
increasing quality.
Data mining is a grab-bag of techniques borrowed from various disciplines.
Like Six Sigma, data mining alternates between generating questions via knowl-
Figure 2.5. The multitiered approach to data warehousing.
From Data Mining Techniques for Marketing, Sales, and Customer Support, byMichael J.A.
Berry and Gordon Lino?, New York, JohnW|ley and Sons, 1997. Used by permission of the
edge discovery, and testing hypotheses via designed experiments. Six Sigma and
data mining both look for the same things in evaluating data, namely classification,
estimation, prediction, a?nity grouping, clustering and description.
However, data mining tends to use a di?erent set of tools than traditional Six
Sigma tools and therefore it o?ers another way to look for improvement opportunities.
Also, where Six Sigma tends to focus on internal business processes,
data mining looks primarily at marketing, sales, and customer support. Since
the object of Six Sigma is, ultimately, to improve customer satisfaction, the
external focus of data mining provides both feed forward data to the Six Sigma
program and feed back data on its success.
Data mining is a process for retrospectively exploring business data. There is
growing agreement on the steps involved in such a process and any differences
relate only to the detailed tasks within each stage.*
Goal definitionLThis involves defining the goal or objective for the data
mining project. This should be a business goal or objective which normally
relates to a business event such as arrears in mortgage repayment,
customer attrition (churn), energy consumption in a process, etc. This
stage also involves the design of how the discovered patterns will result
in action that leads to business improvement.
Data selectionLThis is the process of identifying the data needed for the
data mining project and the sources of these data.
Data preparationLThis involves cleansing the data, joining/merging
data sources and the derivation of new columns (fields) in the data
through aggregation, calculations or text manipulation of existing
data fields. The end result is normally a flat table ready for the
application of the data mining itself (i.e. the discovery algorithms
to generate patterns). Such a table is normally split into two data
Information Systems Requirements 77
Figure 2.6. An OLAP cube.
sets; one set for pattern discovery and one set for pattern verification.
Data explorationLThis involves the exploration of the prepared data to get
a better feel prior to pattern discovery and also to validate the results
of the data preparation. Typically, this involves examining descriptive
statistics (minimum, maximum, average, etc.) and the frequency distribution
of individual data fields. It also involves field versus field scatter
plots to understand the dependency between fields.
Pattern discoveryLThis is the stage of applying the pattern discovery algorithm
to generate patterns. The process of pattern discovery is most
effective when applied as an exploration process assisted by the discovery
algorithm. This allows business users to interact with and to impart
their business knowledge to the discovery process. For example, if
creating a classification tree, users can at any point in the tree construction
examine/explore the data filtering to that path, examine the
recommendation of the algorithm regarding the next data field to use
for the next branch then use their business judgment to decide on the
data field for branching. The pattern discovery stage also involves analyzing
the ability to predict occurrences of the event in data other than
those used to build the model.
Pattern deploymentLThis stage involves the application of the discovered
patterns to solve the business goal of the data mining project. This can
take many forms:
Pattern presentationLThe description of the patterns (or the graphical
tree display) and their associated data statistics are included in a document
or presentation.
Business intelligenceLThe discovered patterns are used as queries against a
database to derive business intelligence reports.
Data scoring and labelingLThe discovered patterns are used to score and/
or label each data record in the database with the propensity and the
label of the pattern it belongs to.
Decision support systemsLThe discovered patterns are used to make components
of a decision support system.
Alarm monitoringLThe discovered patterns are used as norms for a business
process. Monitoring these patterns will enable deviations from
normal conditions to be detected at the earliest possible time. This
can be achieved by embedding the data mining tool as a monitoring
component, or through the use of a classical approach, such as control
Pattern validity monitoringLAs a business process changes over time, the
validity of patterns discovered from historic data will deteriorate. It is
therefore important to detect these changes at the earliest possible time
by monitoring patterns with new data. Significant changes to the
patterns will point to the need to discover new patterns from more
recent data.
OLAP, data mining, and Six Sigma
OLAP is not a substitute for data mining. OLAP tools are a powerful means
for reporting on data, while data mining focuses on finding hidden patterns in
data. OLAP helps users explore theories they already have by quickly presenting
data to confirm or disconfirm ad hoc hypotheses, obviously a valuable
knowledge discovery tool for Six Sigma teams. It is, essentially, a semi-automated
means of analysis. OLAP and data mining are complementary, and both
approaches complement the standard arsenal of tools and techniques used in
Six Sigma. Both OLAP and data mining are used for retrospective studies, that
is, they are used to generate hypotheses by examining past data. Designed
experiments help users design prospective studies, that is, they test the hypotheses
generated by OLAP and data mining. Used together, Six Sigma, data
mining and OLAP comprise a powerful collection of business improvement
Strategies are operationalized by metrics which are displayed on dashboards.
Dashboard displays should be designed to provide the needed information
in a way that is standardized throughout the organization. A process
owner at any level of the organization should be able to look at any dashboard
and quickly recognize the meaning of the data. The purpose of data displays
is to accelerate the learning cycle. The strategy deployment plan is merely a
hypothesis. Science-based management requires that we test this hypothesis
to determine if it is in reasonable agreement with the facts, and take action
or revise the strategy deployment plan or the strategy accordingly. The cycle
works as follows:
1. Formulate a strategy (hypothesis).
2. Develop metrics to operationalize the strategy.
3. Deploy the strategy.
4. Collect data for the metrics.
5. Analyze the data to extract information regarding the e?ectiveness of the
strategy deployment plan. This includes the use of statistical tools and
techniques, graphs and charts, discussion of results, etc.
Dashboard Design 79
6. Think about the result indicated by the information and whether it validates
or invalidates the strategy and/or the metrics used to operationalize
7. Take appropriate action. This may be no action (the null option),
revision of the strategy, revision of the metrics, or some other steps.
This process is illustrated in Figure 2.7.
Dashboard metrics should embody all of the general principles of good
metrics discussed earlier. More specifically, dashboards should:
. Display performance over time.
. Include statistical guidelines to help separate signal (variation from an
identi?able cause) from noise (variation similar to random ?uctuations).
. Show causes of variation when known.
. Identify acceptable and unacceptable performance (defects).
. Be linked to higher-level dashboards (goals and strategies) or lower-level
dashboards (drivers) to guide strategic activity within the organization.
Although all dashboards should conform to these guidelines, different dashboard
formats are needed for data on different scales of measurement (see
Chapter 9). Because of the nature of measurement scales, some data contain
more information than other data. For example, we might be interested in the
size of a hole that will have a bushing pressed into it. If the hole is too large, the
bushing will be loose and it will wear out quickly. If the hole is too small the
bushing won’t fit at all. Assume that there are three different methods available
for checking the hole size.
Method #1 is a hole gage that measures the actual size of the hole. These
data are called ratio data. This measurement scale contains the most information.
Interval data such as time and temperature are often treated as if
they were ratio data, which is usually acceptable for dashboards. In our discussions
of dashboards we will refer to both ratio and interval data as scale
Method #2 is a set of four pin gages. One set of two pins determines if the
hole is smaller than the minimum requirement, or larger than the maximum
requirement. For example, if the hole size requirement is 1.000 to 1.010, then
this set will determine if the hole is smaller than 1.000 or larger than 1.010.
Another set of two pins determines if the hole is in the middle half of the requirement
range. For example, if the hole size requirement is 1.000 to 1.010, then
Figure 2.7. The learning cycle.
this set of pins will determine if the hole is smaller than 1.0025 or larger than
1.0075. Thus, hole sizes will be classified by this measurement system as:
Best:Middle half of requirements.
Acceptable:Not inmiddle half, but stillmeets requirements.
Reject: Does not meet requirements.
These data are called ordinal data. That is, the measurements can be placed in
an order of preference. Ordinal data don’t contain as much information as
ratio data or interval data. For example, we can calculate the precise difference
between two hole sizes if we know their measurements, but we can’t do so if
we only know the hole size classification.
Method #3 is a single pair of go/not-go pin gages: pins will be used to determine
if the hole meets the requirements or not. These are nominal data.
Nominal data have less information than ratio, interval or ordinal data.
Although dashboards are discussed in some detail and examples shown and
interpreted, there are no hard and fast rules for dashboards. Any dashboard
that displays metrics derived from strategic goals and conforming to the principles
of good metrics described above is acceptable, providing it supplies the
information needed to make good decisions in a timely manner. You should
avoid cookbook dashboard development and design dashboards that help you
assure that your strategies are being effectively deployed and accomplishing
your ultimate goals.
Dashboards for scale data
Data on these measurement scales contain the maximum amount of information.
In fact, as the examples above show, it is a simple matter to derive
ordinal or nominal metrics from scale data, but one cannot go in reverse. To
take advantage of the information contained in scale data they should be
viewed in a variety of different ways on the dashboards. Figure 2.8 provides
These different, but related views give a detailed picture of the performance
of the metric. If annotated and linked to driver dashboards, the reason why the
metric did what it did should be relatively easy to determine.
Figure 2.9 shows a dashboard for a fictitious customer service call center. The
metric is speed of answer, which was decided upon based on customer input. It
was determined from actual customer information that customers viewed
Dashboard Design 81
speed-to-answer times in excess of 6 minutes to be unacceptable, so this is
defined as the requirement.
Process central tendency: The chart in Quad I shows the average speed to
answer (ASA) for the most recent month. The dashed lines are statistically
calculated ‘‘limits,’’ which define the normal range of variability in ASA for
this process (for additional information on averages’ charts, see Chapter
12). The chart indicates that the process ASA is stable and averaging 3.2
Distribution of calls: ASA is a good indicator of overall process control, but
individual customers have individual experiences, so the average doesn’t mean
much to them. The histogram in Quad II gives a better indication of the individual
customer’s experience in the most recent week. The bar labeled ‘‘More’’
indicates the calls not meeting the customer requirement; it’s a pretty big bar.
Process Behavior Chart of the metric
in time order. E.g., by hour, day, week,
or month. Generally an averages chart
or equivalent. Learn by studying timerelated
patterns, outliers, etc.
Histogram showing distribution of the
metric for the most recent time
period. E.g., histogram of the most
recent month. The requirement
should be drawn on the histogram to
identify defects. Learn by studying the
frequency distribution pattern.
A Process Behavior Chart of the rate
of defects or defectives in time-order.
Learn by studying defect patterns.
A chart showing the distribution of
individual defectives, such as a dotplot
or box-and-whiskers charts. I.e., an
‘‘ungrouping’’ of the defect data
shown as grouped by the histogram.
Learn by studying process changes
designed to improve extremely poor
customer experiences.
Figure 2.8. Layout of a scale data dashboard.
Despite a process average of 3.2 minutes, approximately 15% of the calls
answered this week were not answered within 6 minutes.
Defectives over time: Okay, so we failed to meet the customer requirements
15% of the time during this week. Is that due to some special circumstance
that only happened once? The chart in Quad III shows the pattern of defectives
over time; the dashed lines are limits for the percent not meeting
requirements. The chart shows that the defective rate is stable and averaging
Outlier or tails perspective: Finally, a box-and-whiskers chart, or boxplots
(see Chapter 11) is shown in Quad IV. The time scale is chopped off (truncated)
at 6 minutes, since we are interested in learning about calls that failed
to meet the requirement. In the plot, an ‘‘*’’ or an ‘‘o’’ is an individual call
that is considered an outlier. The chart indicates that some people wait a
very long time for the phone to be answered, but most get an answer within
10 minutes.
Dashboard Design 83
Figure 2.9. Example of a scale data dashboard.
It is important to note that although the speed-to-answer metric indicates a
stable process mean, this may not be what management wants to see. In this
case, it’s very unlikely that management will be satisfied with a process that consistently
makes 15% of the customers wait too long for their call to be answered.
This metric is a good candidate for one or more Six Sigma projects. The project
Black Belt will no doubt drill down to lower level dashboards to attempt to identify
drivers to address during the project. For example, are there differences by
department? By technician?
The interpretation of the dashboard also depends on the strategy it operationalizes.
In this example, if the company is pursuing a strategy where process
excellence is a primary driver, then the 6 minute requirement might not be sufficient.
Instead of asking customers what level of service would satisfy them, the
focus might be on the level of service that would delight them. Perhaps the company’s
leadership believes that some customers (e.g., professionals) would pay
a premium price to have the phone answered immediately, in which case the
metric of interest might be ‘‘percentage of calls not answered within three
rings.’’ Staff levels could be increased because of the increased prices paid by
these premier customers.
Dashboards for ordinal data
Figure 2.10 provides guidelines for ordinal data dashboards. These different,
but related views give a detailed picture of the performance of the metric. If
annotated and linked to driver dashboards, the reason why the metric did what
it did should be relatively easy to determine.
Figure 2.11 shows a dashboard from a customer survey metric measuring
how easy it is for the customer to contact the call center.
Process central tendency: The chart in Quad I shows the weekly average ease
of contact scores. This is the same type of chart used in Quad I for scale data
dashboards, and it is used the same way.
Distribution of ratings: Quad II shows the distribution of actual customer
ratings. The customers had to choose a response on a five-point scale; higher
numbered responses are better. Previous research had determined that customers
who gave a response below 4 were less likely to remain customers than
those who scored easy-to-contact a 4 or better, so that’s where the defect line is
drawn on the bar chart. Note that this isn’t a histogram because the data are
not grouped. The bar at, say, 2 indicates the number of customers who rated
easy-to-contact a 2. A large number of customers in this example rate ease of
contact very low.
Defectives over time: Quad III for ordinal data is a process behavior
chart of the defective rate, just like it is in the scale data dashboard. Since
defective rates are pass/fail data (did or did not meet the requirement), it
is a nominal measurement and we can convert the ordinal data into nominal
data for this chart. This chart shows that the defective rate for this
call center is stable at 48%; i.e., we consistently fail to meet the customer
Outlier or tails perspective: Since we are analyzing the voice of the customer,
leadership wanted to use Quad IV to display some actual customer comments.
(Comments can also be statistically analyzed.) The comments highlight some
problem areas that could become Six Sigma projects. Other problem areas
have obvious solutions and can be addressed without Six Sigma projects.
Don’t use a hammer to swat a fly!
Dashboard Design 85
Process Behavior Chart of the metric
in time order. E.g., by hour, day, week,
or month. Generally an averages chart
or equivalent. Learn by studying timerelated
patterns, outliers, etc.
Bar chart showing distribution of the
metric for the most recent time
period. E.g., bar chart of the most
recent month. The requirement
should be drawn on the bar chart to
identify defects. Learn by studying the
frequency distribution pattern.
A Process Behavior Chart of the rate
of defects or defectives in time-order.
Learn by studying defect patterns.
Text, drawings, links to other
documents, photos, maps or other
information that lends meaning to the
Figure 2.10. Layout of an ordinal data dashboard.
Like speed-to-answer, the dashboard for easy-to-contact indicates a stable
process mean at an unacceptable performance level. It may be that management
is responsible for this situation, which is usually the case when a process
is stable. The Six Sigma project may require a new process design. Stability
means that there’s no point in looking for ‘‘problems.’’ If management isn’t
happy with the results, the problem is the process itself. The process redesign
should be linked to the overall strategy. This means, for example, if the strategy
is to make the customer relationship a differentiator, then the goal for
easy-to-contact should be set at or near a benchmark level. If it is a requirement,
then the goal should be set near the industry average. However, keep
in mind that if the industry average is awful, then differentiation should be
relatively easy to attain.

Figure 2.11. Example of an ordinal data dashboard.
Dashboards for nominal data
Nominal data, such as pass-fail, yes-no, acceptable-unacceptable, met goaldidn’t
meet goal, are based on rates (e.g., failure rates), counts or proportions
of counts. Unlike scale or ordinal data, nominal data can’t be further broken
down into numbers on other scales. Typically, nominal data dashboards show
defect metrics or, equivalently, success metrics. In many cases where defectives
or failures are measured, non-defectives or non-failures provide identical information.
Figure 2.12 provides guidelines for nominal data dashboards. Since there are
limits to how much information can be obtained by analyzing nominal data
directly, nominal data dashboards focus on providing background details.
Also, since nominal metrics are so often measures of process failure, you may
wish to devote part of the dashboard to descriptions of action being taken to
improve the metric.
A Process Behavior Chart of the rate
of defects or defectives in time-order.
Learn by studying defect patterns.
A graph or table providing additional
information regarding the metric.
A graph or table providing additional
information regarding the metric.
Action plans, responsibilities,
timetables, etc. for improving the
Figure 2.12. Layout of a nominal data dashboard.
Figure 2.13 shows a dashboard from a customer survey metric measuring
the rate at which customer issues were unresolved after they spoke with a
technician. Obviously, this can be considered a failure in terms of customer
Defectives over time: Quad I for nominal data is a process behavior chart of
the defective rate. This chart appears on the dashboards for scale and ordinal
data too, but in later quads. In this example the chart shows that the non-resolution
rate for this call center is not stable. During one week the rate of unresolved
Dashboard Design 87
problems dropped below the lower control limit. An investigation revealed that
this was due to the fact that the problem mix was influenced by a national holiday.
This information provided a hint to management: problem resolution is
influenced by the type of problem. Excluding the unusual week, the chart
shows that on average about 11% of customer issues are unresolved.
Bar chart of calls needed to resolve issues: Quad II looks more deeply into
the process failure. Rather than simply count the number of unresolved customer
issues, this chart shows how many attempts customers made before
their issues were resolved. This trial-and-error approach is frustrating to customers
and costly to the enterprise. It indicates a fairly consistent drop-off in
the frequency of customers who call back two or more times. Is this due to
their problems being resolved, or frustration and eventual abandoning of the
effort? This is a good question for a Black Belt to address as part of a Six
Sigma project, but it isn’t answered by information on this particular dashboard.
Figure 2.13. Problem resolution dashboard.
Pareto analysis by call type: Tipped off by the holiday week outlier, leadership
asked for an analysis by call type. This information is shown in Quad III.
Corrupt data files account for nearly half of the unresolved customer issues,
suggesting a possible area to be addressed by a Six Sigma project.
Process FMEA: FMEA stands for Failure Mode and Effects Analysis. As the
name suggests, FMEA is a tool that can be used to identify the way in which
the process fails. Quad IV could be used to display such information. Once a
project is underway to address issues identified by the FMEA, Quad IV could
be used to track the project’s progress. FMEA is discussed in Chapter 16.
This dashboard shows that there is a real opportunity to improve customer
satisfaction. The key information is contained in the process behavior chart in
Quad I. The long-term rate of unresolved issues of 11% is costing the company
a lot of money, and frustrating customers. Each call beyond the first is pure
Plans, budgets, goals and targets are key requirements set by the leadership
for the organization. If not done properly, the behavior driven by these key
requirements may not be anywhere close to what the leadership desires, or
expects. Key requirements are used to assess employee performance, which is
linked to promotions, pay increases, bonuses and many other things that people
care about a great deal. People will try hard to meet the key requirements, but
if the process they must work with makes it impossible to do so they will often
cheat (see sidebar, Gaming the System).
The most common flaw in goal setting, in my opinion, is the tendency to
set goals that are merely wishes and hopes. The leadership looks at a metric
and pontificates on what a ‘‘good’’ level of performance would be for it.
If enough heads nod around the conference table, this becomes that metric’s
A better way to arrive at goals for key requirements is to examine the actual
history of the metric over time. This information should be plotted on a process
behavior chart. If the metric falls within the calculated limits the bulk of the
time, then the process is considered predictable. Typically, unless the metric is
operationalizing a differentiation strategy, the goal for predictable processes
will be to maintain the historical levels. These metrics will not appear on the
dashboards that the leadership reviews on a routine basis. However, their
performance is monitored by process owners and the leadership is informed if
Setting Organizational Key Requirements 89
Gaming the System
It has been said that the managers of factories in the former Soviet Union didn’t fail
to meet their numerical targets, rather that they met them too well. Is the quota for
my shoe factory 100,000 pairs of shoes? Here are 100,000 pairs of baby shoes, the easiest
to make. Change the quota to 10,000 pounds of shoes and the shoe factory manager
will deliver 500 pairs of concrete boots. One advantage of balanced scorecards is that
they make it more di?cult to get away with ‘‘gaming the system’’ like this. Gaming the
system involves the manipulation of metrics to reach numerical targets, rather than
actually achieving the goals themselves. Of course, balanced scorecards can’t solve the
problems inherent in communism, and the situation in our companies is nowhere near
as bad. But we’ve all seen similar behavior when managers are given metrics for their
local area and inadequate information about how the metrics ?t into the grand scheme
of things.
Another common game that gets played is ‘‘denominator management.’’ Denominator
management is the practice of manipulating the base of a metric, rather than doing the
work necessary to change the underlying reality. Is my metric defects-per-million-opportunities?
Well, reducing defects is di?cult and time-consuming. I’ll just manipulate the number
of opportunities. Table 2.3 below shows a few examples of this, creative gamers can
come up with many more!
Table 2.3. Examples of denominator management.
Intent Denominator Management Approach
1. Reduce defective circuit
2. Reduce customer billing
3. Improve time to introduce
new products
4. Reduce setup costs
5. Improve process yields
1. Count the number of components not
the number of circuit boards
2. Count the lines (or words) on billing
statements not the number of
3. Develop simple products
4. Increase batch sizes
5. Produce more simple parts
By linking the metric to the strategic goal and by simultaneously monitoring all key differentiator
metrics on the stakeholder dashboards, gaming is minimized. Six Sigma’s rationalization
of management also makes gaming more di?cult. If, despite repeated warnings,
some people persist in gaming the system, and assuming they are given the resources they
need to meet the key requirements, they should be disciplined or terminated. Gaming is
inherently dishonest and it has no place in the Six Sigma organization.
the process behavior becomes unpredictable. In other words, the reporting is on
an exception basis.
If process behavior charts indicate that a key requirement metric is not
predictable, an investigation into the reason should ensue and the cause of the
unpredictability should be corrected.
Finally, if the key requirement metric is so far from generally accepted
standards of performance that it demands action, a short-term project
should be commissioned to address the issue. The project team should
focus on identifying why performance is so far below the norm and on
what needs to be done to remedy the situation. This is not a strategic
focus, but a remedial one, and it should not distract the leadership from
pursuing their vision.
Goal setting for differentiators is another matter entirely. Unlike key
requirements, the historical level of performance for differentiators is unacceptable
by definition. Leadership doesn’t want to maintain differentiator performance,
it wants to improve it dramatically. Setting goals for differentiators is
discussed next.
Benchmarking is a topic of general interest in Six Sigma. Thus, the discussion
here goes beyond the use of benchmarking in project management
Benchmarking is a popular method for developing requirements and setting
goals. In more conventional terms, benchmarking can be defined as measuring
your performance against that of best-in-class companies, determining how
the best-in-class achieve those performance levels, and using the information
as the basis for your own company’s targets, strategies, and implementation.
Benchmarking involves research into the best practices at the industry, firm,
or process level. Benchmarking goes beyond a determination of the ‘‘industry
standard;’’ it breaks the firm’s activities down to process operations and looks
for the best-in-class for a particular operation. For example, to achieve improvement
in their parts distribution process Xerox Corporation studied the retailer
L.L. Bean.
Benchmarking goes beyond the mere setting of goals. It focuses on practices
that produce superior performance. Benchmarking involves setting up partnerships
that allow both parties to learn from one another. Competitors can also
engage in benchmarking, providing they avoid proprietary issues.
Benchmarking projects are like any other major project. Benchmarking must
have a structured methodology to ensure successful completion of thorough
and accurate investigations. However, it must be flexible to incorporate new
Setting Organizational Key Requirements 91
and innovative ways of assembling difficult-to-obtain information. It is a discovery
process and a learning experience. It forces the organization to take an
external view, to look beyond itself.
Camp (1989) lists the following steps for the benchmarking process:
1. Planning
1.1. Identify what is to be benchmarked
1.2. Identify comparative companies
1.3. Determine data collection method and collect data
2. Analysis
2.1. Determine current performance ‘‘gap’’
2.2. Project future performance levels
3. Integration
3.1. Communicate benchmark ?ndings and gain acceptance
3.2. Establish functional goals
4. Action
4.1. Develop action plans
4.2. Implement speci?c actions and monitor progress
4.3. Recalibrate benchmarks
5. Maturity
5.1. Leadership position attained
5.2. Practices fully integrated into process
The first step in benchmarking is determining what to benchmark. To focus
the benchmarking initiative on critical issues, begin by identifying the process
outputs most important to the customers of that process (i.e., the key quality
characteristics). This step applies to every organizational function, since each
one has outputs and customers. The QFD/customer needs assessment is a
natural precursor to benchmarking activities.
The essence of benchmarking is the acquisition of information. The process
begins with the identification of the process that is to be benchmarked. The
process chosen should be one that will have a major impact on the success of
the business.
Once the process has been identified, contact a business library and request
a search for the information relating to your area of interest. The library will
identify material from a variety of external sources, such as magazines,
journals, special reports, etc. You should also conduct research using the
Internet and other electronic networking resources. However, be prepared to
pare down what will probably be an extremely large list of candidates (e.g.,
an Internet search on the word ‘‘benchmarking’’ produced 20,000 hits).
Don’t forget your organization’s internal resources. If your company has an
‘‘Intranet’’ use it to conduct an internal search. Set up a meeting with people
in key departments, such as R&D. Tap the expertise of those in your company
who routinely work with customers, competitors, suppliers, and other ‘‘outside’’
organizations. Often your company’s board of directors will have an
extensive network of contacts.
The search is, of course, not random. Look for the best of the best, not the
average firm. There are many possible sources for identifying the elites. One
approach is to build a compendium of business awards and citations of merit
that organizations have received in business process improvement. Sources to
consider are Industry Week’s Best Plant’s Award, National Institute of
Standards and Technology’s Malcolm Baldrige Award, USA Today and
Rochester Institute of Technology’s Quality Cup Award, European
Foundation for Quality Management Award, Occupational Safety and Health
Administration (OSHA), Federal Quality Institute, Deming Prize,
Competitiveness Forum, Fortune magazine, United States Navy’s Best
Manufacturing Practices, to name just a few. You may wish to subscribe to an
‘‘exchange service’’ that collects benchmarking information and makes it available
for a fee. Once enrolled, you will have access to the names of other subscribers
La great source for contacts.
Don’t overlook your own suppliers as a source for information. If your company
has a program for recognizing top suppliers, contact these suppliers and
see if they are willing to share their ‘‘secrets’’ with you. Suppliers are predisposed
to cooperate with their customers; it’s an automatic door-opener. Also
contact your customers. Customers have a vested interest in helping you do a
better job. If your quality, cost, and delivery performance improve, your customers
will benefit. Customers may be willing to share some of their insights
as to how their other suppliers compare with you. Again, it isn’t necessary that
you get information about direct competitors. Which of your customer’s
suppliers are best at billing? Order fulfillment? Customer service? Keep your
focus at the process level and there will seldom be any issues of confidentiality.
An advantage to identifying potential benchmarking partners through your customers
is that you will have a referral that will make it easier for you to start
the partnership.
Another source for detailed information on companies is academic research.
Companies often allow universities access to detailed information for research
purposes. While the published research usually omits reference to the specific
companies involved, it often provides comparisons and detailed analysis of
Setting Organizational Key Requirements 93
what separates the best from the others. Such information, provided by experts
whose work is subject to rigorous peer review, will often save you thousands
of hours of work.
After a list of potential candidates is compiled, the next step is to choose the
best three to five targets. A candidate that looked promising early in the process
might be eliminated later based on the following criteria (Vaziri, 1992):
. Not the best performer
. Unwilling to share information and practices (i.e., doesn’t view the benchmarking
process as a mutually bene?cial learning opportunity)
. Low availability and questionable reliability of information on the
As the benchmarking process evolves, the characteristics of the most desirable
candidates will be continually refined. This occurs as a result of a clearer
understanding of your organization’s key quality characteristics and critical
success factors and an improved knowledge of the marketplace and other
This knowledge and the resulting actions tremendously strengthen an
The causes of failed benchmarking projects are the same as those for other
failed projects (DeToro, 1995):
. Lack of sponsorshipLA team should submit to management a one- to
four-page benchmarking project proposal that describes the project, its
objectives, and potential costs. If the team can’t gain approval for the project
or get a sponsor, it makes little sense to proceed with a project that’s
not understood or appreciated or that is unlikely to lead to corrective
action when completed.
. Wrong people on teamLWho are the right people for a benchmarking
team? Individuals involved in benchmarking should be the same ones
who own or work in the process. It’s useless for a team to address
problems in business areas that are unfamiliar or where the team has no
control or in?uence.
. Teams don’t understand their work completelyLIf the benchmarking
team didn’t map, ?owchart, or document its work process, and if it didn’t
benchmark with organizations that also documented their processes,
there can’t be an e?ective transfer of techniques. The intent in every
benchmarking project is for a team to understand how its process works
and compare it to another company’s process at a detailed level. The
exchange of process steps is essential for improved performance.
. Teams take on too muchLThe task a team undertakes is often so broad
that it becomes unmanageable. This broad area must be broken into smaller,
more manageable projects that can be approached logically. A suggested
approach is to create a functional ?owchart of an entire area, such
as production or marketing, and identify its processes. Criteria can then
be used to select a process to be benchmarked that would best contribute
to the organization’s objectives.
. Lack of long-term management commitmentLSince managers aren’t
as familiar with speci?c work issues as their employees, they tend to
underestimate the time, cost, and e?ort required to successfully complete
a benchmarking project. Managers should be informed that while it’s
impossible to know the exact time it will take for a typical benchmarking
project, there is a rule of thumb that a team of four or ?ve individuals
requires a third of their time for ?ve months to complete a project.
. Focus on metrics rather than processesLSome ?rms focus their benchmarking
e?orts on performance targets (metrics) rather than processes.
Knowing that a competitor has a higher return on assets doesn’t mean
that its performance alone should become the new target (unless an understanding
exists about how the competitor di?ers in the use of its assets
and an evaluation of its process reveals that it can be emulated or surpassed).
. Not positioning benchmarking within a larger strategyLBenchmarking
is one of many Six Sigma toolsLsuch as problem solving, process
improvement, and process reengineeringLused to shorten cycle time,
reduce costs, and minimize variation. Benchmarking is compatible with
and complementary to these tools, and they should be used together for
maximum value.
. Misunderstanding the organization’s mission, goals, and objectivesL
All benchmarking activity should be launched by management as part of
an overall strategy to ful?ll the organization’s mission and vision by ?rst
attaining the short-term objectives and then the long-term goals.
. Assuming every project requires a site visitLSu?cient information is
often available from the public domain, making a site visit unnecessary.
This speeds the benchmarking process and lowers the cost considerably.
. Failure to monitor progressLOnce benchmarking has been completed
for a speci?c area or process benchmarks have been established and
process changes implemented, managers should review progress in
implementation and results.
The issues described here are discussed in other parts of this chapter and in
other parts of this book. The best way of dealing with them is to prevent their
occurrence by carefully planning and managing the project from the outset.
Setting Organizational Key Requirements 95
This list can be used as a checklist to evaluate project plans; if the plans don’t
clearly preclude these problems, then the plans are not complete.
The benefits of competitive benchmarking include:
. Creating a culture that values continuous improvement to achieve excellence
. Enhancing creativity by devaluing the not-invented-here syndrome
. Increasing sensitivity to changes in the external environment
. Shifting the corporate mind-set from relative complacency to a strong
sense of urgency for ongoing improvement
. Focusing resources through performance targets set with employee input
. Prioritizing the areas that need improvement
. Sharing the best practices between benchmarking partners
Benchmarking is based on learning from others, rather than developing new
and improved approaches. Since the process being studied is there for all to
see, benchmarking cannot give a firm a sustained competitive advantage.
Although helpful, benchmarking should never be the primary strategy for
Competitive analysis is an approach to goal setting used by many firms. This
approach is essentially benchmarking confined to one’s own industry.
Although common, competitive analysis virtually guarantees second-rate
quality because the firm will always be following their competition. If the entire
industry employs the approach it will lead to stagnation for the entire industry,
setting them up for eventual replacement by outside innovators.
^ ^ ^
Creating Customer-Driven
The proper place of the customer in the organization’s hierarchy is illustrated
in Figure 3.1.
Note that this perspective is precisely the opposite of the traditional view of
the organization. The difficulties involved in making such a radical change
should not be underestimated.
Figure 3.1. The ‘‘correct’’ view of the company organization chart.
From Marketing Management: Analysis, Planning, Implementation, and Control, Figure 1^7,
p. 21, by Philip Kotler, copyright#1991 by Prentice-Hall, Inc. Reprinted by permission.
Copyright 2003 by The McGraw-Hill Companies, Inc. Click Here for Terms of Use.
Becoming a customer- and market-driven
Edosomwan (1993) defines a customer- and market-driven enterprise as one
that is committed to providing excellent quality and competitive products and
services to satisfy the needs and wants of a well-defined market segment. This
approach is in contrast to that of the traditional organization, as shown in
Table 3.1.
The journey from a traditional to a customer-driven organization has been
made by enough organizations to allow us to identify a number of distinct milestones
that mark the path to success. Generally, the journey begins with recognition
that a crisis is either upon the organization, or imminent. This wrenches
the organization’s leadership out of denial and forces them to abandon the
status quo.
When the familiar ways of the past are no longer acceptable, the result is a
feeling of confusion among the leaders. At this stage the leadership must answer
some very basic questions:
. What is the organization’s purpose?
. What are our values?
. What does an organization with these values look like?
A ‘‘value’’ is that which one acts to gain and/or keep. It presupposes an entity
capable of acting to achieve a goal in the face of an alternative. Values are not
simply nice-sounding platitudes, they represent goals. Pursuing the organization’s
values implies building an organization which embodies these values.
This is the leadership’s vision, to create a reality where their values have been
After the vision has been clearly developed, the next step is to develop a
strategy for building the new organization (see Chapter 1). The process of
implementing the strategic plan is the turnaround stage.
Elements of the transformed organization
Customer-driven organizations share certain common features.
. Flattened hierarchiesLGetting everyone closer to the customer involves
reducing the number of bureaucratic ‘‘layers’’ in the organization structure.
It also involves the ‘‘upside-down’’ perspective of the organization
structure shown in Figure 3.1. The customer comes ?rst, not the boss.
Everyone serves the customer.
. Risk-takingLCustomers’ demands tend to be unpredictable.
Responsiveness requires that organizations be willing to change quickly,
Elements of Customer-Driven Organizations 99
Table 3.1. Traditional organizations vs. customer-driven organizations.
From Customer and Market-Driven Quality Management, Table 1.1, by Johnson A.
Edosomwan, copyright#1993 by ASQ. Reprinted by permission.
Product and service
^ Short-term focus
^ Reactionary management
^Management by objectives
planning process
^ L ong-termfocus
^ Prevention-based management
^ Customer-driven strategic
planning process
Measures of
^ Bottom-line ?nancial results
^ Quick return on investment
^ Customer satisfaction
^Market share
^ Long-term pro?tability
^ Quality orientation
^ Total productivity
Attitudes toward
^ Customers are irrational and
a pain
^ Customers are a bottleneck
to pro?tability
^ Hostile and careless
^ ‘‘Take it or leave it’’ attitude
^ Voice of the customer is
^ Professional treatment and
attention to customers is
^ Courteous and responsive
^ Empathy and respectful
Quality of products
and services
^ Provided according to
organizational requirements
^ Provided according to
customer requirements and
Marketing focus ^ Seller’s market
^ Careless about lost
customers due to poor
customer satisfaction
^ Increased market share and
?nancial growth achieved
^ Focus on error and defect
^ Focus on error and defect
which involves uncertainty and risk Customer-driven organizations
encourage risk-taking in a variety of ways. One important aspect is to
celebrate mistakes made by individuals who engage in risky behavior.
Bureaucratic impediments such as excessive dependence on written procedures
are minimized or eliminated. Employees are encouraged to act
on their own best judgments and not to rely on formal approval
. CommunicationLDuring the transformation the primary task of the
leadership team is the clear, consistent, and unambiguous transmission
of their vision to others in the organization. One way this is done is
through ‘‘internal marketing’’ which involves using the principles of marketing
to get the message to the target ‘‘market’’: the employees. It is vital
that the leaders’ actions are completely consistent with their words. The
Product and service
delivery attitude
^ It is OK for customers to
wait for products and
^ It is best to provide fast timeto
market products and
People orientation ^ People are the source of
problems and are burdens on
the organization
^ People are an organization’s
greatest resource
Basis for decisionmaking
^ Product-driven
^Management by opinion
^Management by data
^ Crisis management
^Management by fear and
^ Continuous process
^ Total process management
Mode of operation ^ Career-driven and
independent work
^ Customers, suppliers, and
process owners have nothing
in common
^ Teamwork between suppliers,
process owners, and customers
Table 3.1 continued
assistance of outside consultants may be helpful in identifying inconsistencies.
Leaders should realize that their behavior carries tremendous symbolic
meaning. This can contribute to the failure of convincing employees; a single
action which is inconsistent with the stated message is su?cient to
destroy all credibility.Onthe plus side, an action that clearly shows a commitment
to the vision can help spread the word that ‘‘They’re serious this
time.’’ The leadership should seek out stories that capture the essence of
the new organization and repeat these stories often. For example,
Nordstrom employees all hear the story of the sales clerk who allowed
the customer to return a tire (Nordstrom’s doesn’t sell tires). This story
captures the essence of the Nordstrom ‘‘rule book’’ which states:
Rule #1LUse your own best judgment.
Rule #2LThere are no other rules.
Leaders should expect to devote a minimum of 50% of their time to communication
during the transition.
. Boards of directorsLIt is vital to obtain the enthusiastic endorsement of
the new strategy by the board. Management cannot focus their attention
until this support has been received. This will require that management
educate their board and ask them for their approval. However, boards
are responsible for governance, not management. Don’t ask the board to
approve tactics. This bogs down the board, sti?es creativity in the ranks,
and slows the organization down.
. UnionsLIn the transformed organization, everyone’s job changes. If
the organization’s employees are unionized, changing jobs requires
that the union become management’s partner in the transformation
process. In the ?at organization union employees will have greater
authority. Union representatives should be involved in all phases of
the transformation, including planning and strategy development. By
getting union input, the organization can be assured that during collective
bargaining the union won’t undermine the company’s ability to
compete or sabotage the strategic plan. Unions also play a role in auditing
the company’s activities to assure that they comply with contracts
and labor laws.
. Measuring resultsLIt is important that the right things be measured.
The ‘‘right things’’ are measurements that determine that you are delivering
on your promises to customers, investors, employees, and other
stakeholders. You must also measure for the right reasons. This means
that measurements are used to learn about how to improve, not for
judgment. Finally, you must measure the right way. Measurements
Elements of Customer-Driven Organizations 101
should cover processes as well as outcomes. Data must be available
quickly to the people who use them. Measurements must be easy to
. Rewarding employeesLCare must be taken to avoid punishing with
rewards. Rewarding individuals with ?nancial incentives for simply
doing their jobs well implies that the employee wouldn’t do the job
without the reward. It is inherently manipulative. The result is to
destroy the very behavior you seek to encourage (Kohn, 1993). The message
is that rewards should not be used as control mechanisms.
Employees should be treated like adults and provided with adequate
and fair compensation for doing their jobs. Recognizing exceptional performance
or e?ort should be done in a way that encourages cooperation
and team spirit, such as parties and public expressions of appreciation.
Leaders should assure fairness: e.g., management bonuses and worker
pay cuts don’t mix.
There are any number of reasons why a firm may wish to communicate with
its customers. A primary reason is the evaluation of the customer’s perception
of the firm’s product and service quality and its impact on customer satisfaction.
The purpose may be to get an idea of the general condition of quality and satisfaction,
or a comparison of the current levels with the firm’s goals. A firm
might wish to conduct employee surveys and focus groups to assess the organization’s
quality structure.
Strategies for communicating with customers and
There are four primary strategies commonly used to obtain information
from or about customers and employees:
. sample surveys
. case studies
. ?eld experiments
. available data
With sample surveys, data are collected from a sample of a universe to
estimate the characteristics of the universe, such as their range or dispersion,
the frequency of occurrence of events, or the expected values of important
universe parameters. The reader should note that these terms are consistent
with the definition of enumerative statistical studies described in Chapter 9.
This is the traditional approach to such surveys. However, if survey results
are collected at regular intervals, the results can be analyzed using the quality
control tools described in Part II to obtain information on the underlying
process. The process excellence leader should not be reticent in recommending
that survey budgets be allocated to conducting small, routine, periodic
surveys rather than infrequent ‘‘big studies.’’ Without the information available
from time-ordered series of data, it will not be possible to learn about
processes which produce changes in customer satisfaction or perceptions of
A case study is an analytic description of the properties, processes, conditions,
or variable relationships of either single or multiple units under study.
Sample surveys and case studies are usually used to answer descriptive questions
(‘‘How do things look?’’) and normative questions (‘‘How well do things
compare with our requirements?’’). A field experiment seeks the answer to a
cause-and-effect question (‘‘Did the change result in the desired outcome?’’).
Use of available data as a strategy refers to the analysis of data previously collected
or available from other sources. Depending on the situation, available
data strategies can be used to answer all three types of questions: descriptive,
normative, and cause-and-effect. Original data collection strategies such as
mail questionnaires are often used in sample surveys, but they may also be
used in case studies and field experiments.
Survey development consists of the following major tasks (GAO, 1986, p. 15):
1. initial planning of the questionnaire
2. developing the measures
3. designing the sample
4. developing and testing the questionnaire
5. producing the questionnaire
6. preparing and distributing mailing materials
7. collecting data
8. reducing the data to forms that can be analyzed
9. analyzing the data
Figure 3.2 shows a typical timetable for the completion of these tasks.
The axiom that underlies the guidelines shown below is that the questionwriter(
s) must be thoroughly familiar with the respondent group and must
understand the subject matter from the perspective of the respondent group.
Surveys and Focus Groups 103
This is often problematic for the employee when the respondent group is the
customer; methods for dealing with this situation are discussed below. There
are eight basic guidelines for writing good questions:
1. Ask questions in a format that is appropriate to the questions’ purpose
and the information required.
2. Make sure the questions are relevant, proper, and quali?ed as needed.
3. Write clear, concise questions at the respondent’s language level.
4. Give the respondent a chance to answer by providing a comprehensive
list of relevant, mutually exclusive responses from which to choose.
5. Ask unbiased questions by using appropriate formats and item constructions
and by presenting all important factors in the proper sequence.
6. Get unbiased answers by anticipating and accounting for various
respondent tendencies.
7. Quantify the response measures where possible.
8. Provide a logical and unbiased line of inquiry to keep the reader’s
attention and make the response task easier.
The above guidelines apply to the form of the question. Using the critical
incident technique to develop good question content is described below.
Figure 3.2. Typical completion times for major questionnaire tasks.
From Developing and Using Questionnaires, Transfer Paper 7, p. 15,Washington, DC:
GAO Program Evaluation and Methodology Division.
There are several commonly used types of survey responses.
. Open-ended questionsLThese are questions that allow the
respondents to frame their own response without any restrictions
placed on the response. The primary advantage is that such questions
are easy to form and ask using natural language, even if the
question writer has little knowledge of the subject matter.
Unfortunately, there are many problems with analyzing the
answers received to this type of question. This type of question is
most useful in determining the scope and content of the survey,
not in producing results for analysis or process improvement.
. Fill-in-the-blank questionsLHere the respondent is provided with
directions that specify the units in which the respondent is to answer.
The instructions should be explicit and should specify the answer units.
This type of question should be reserved for very speci?c requests, e.g.,
‘‘What is your age on your last birthday?_____________ (age in years).’’
. Yes/No questionsLUnfortunately, yes/no questions are very popular.
Although they have some advantages, they have many problems and few
uses. Yes/no questions are ideal for dichotomous variables, such as defective
or not defective. However, too often this format is used when the measure
spans a range of values and conditions, e.g., ‘‘Were you satis?ed with
the quality of your new car (yes/no)?’’ A yes/no response to such questions
contains little useful information.
. Ranking questionsLThe ranking format is used to rank options according
to some criterion, e.g., importance. Ranking formats are di?cult to
write and di?cult to answer. They give very little real information and
are very prone to errors that can invalidate all the responses. They should
be avoided whenever possible in favor of more powerful formats and formats
less prone to error, such as rating. When used, the number of ranking
categories should not exceed ?ve.
. Rating questionsLWith this type of response, a rating is assigned on the
basis of the score’s absolute position within a range of possible values.
Rating scales are easy to write, easy to answer, and provide a level of quanti
?cation that is adequate for most purposes. They tend to produce reasonably
valid measures. Here is an example of a rating format:
For the following statement, checkthe appropriate box:
The workmanship standards provided by the purchaser are
& Clear
& Marginally adequate
& Unclear
Surveys and Focus Groups 105
. Guttman formatLIn the Guttman format, the alternatives increase in
comprehensiveness; that is, the higher-valued alternatives include the
lower-valued alternatives. For example,
Regarding the benefit received from training in quality improvement:
& No bene?t identi?ed
& Identi?ed bene?t
& Measured bene?t
& Assessed bene?t value in dollar terms
& Performed cost/bene?t analysis
. Likert and other intensity scale formatsLThese formats are usually
used to measure the strength of an attitude or an opinion. For example,
Please checkthe appropriate box in response to the following statement:
‘‘The customer service representative was knowledgeable.’’
& Strongly disagree
& Disagree
& Neutral
& Agree
& Strongly agree
Intensity scales are very easy to construct. They are best used when
respondents can agree or disagree with a statement. A problem is that
statements must be worded to present a single side of an argument. We
know that the respondent agrees, but we must infer what he believes. To
compensate for the natural tendency of people to agree, statements are
usually presented using the converse as well, e.g., ‘‘The customer service
representative was not knowledgeable.’’
When using intensity scales, use an odd-numbered scale, preferably
with ?ve or seven categories. If there is a possibility of bias, order the
scale in a way that favors the hypothesis you want to disprove and handicaps
the hypothesis you want to con?rm. This way you will con?rm the
hypothesis with the bias against youLa stronger result. If there is no
bias, put the most undesirable choices ?rst.
. Semantic di?erential formatLIn this format, the values that span the
range of possible choices are not completely identi?ed; only the end
points are labeled. For example,
Indicate the number of times you initiated communication with your customer
in the past month.
few & & & & & & & many
The respondent must infer that the range is divided into equal intervals.
The range seems to work well with seven categories.
Semantic di?erentials are very usefulwhenwedo not have enough information
to anchor the intervals between the poles. However, they are very
di?cult to write well and if not written well the results are ambiguous.
This actual case study involves the development of a mail survey at a community
hospital. The same process has been successfully used by the author to develop
customer surveys for clientele in a variety of industries.
The study of service quality and patient satisfaction was performed at a 213 bed
community hospital in the southwestern United States. The hospital is a nonpro
?t, publicly funded institution providing services to the adult community;
pediatric services are not provided. The purpose of the study was to:
1. Identify the determinants of patient quality judgments.
2. Identify internal service delivery processes that impacted patient quality
3. Determine the linkage between patient quality judgments and intent-topatronize
the hospital in the future or to recommend the hospital to
To conduct the study, the author worked closely with a core team of hospital
employees, and with several ad hoc teams of hospital employees. The core team
included the Nursing Administrator, the head of the Quality Management
Department, and the head of Nutrition Services.**
The team decided to develop their criteria independently. It was agreed that
the best method of getting information was directly from the target group, inpatients.
Due to the nature of hospital care services, focus groups were not
deemed feasible for this study. Frequently, patients must spend a considerable
period of time convalescing after being released from a hospital, making it
impossible for them to participate in a focus group soon after discharge. While
the patients are in the hospital, they are usually too sick to participate. Some
patients have communicable diseases, which makes their participation in focus
groups inadvisable.
Since memories of events tend to fade quickly (Flanagan, 1954, p. 331), the team
decided that patients should be interviewed within 72 hours of discharge. The
Surveys and Focus Groups 107
*The survey referenced by this case study is located in the Appendix.
**The nutrition services manager was very concerned that she get sufficient detail on her particular service. Thus, the critical
incident interview instrument she used included special questions relating to food service.
target patient population was, therefore, all adults treated as in-patients and discharged
to their homes. The following groups were not part of the study: families
of patients who died while in the hospital, patients discharged to nursing homes,
patients admitted for psychiatric care.*
The team used the Critical Incident Technique (CIT) to obtain patient comments.
The CIT was ?rst used to study procedures for selection and classi?cation
of pilot candidates in World War II (Flanagan, 1954). A bibliography assembled
in 1980 listed over seven hundred studies about or using the CIT (Fivars, 1980).
Given its popularity, it is not surprising that the CIT has also been used to evaluate
service quality.
CIT consists of a set of speci?cally de?ned procedures for collecting observations
of human behavior in such a way as to make them useful in addressing practical
problems. Its strength lies in carefully structured data collection and data
classi?cation procedures that produce detailed information not available through
other research methods. The technique, using either direct observation or recalled
information collected via interviews, enables researchers to gather ?rsthand
patient-perspective information. This kind of self-report preserves the richness of
detail and the authenticity of personal experience of those closest to the activity
being studied. Researchers have concluded that the CIT produces information
that is both reliable and valid.
This study attempted to follow closely the ?ve steps described by Flanagan as
crucial to the CIT: 1) establishment of the general aim of the activity studied;
2) development of a plan for observers or interviewers; 3) collection of data; 4) analysis
(classi?cation) of data; and 5) interpretation of data.
Establishment of the general aim of the activity studied
The general aim is the purpose of the activity. In this case the activity
involves the whole range of services provided to in-patients in the hospital.
This includes every service activity between admission and discharge.** From
the service provider’s perspective the general aim is to create and manage service
delivery processes in such a way as to produce a willingness by the patient to
utilize the provider’s services in the future. To do this the service provider
must know which particular aspects of the service are remembered by the
Our general aim was to provide the service provider with information on what
patients remembered about their hospital stay, both pleasant and unpleasant.
This information was to be used to construct a new patient survey instrument
*The team was unable to obtain a Spanish-speaking interviewer, which meant that some patients that were candidates were
not able to participate in the survey.
**Billing was not covered in the CIT phase of the study because patients had not received their bills within 72 hours.
that would be sent to recently discharged patients on a periodic basis. The information
obtained would be used by the managers of the various service processes as
feedback on their performance, from the patient’s perspective.
Interview plan
Interviewers were provided with a list of patients discharged within the past 3
days. The discharge list included all patients. Non-psychiatric patients who
were discharged to ‘‘home’’ were candidates for the interview. Home was de?ned
as any location other than the morgue or a nursing home. Interviewers were
instructed to read a set of predetermined statements. Patients to be called were
selected at random from the discharge list. If a patient could not be reached,
the interviewer would try again later in the day. One interview form was prepared
per patient. To avoid bias, 50% of the interview forms asked the patient
to recall unpleasant incidents ?rst and 50% asked for pleasant incidents ?rst.
Interviewers were instructed to record the patient responses using the patient’s
own words.
Collection of data
Four interviewers participated in the data collection activity, all were management
level employees of the hospital. Three of the interviewers were female, one
was male. The interviews were conducted when time permitted during the interviewer’s
normal busy work day. The interviews took place during the September
1993 time period. Interviewers were given the instructions recommended by
Hayes (1992, pp. 14^15) for generating critical incidents.
A total of 36 telephone attempts were made and 23 patients were reached. Of
those reached, three spoke only Spanish. In the case of one of the Spanish-speaking
patients a family member was interviewed. Thus, 21 interviews were conducted,
which is slightly greater than the 10 to 20 interviews recommended by Hayes
(1992, p. 14). The 21 interviews produced 93 critical incidents.
Classi?cation of data
The Incident Classi?cation System required by CIT is a rigorous, carefully
designed procedure with the end goal being to make the data useful to the problem
at hand while sacri?cing as little detail as possible (Flanagan, 1954, p. 344). There
are three issues in doing so: 1) identi?cation of a general framework of reference
that will account for all incidents; 2) inductive development of major area and subarea
categories that will be useful in sorting the incidents; and 3) selection of the
most appropriate level of speci?city for reporting the data.
The critical incidents were classi?ed as follows:
1. Each critical incident was written on a 35 card, using the patient’s own
Surveys and Focus Groups 109
2. The cards were thoroughly shu?ed.
3. Ten percent of the cards (10 cards) were selected at random, removed from
the deck and set aside.
4. Two of the four team members left the room while the other two grouped the
remaining 83 cards and named the categories.
5. The ten cards originally set aside were placed into the categories found in
step 4.
6. Finally, the two members not involved in the initial classi?cation were told
the names of the categories. They then took the reshu?ed 93 cards and
placed them into the previously determined categories.
The above process produced the following dimensions of critical incidents:
. Accommodations (5 critical incidents)
. Quality of physician (14 critical incidents)
. Care provided by sta? (20 critical incidents)
. Food (26 critical incidents)
. Discharge process (1 critical incident)
. Attitude of sta? (16 critical incidents)
. General (11 critical incidents)
Interpretation of data
Interjudge agreement, the percentage of critical incidents placed in the same
category by both groups of judges, was 93.5%. This is well above the 80% cuto?
value recommended by experts. The setting aside of a random sample and trying
to place them in established categories is designed to test the comprehensiveness
of the categories. If any of the withheld items were not classi?able it would be an
indication that the categories do not adequately span the patient satisfaction
space. However, the team experienced no problem in placing the withheld critical
incidents into the categories.
Ideally, a critical incident has two characteristics: 1) it is speci?c and 2) it
describes the service provider in behavioral terms or the service product with speci-
?c adjectives (Hayes, 1992, p. 13). Upon reviewing the critical incidents in the
General category, the team determined that these items failed to have one or both
of these characteristics. Thus, the 11 critical incidents in the General category
were dropped. The team also decided to merge the two categories ‘‘Care provided
by sta?’’ and ‘‘Attitude of sta?’’ into the single category ‘‘Quality of sta? care.’’
Thus, the ?nal result was a ?ve dimension model of patient satisfaction judgments:
Food, Quality of physician, Quality of sta? care, Accommodations, and
Discharge process.
A rather obvious omission in the above list is billing. This occurred because the
patients had not yet received their bill within the 72 hour time frame. However,
the patient’s bill was explained to the patient prior to discharge. This item is
included in the Discharge process dimension. The team discussed the billing issue
and it was determined that billing complaints do arise after the bills are sent, suggesting
that billing probably is a satisfaction dimension. However, the team
decided not to include billing as a survey dimension because 1) the time lag was so
long that waiting until bills had been received would signi?cantly reduce the ability
of the patient to recall the details of their stay; 2) fear that the patient’s judgments
would be overwhelmed by the recent receipt of the bill; and 3) a system
already existed for identifying patient billing issues and adjusting the billing
process accordingly.
Survey item development
As stated earlier, the general aim was to provide the service provider with
information on what patients remembered about their hospital stay, both pleasant
and unpleasant. This information was then to be used to construct a new
patient survey instrument that would be sent to recently discharged patients on
a periodic basis. The information obtained would be used by the managers of
the various service processes as feedback on their performance, from the
patient’s perspective.
The core team believed that accomplishing these goals required that the managers
of key service processes be actively involved in the creation of the survey
instrument. Thus, ad hoc teams were formed to develop survey items for each of
the dimensions determined by the critical incident study. The teams were given
brief instruction by the author in the characteristics of good survey items. Teams
were required to develop items that, in the opinion of the core team, met ?ve criteria:
1) relevance to the dimension being measured; 2) concise; 3) unambiguous;
4) one thought per item; and 5) no double negatives. Teams were also shown the
speci?c patient comments that were used as the basis for the categories and
informed that these comments could be used as the basis for developing survey
Writing items for the questionnaire can be di?cult. The process of developing
the survey items involved an average of three meetings per dimension, with each
meeting lasting approximately two hours. Ad hoc teams ranged in size from four
to eleven members. The process was often quite tedious, with considerable debate
over the precise wording of each item.
The core team discussed the scale to be used with each ad hoc team. The core
team’s recommended response format was a ?ve point Likert-type scale. The consensus
was to use a ?ve point agree-disagree continuum as the response format.
Item wording was done in such a way that agreement represented better performance
from the hospital’s perspective.
In addition to the response items, it was felt that patients should have an
opportunity to respond to open-ended questions. Thus, the survey also included
Surveys and Focus Groups 111
general questions that invited patients to comment in their own words. The bene-
?ts of having such questions is well known. In addition, it was felt that these questions
might generate additional critical incidents that would be useful in
validating the survey.
The resulting survey instrument contained 50 items and three open-ended questions
and is included in the Appendix.
Survey administration and pilot study
The survey was to be tested on a small sample. It was decided to use the total
design method (TDM) to administer the survey (Dillman, 1983). Although the
total design method is exacting and tedious, Dillman indicated that its use would
assure a high rate of response. Survey administration would be handled by the
Nursing Department.
TDM involves rather onerous administrative processing. Each survey form is
accompanied by a cover letter, which must be hand-signed in blue ink. Follow up
mailings are done 1, 3 and 7 weeks after the initial mailing. The 3 and 7 week follow
ups are accompanied by another survey and another cover letter. No ‘‘bulk processing’’
is allowed, such as the use of computer-generated letters or mailing labels.
Dillman’s research emphasizes the importance of viewing the TDM as a completely
integrated approach (Dillman, 1983, p. 361).
Because the hospital in the study is small, the author was interested in obtaining
maximum response rates. In addition to following the TDM guidelines, he recommended
that a $1 incentive be included with each survey. However, the hospital
administratorwas not convinced that the additional $1 per survey was worthwhile.
It was ?nally agreed that to test the e?ect of the incentive on the return rate $1
would be included in 50% of the mailings, randomly selected.
The hospital decided to perform a pilot study of 100 patients. The patients
selected were the ?rst 100 patients discharged to home starting April 1, 1994. The
return information is shown in Table 3.2.
Although the overall return rate of 49% is excellent for normal mail-survey
procedures, it is substantially below the 77% average and the 60% ‘‘minimum’’
reported by Dillman. As possible explanations, the author conjectures that
there may be a large Spanish-speaking constituency for this hospital. As mentioned
above, the hospital is planning a Spanish version of the survey for the
The survey respondent demographics were analyzed and compared to the
demographics of the non-respondents to assure that the sample group was representative.
A sophisticated statistical analysis was performed on the responses to
evaluate the reliability and validity of each item. Items with low reliability coe?-
cients or questionable validity were reworded or dropped.
Focus groups
The focus group is a special type of group in terms of purpose, size, composition,
and procedures. A focus group is typically composed of seven to ten participants
who are unfamiliar with each other. These participants are selected
because they have certain characteristic(s) in common that relate to the topic
of the focus group.
The researcher creates a permissive environment in the focus group that
nurtures different perceptions and points of view, without pressuring participants
to vote, plan, or reach consensus. The group discussion is conducted
several times with similar types of participants to identify trends and patterns
in perceptions. Careful and systematic analyses of the discussions provide
clues and insights as to how a product, service, or opportunity is
A focus group can thus be defined as a carefully planned discussion
designed to obtain perceptions on a defined area of interest in a permissive,
non-threatening environment. The discussion is relaxed, comfortable, and
often enjoyable for participants as they share their ideas and perceptions.
Group members influence each other by responding to ideas and comments
in the discussion.
In Six Sigma, focus groups are useful in a variety of situations:
. prior to starting the strategic planning process
. generate information for survey questionnaires
Surveys and Focus Groups 113
Table 3.2. Pilot patient survey return information.
A. NUMBERS incentive: 55%
Surveys mailed: 100 Number delivered that had no $1
Surveys delivered: 92 incentive: 45
Surveys returned as undeliverable: 8 Number returned that had no $1
Survey returned, needed Spanish
version: 1
incentive: 19
Percentage returned that had no $1
incentive: 42%
Total surveys returned: 45
Percentage of surveys delivered
returned: 49%
Number of surveys returned after:
Number delivered that had $1
incentive: 47 Initial mailing: 12
Number returned that had $1 One week follow up: 16
incentive: 26 Three week follow up: 8
Percentage returned that had $1 Seven week follow up: 9
. needs assessment, e.g., training needs
. test new program ideas
. determine customer decision criteria
. recruit new customers
The focus group is a socially oriented research procedure. The advantage of
this approach is that members stimulate one another, which may produce a
greater number of comments than would individual interviews. If necessary,
the researcher can probe for additional information or clarification. Focus
groups produce results that have high face validity, i.e., the results are in the participant’s
own words rather than in statistical jargon. The information is
obtained at a relatively low cost, and can be obtained very quickly.
There is less control in a group setting than with individual interviews. When
group members interact, it is often difficult to analyze the resulting dialogue.
The quality of focus group research is highly dependent on the qualifications
of the interviewer. Trained and skilled interviewers are hard to find. Group-togroup
variation can be considerable, further complicating the analysis. Finally,
focus groups are often difficult to schedule.
Other customer information systems
. Complaint and suggestion systems typically provide all customers with an
easy-to-use method of providing favorable or unfavorable feedback to
management. Due to selection bias, these methods do not provide statistically
valid information. However, because they are a census rather than a
sample, they provide opportunities for individual customers to have
their say. These are moments of truth that can be used to increase customer
loyalty. They also provide anecdotes that have high face validity
and are often a source of ideas for improvement.
. Customer panels are composed of a representative group of customers
who agree to communicate their attitudes periodically via phone calls or
mail questionnaires. These panels are more representative of the range of
customer attitudes than customer complaint and suggestion systems. To
be e?ective, the identity of customers on the panel must be withheld
from the employees serving them.
. Mystery shoppers are employees who interact with the system as do real
Once customer feedback has been obtained, it must be used to improve process
and product quality. A system for utilizing customer feedback is shown in
Figure 3.3.
1. Local managers and employees serve customers’ needs on a daily basis,
using locally modi?ed procedures along with general corporate policies
and procedures.
2. By means of a standardized and locally sensitive questionnaire, determine
the needs and attitudes of customers on a regular basis.
3. Comparing ?nancial data, expectations, and past attitude information,
determine strengths and weaknesses and their probable causes.
4. Determine where and how e?ort should be applied to correct weaknesses
and preserve strengths. Repeat the process by taking actionL
step 1Land maintain it to attain a steady state or to evolve in terms of
customer changes.
Surveys and Focus Groups 115
Figure 3.3. System for utilizing customer feedback.
From Daltas, A.J. (1977). ‘‘Protecting service markets with consumer feedback,’’
Cornell Hotel and Restaurant Administration Quarterly, May, pp. 73^77.
5. A similar process can take place at higher levels, using aggregated data
from the ?eld and existing policy ?ows of the organization.
Although this system was developed by marketing specialists, note that it
incorporates a variation of the classical Shewhart quality improvement PDCA
(Plan-Do-Check-Act) cycle (see Chapter 7).
Customers have value. This simple fact is obvious when one looks at a
customer making a single purchase. The transaction provides revenue and
profit to the firm. However, when the customer places a demand on the
firm, such as a return of a previous purchase or a call for technical support,
there is a natural tendency to see this as a loss. At these times it is important
to understand that customer value must not be viewed on a shortterm
transaction-by-transaction basis. Customer value must be measured
over the lifetime of the relationship. One method of calculating the lifetime
value of a loyal customer, based on work by Frederick Reichheld of Bain
and Co. and the University of Michigan’s Claes Fornell, is as follows
(Stewart, 1995):
1. Decide on ameaningful period of time over which to do the calculations.
This will vary depending on your planning cycles and your business: A
life insurer should track customers for decades, a disposable diaper
maker for just a few years, for example.
2. Calculate the pro?t (net cash ?ow) customers generate each year.
Track several samplesLsome newcomers, some old-timersLto ?nd
out how much business they gave you each year, and how much it
cost to serve them. If possible, segment them by age, income, sales
channel, and so on. For the ?rst year, be sure to subtract the cost of
acquiring the pool of customers, such as advertising, commissions,
back-o?ce costs of setting up a new account. Get speci?c numbersL
pro?t per customer in year one, year two, etc.Lnot averages for all
customers or all years. Long-term customers tend to buy more, pay
more (newcomers are often lured by discounts), and create less bad
3. Chart the customer ‘‘life expectancy,’’ using the samples to ?nd out how
much your customer base erodes each year. Again, speci?c ?gures are
better than an average like ‘‘10% a year’’; old customers are much less
likely to leave than freshmen. In retail banking, 26% of account holders
defect in the ?rst year; in the ninth year, the rate drops to 9%.
4. Once you know the pro?t per customer per year and the customerretention
?gures, it’s simple to calculate net present value (NPV).
Pick a discount rateLif you want a 15% annual return on assets, use
that. In year one, the NPV will be pro?t 	 1.15. Next year, NPV =
(year-two pro?t  retention rate) 	 (1.15)2. In year n, the last year in
your ?gures, the NPV is the n year’s adjusted pro?t 	 (1.15)n. The
sum of the years one through n is how much your customer is
worthLthe net present value of all the pro?ts you can expect from
his tenure.
This is very valuable information. It can be used to find out how much to
spend to attract new customers, and which ones. Better still, you can exploit
the leverage customer satisfaction offers. Take your figures and calculate how
much more customers would be worth if you increased retention by 5%.
Figure 3.4 shows the increase in customer NPV for a 5% increase in retention
for three industries.
Once the lifetime value of the customer is known, it forms the basis of loyalty-
based managementSM of the customer relationship. According to
Reichheld (1996), loyalty-based management is the practice of carefully selecting
customers, employees, and investors, and then working hard to retain
them. There is a tight, cause-and-effect connection between investor, employee
and customer loyalty. These are the human assets of the firm.
Calculating the Value of Retention of Customers 117
Figure 3.4. Increase in customer NPV for a 5% increase in customer retention.
Complaint handling
When a customer complaint has been received it represents an opportunity
to increase customer loyalty, and a risk of losing the customer. The way the complaint
is handled is crucial. The importance of complaint handling is illustrated
in Figure 3.5. These data illustrate that the decision as to whether a customer
who complains plans to repurchase is highly dependent on how well they felt
their complaint was handled. Add to this the fact that customers who complain
are likely to tell as many as 14 others of their experience, and the importance
of complaint handling in customer relations becomes obvious.
Despite the impressive nature of Figure 3.5, even these figures dramatically
understate the true extent of the problem. Complaints represent people who
were not only unhappy, they notified the company. Research indicates that up
to 96% of unhappy customers never tell the company. This is especially unfortunate
since it has been shown that customer loyalty is increased by proper resolution
of complaints. Given the dramatic impact of a lost customer, it makes
sense to maximize the opportunity of the customer to complain. Complaints
should be actively sought, an activity decidedly against human nature. This suggests
that a system must be developed and implemented to force employees to
seek out customer complaints. In addition to actively soliciting customer complaints,
the system should also provide every conceivable way for an unhappy
Figure 3.5. Percent planning to repurchase vs. how complaint was handled. (*Note: The
large durable goods survey did not include a response category of ‘‘satis?ed.’’)
customer to contact the company on their own, including toll-free hotlines,
email, comment cards, etc.
Customer expectations, priorities, needs, and
Although customers seldom spark true innovation (for example, they are
usually unaware of state-of-the art developments), their input is extremely valuable.
Obtaining valid customer input is a science itself. Market research firms
use scientific methods such as critical incident analysis, focus groups, content
analysis and surveys to identify the ‘‘voice of the customer.’’ Noritaki Kano
developed the following model of the relationship between customer satisfaction
and quality (Figure 3.6).
The Kano model shows that there is a basic level of quality that customers
assume the product will have. For example, all automobiles have windows and
tires. If asked, customers don’t even mention the basic quality items, they take
them for granted. However, if this quality level isn’t met the customer will be
dissatisfied; note that the entire ‘‘Basic Quality’’ curve lies in the lower half of
the chart, representing dissatisfaction. However, providing basic quality isn’t
enough to create a satisfied customer.
Kano Model of Customer Expectations 119
Figure 3.6. Kano model.
The ‘‘Expected Quality’’ line represents those expectations which customers
explicitly consider. For example, the length of time spent waiting in line at a
checkout counter. The model shows that customers will be dissatisfied if their
quality expectations are not met; satisfaction increases as more expectations
are met.
The ‘‘Exciting Quality’’ curve lies entirely in the satisfaction region. This is
the effect of innovation. Exciting quality represents unexpected quality items.
The customer receives more than they expected. For example, Cadillac pioneered
a system where the headlights stay on long enough for the owner to
walk safely to the door. When first introduced, the feature excited people.
Competitive pressure will constantly raise customer expectations. Today’s
exciting quality is tomorrow’s basic quality. Firms that seek to lead the market
must innovate constantly. Conversely, firms that seek to offer standard
quality must constantly research customer expectations to determine the
currently accepted quality levels. It is not enough to track competitors since
expectations are influenced by outside factors as well. For example, the
quality revolution in manufacturing has raised expectations for service quality
as well.
Garden variety Six Sigma only addresses half of
the Kano customer satisfaction model
Some people, including your author, believe that even Six Sigma doesn’t go
far enough. In fact, even ‘‘zero defects’’ falls short. Defining quality as only the
lack of non-conforming product reflects a limited view of quality. Motorola, of
course, never intended to define quality as merely the absence of defects.
However, some have misinterpreted the Six Sigma program in this way.
One problem with ‘‘garden variety’’ Six Sigma is that it addresses only half
of the Kano model. By focusing on customer expectations and prevention of
non-conformances and defects, Six Sigma addresses the portion of the Kano
model on and below the line labeled ‘‘Expected Quality.’’ While there is nothing
wrong with improving these aspects of business performance, it will not
assure that the organization remains viable in the long term. Long-term success
requires that the organization innovate. Innovation is the result of creative
activity, not analysis. Creativity is not something that can be done ‘‘by
the numbers.’’ In fact, excessive attention to a rigorous process such as Six
Sigma can detract from creative activities if not handled carefully. As discussed
above, the creative organization is one which exhibits variability,
redundancy, quirky design, and slack. It is vital that the organization keep
this Paradox in mind.
Once information about customer expectations has been obtained, techniques
such as quality function deployment (QFD) can be used to link the
voice of the customer directly to internal processes.
Tactical quality planning involves developing an approach to implementing
the strategic quality plan. One of the most promising developments in this
area has been policy deployment. Sheridan (1993) describes policy deployment
as the development of a measurement-based system as a means of planning for
continuous quality improvement throughout all levels of an organization.
Originally developed by the Japanese, American companies also use policy
deployment because it clearly defines the long-range direction of company
development, as opposed to short-term.
QFD is a customer-driven process for planning products and services. It
starts with the voice of the customer, which becomes the basis for setting
requirements. QFD matrices, sometimes called ‘‘the house of quality,’’ are
graphical displays of the result of the planning process. QFD matrices
vary a great deal and may show such things as competitive targets and
process priorities. The matrices are created by interdepartmental teams,
thus overcoming some of the barriers which exist in functionally organized
QFD is also a system for design of a product or service based on customer
demands, a system that moves methodically from customer requirements to
specifications for the product or service. QFD involves the entire company in
the design and control activity. Finally, QFD provides documentation for the
decision-making process. The QFD approach involves four distinct phases
(King, 1987):
Organization phaseLManagement selects the product or service to be
improved, the appropriate interdepartmental team, and de?nes the focus
of the QFD study.
Descriptive phaseLThe team de?nes the product or service from several different
directions such as customer demands, functions, parts, reliability,
cost, and so on.
Breakthrough phaseLThe team selects areas for improvement and ?nds
ways to make them better through new technology, new concepts, better
reliability, cost reduction, etc., and monitors the bottleneck process.
Implementation phaseLThe team de?nes the new product and how it will be
QFD is implemented through the development of a series of matrices. In its
simplest form QFD involves a matrix that presents customer requirements as
rows and product or service features as columns. The cell, where the row and
Quality Function Deployment (QFD) 121
column intersect, shows the correlation between the individual customer
requirement and the product or service requirement. This matrix is sometimes
called the ‘‘requirement matrix.’’ When the requirement matrix is enhanced by
showing the correlation of the columns with one another, the result is called
the ‘‘house of quality.’’ Figure 3.7 shows one commonly used house of quality
The house of quality relates, in a simple graphical format, customer requirements,
product characteristics, and competitive analysis. It is crucial that this
matrix be developed carefully since it becomes the basis of the entire QFD process.
By using the QFD approach, the customer’s demands are ‘‘deployed’’ to
the final process and product requirements.
One rendition of QFD, called the Macabe approach, proceeds by developing
a series of four related matrices (King, 1987): product planning matrix, part
deployment matrix, process planning matrix, and production planning matrix.
Each matrix is related to the previous matrix as shown in Figure 3.8.
Figure 3.9 shows an example of an actual QFD matrix.
Figure 3.7. The house of quality.
Data collection and review of customer
expectations, needs, requirements, and
Another approach to QFD is based on work done by Yoji Akao. Akao (1990,
pp. 7^8) presents the following 11-step plan for developing the quality plan
and quality design, using QFD.
1. First, survey both the expressed and latent quality demands of consumers
in your target marketplace. Then decide what kinds of ‘‘things’’
to make.
2. Study the other important characteristics of your target market and
make a demanded quality function deployment chart that re?ects
both the demands and characteristics of that market.
3. Conduct an analysis of competing products on the market, which we
call a competitive analysis. Develop a quality plan and determine the
selling features (sales points).
Quality Function Deployment (QFD) 123
Figure 3.8. QFD methodology: Macabe approach.
Figure 3.9. QFD matrix for an aerospace ?rm.
From Wahl, P.R. and Bersbach, P.L. (1991), ‘‘TQM AppliedLCradle to Grave,’’
ASQ 45th Quality Congress Transactions. Reprinted with permission.
4. Determine the degree of importance of each demanded quality.
5. List the quality elements and make a quality elements deployment
6. Make a quality chart by combining the demanded quality deployment
chart and the quality elements deployment chart.
7. Conduct an analysis of competing products to see how other companies
perform in relation to each of these quality elements.
8. Analyze customer complaints.
9. Determine the most important quality elements as indicated by customer
quality demands and complaints.
10. Determine the speci?c design quality by studying the quality characteristics
and converting them into quality elements.
11. Determine the quality assurance method and the test methods.
I am often asked ‘‘Will Six Sigma work for. . .’’ where the blank is ‘‘health
care,’’ ‘‘oil exploration,’’ ‘‘custom-built homes,’’ etc. The list is unending. My
typical response is that, if a process is involved, Six Sigma may be able to help
you improve it. Personally, I don’t believe that everything will benefit from the
application of Six Sigma rigor. There are some things that aren’t processes,
such as pure creativity, love and unstructured play. I don’t believe a chess grandmaster
would benefit from the advice of a Black Belt applying DMAIC to his
moves, nor would his equivalent in the R&D area. There are other things that
are processes, but processes so poorly understood that we don’t know enough
about them to use the Six Sigma approach to improve them, such as pure
research, social relationships, criminal behavior, or curing substance abuse.
However, the vast majority of processes encountered in business, non-profit
organizations, and government services fall into the category of processes that
can be improved by the application of Six Sigma methods.
But what exactly is a ‘‘process’’? There is a tendency to narrowly interpret the
term process to refer to a manufacturing operation that converts raw materials
into finished products. That’s true, of course. But as I use the term process
throughout this book it has a much broader meaning. In this book process refers
to any activity or set of activities that transform inputs to create values for stakeholders.
The inputs can be labor, expertise, raw materials, products, transactions,
or services that someone is willing to pay more for than they cost to
create. In other words, the process adds value to the inputs. Said another way,
the process is the act of creating value. The value can be a cured disease, a tasty
banana split, a great movie, a successfully completed credit card transaction,
or a cold soda purchased at a convenience store.
The Six Sigma Process Enterprise 125
Reengineering, the process redesign fad so popular in the early 1990s, has
become associated in the minds of many with brutal downsizing. Many academics
condemned it as heartless and cold. But the problem wasn’t caused
by reengineering. Reengineering (and Six Sigma) focus attention on broken
and inefficient processes. The truth is, this focus enabled companies to operate
faster and more efficiently and to use information technology more productively.
It gave employees more authority and a clearer view of how their
work fit into the broader scheme of things. Customers benefited from lower
prices, higher quality and better services, and investors enjoyed a higher rate
of return. And, more germane to our discussion of processes, reengineering
taught business leaders to see their organizations not as control structures,
but as processes that deliver value to customers in a way that creates profits
for shareholders.
Examples of processes
Many business leaders think of their organizations as extremely complex.
From a process perspective, this is seldom the case, at least at the high levels.
For example, Texas Instruments was able to break its $4 billion semiconductor
business into six core processes:
1. Strategy development.
2. Product development.
3. Customer design and support.
4. Manufacturing capability development.
5. Customer communication.
6. Order ful?llment.
A large financial software company described its four core processes in plain
English as:
1. Provide good products at good prices.
2. Acquire customers and maintain good relations with them.
3. Make it easy to buy from us.
4. Provide excellent service and support after the sale.
Both of these companies have thousands of employees and generate billions
of dollars in sales. Yet what they do for customers is really very simple. Once
the basic (core) processes have been identified, the relationship between them
should be determined and drawn on a process map. (Process mapping is discussed
in greater detail in Part II of this handbook.) The process map presents
employees with a simple picture that illustrates how the enterprise serves its
customers. It is the basis for identifying subprocesses and, eventually, Six
Sigma projects. Table 3.3 gives some examples of high-level processes and subprocesses.
The truth is, it’s the organizational structure that’s complicated, not the
business itself. The belief that the business is complicated results from a misplaced
internal perspective by its leaders and employees. In a traditional
organization tremendous effort is wasted trying to understand what needs
to be done when goals are not well defined and people don’t know how
their work relates to the organization’s purpose. A process focus is often
the first real ‘‘focus’’ an employee experiences, other than pleasing one’s
The Six Sigma Process Enterprise 127
Table 3.3. Examples of high-level processes and subprocesses.
Core Process Subprocess
Product development . R&D
. Design creation
. Prototype development
. Design production support
Marketing . Inspiration, concept discovery
. Customer identi?cation
. Developing market strategies
. Concept production support
. Customer acquisition and maintenance
Product creation . Manufacturing
. Procurement
. Installation
Sales and service . Ful?llment (order cpayment)
. Pre-sale customer support
. Installation and front-line service
. Usage
Meta-processes . Process excellence (Six Sigma)
. Voice of customer
. Voice of shareholder
. Voice of employee
The source of conflict
Management structures, since the time of Alfred P. Sloan in the 1920s and
1930s, are designed to divide work into discrete units with clear lines of responsibility
and authority. While this approach produced excellent results for a
time, it has inherent flaws that became quite apparent by 1980. Organizations
put leadership atop a pyramid-shaped control system designed to carry out
their strategies. Control of the resources needed to accomplish this resided in
the vertical pillars, known as ‘‘functions’’ or ‘‘divisions.’’ This command-andcontrol
approach is depicted in Figure 3.10.
This arrangement creates ‘‘turf’’ where, much like caste systems, activities
within a given area are the exclusive domain of that area. Personnel in engineering,
for example, are not allowed to engage in activities reserved to the finance
group, nor is finance allowed to ‘‘meddle’’ in engineering activities. These turfs
are jealously guarded. In such a structure employees look to the leadership to
tell them what to do and to obtain the resources needed to do it. This upwardinward
focus is the antithesis of an external-customer focus. As Figure 3.10 also
shows, customer value is created by processes that draw resources from several
different parts of the organization and end at a customer contact point. If an organization
wants to be customer-focused, then it must change the traditional structure
so its employees look across the organization at processes. As you might
expect, this calls for a radical rethinking of the way the enterprise operates.
As long as control of resources and turf remain entirely with the functional
units, the enterprise will remain focused inwardly. Goals will be unit-based,
rather than process-based. In short, Six Sigma (or any other process-oriented
initiative) will not work. Functional department leaders have both the incentive
and the ability to thwart cross-functional process improvement efforts. This
doesn’t mean that these people are ‘‘bad.’’ It’s simply that their missions are
Figure 3.10. Traditional command-and-control organizational structure.
defined in such a way that they are faced with a dilemma: pursue the mission
assigned to my area to the best of my ability, or support an initiative that
detracts from it but benefits the enterprise as a whole. Social scientists call this
‘‘the tragedy of the commons.’’ It is in the best interest of all fishermen not to
overharvest the fishing grounds, but it is in the interest of each individual fisherman
to get all he can from this ‘‘common resource.’’ Similarly, it is in the best
interest of the enterprise as a whole to focus on customers, but it is in each functional
leader’s best interest to pursue his or her provincial self-interest. After
all, if every other functional manager tries to maximize the resources devoted
to their area and I don’t, I’ll lose my department’s share of the resources. Selfinterest
wins hands down.
A resolution to the conflict
Some companiesLsuch as IBM, Texas Instruments, Owens Corning, and
DukePowerLhavesuccessfully madethe transition from the traditional organizational
structure to an alternative system called the ‘‘Process Enterprise’’
(Hammer and Stanton, 1999). In these companies the primary organizational
unit is not the functional department, but the process development team. These
cross-functional teams, like the reengineering teams of old, have full responsibility
for a major business process. For example, a product development team
would work together in the same location to build the product development process
from concept to launch. They would produce the design, documentation,
training materials, advertising, etc. In a Process Enterprise authority and control
of resources is redistributed in a manner that achieves a balance of power
between the process-focused and structure-focused aspects of the enterprise.
ThedifferencesbetweenProcessEnterprisesandtraditional organizations are
fundamental. In the Process Enterprise a new position is created, that of Process
Owner or Business Process Executive (BPE). The BPE position is permanent.
BPEs are assigned from the senior-most executive body and given responsibility
for designing and deploying the process, as well as control over all expenditures
and supporting technology. They establish performance metrics, set and distribute
budgets, and train the front-line workers who perform the process work.
However, the people who perform the process work report to unit heads, not
BPEs. In the Process Enterprise process goals are emphasized over unit goals.
Process performance is used as a basis for compensation and advancement.
In a Process Enterprise lines of authority are less well defined. BPEs and
functional unit managers are expected to work together to resolve disagree-
The Six Sigma Process Enterprise 129
ments. The BPE doesn’t exert direct control over the workers, but because he
controls budgets and sets goals by which unit managers will be evaluated, he
does have a good deal of influence. The unit managers have to see to it that the
process designs are sound, the resource allocation sufficient, and the goals
clear and fair. In short, managing in a Process Enterprise places a premium on
collaboration and cooperation.
One tool that has been developed to help clarify the different roles and
responsibilities is the Decision Rights Matrix (Hammer and Stanton, 1999,
p. 113). This matrix specifies the roles the different managers play for each
major decision, such as process changes, personnel hiring, setting budgets, and
so on. For example, on a given decision must a given manager:
. Make the decision?
. Be noti?ed in advance?
. Be consulted beforehand?
. Be informed after the fact?
The Decision Rights Matrix serves as a roadmap for the management team,
especially in the early stages of the transition from traditional organization
to Process Enterprise. Eventually team members will internalize the matrix
BPEs must also work together. Processes overlap and process handoffs are
critical. Often the same worker works with different processes. To avoid ‘‘horizontal
turf wars’’ senior leadership needs to set enterprise goals and develop
compensation and incentive systems that promote teamwork and cooperation
between process owners.
Process excellence
The need for interprocess cooperation highlights the fact that no process is
an island. From the customer’s perspective, it’s all one process. Overall excellence
requires that the entire business be viewed as the customer sees it. One
way to accomplish this is to set up a separate process with a focus of overall
process excellence. For the sake of discussion, let’s call this Process
Excellence (PEX). PEX will have a BPE and it will be considered another
core business process. The mission of PEX is to see to it that all business processes
accomplish the enterprise goals as they relate to customers, shareholders,
and employees. PEX is also concerned with helping BPEs improve
their processes, both internally and across the process boundaries. In other
words, PEX is a meta-process, a process of processes. BPEs, unit managers,
and Process Excellence leaders work together through Process Excellence
Leadership Teams (PELTs) to assure that the needs of the major stakeholder
groups are met (Figure 3.11).
Once the decision is made to become a Six Sigma Process Enterprise, the
question of how to integrate the Six Sigma infrastructure will arise. Here are
my recommendations:
1. Designate Process Excellence (PEX) as one of the enterprise’s core
processes and select a BPE.
2. Master Black Belts will report to PEX. The Master Black Belts will have
an enterprise-wide perspective. They will be concerned with the internal
processes in PEX, as well as the overall value creation and delivery
produced by the cooperative e?orts of the core processes.
3. Black Belts will report to the BPEs, but the budget for the Black Belts
comes from Process Excellence. This gives PEX in?uence which helps
maintain the enterprise perspective, but leaves day-to-day management
and control with the Black Belt’s customers, the BPEs.
4. BPEs have PEX goals, tied to incentives. PEX incentives are in the PEX
5. Unit managers have process-based incentives. Process incentives are in
the BPE’s budgets.
The Six Sigma Process Enterprise 131
Figure 3.11. Process Enterprise roles and responsibilities.
6. The PEX leader and BPEs should collaboratively create a Decision
Rights Matrix identifying:
. The roles and responsibilities of PEX, BPEs, and unit managers. For
example, hiring, budgets, project selection.
. Who makes the decision in the areas just described?
. Who must be consulted in decision-making?
. What is the communication plan?
7. PEX helps develop a BPE Coordination Plan addressing such interprocess
issues as:
. Where do the core processes overlap?
. How will cross-process Six Sigma projects be chartered and coordinated?
. Who will assure that overlapping activities and hando?s are coordinated?
(PEX plays a facilitation role here.)
. When is standardization across processes best and when isn’t it? The
process intersections should be invisible to customers (e.g., customers
shouldn’t have to provide the same information more than
once; single form information for ordering of products, support
plans, registration, etc.). However, diversity may be necessary to
serve unique customer needs.
You may have noticed that having Black Belts reporting to BPEs instead of to
PEX seems to contradict the advice given in the first chapter where I strongly
recommended having the Black Belts report centrally. However, there is a critical
difference. The traditional organizational structure was assumed in
Chapter 1, so if the Black Belts didn’t report to the Six Sigma organization
(referred to here as PEX) they would have been reporting to the unit managers.
I am not recommending that they report to unit managers, but to BPEs. BPEs
are process owners, which gives them a much different perspective than the
unit manager. This perspective, unlike that of unit managers, meshes very well
with the Six Sigma focus on process improvement.
A common problem with Six Sigma is that there is a cognitive disconnect
between the Six Sigma projects and top leadership’s strategic goals. In the previous
chapter we discussed the development of Strategy Deployment Plans.
Strategy Deployment Plans are simple maps showing the linkage between stakeholder
satisfaction, strategies, and metrics. However, these maps are inadequate
guides to operational personnel trying to relate their activities to the vision of
their leadership. Unfortunately, more complexity is required to communicate
the strategic message throughout the organization. We will use QFD for this
purpose. An example, based on the Strategy Deployment Plan shown in
Chapter 2 (Figure 2.4, page 72), will illustrate the process.
The strategy deployment matrix
The first QFD matrix will be based directly on the Strategy Deployment
Plan. If you take a more detailed look at the Strategy Deployment Plan you’ll
notice that the situation is oversimplified. For example, the strategy for operational
excellence is related to operations and logistics, but the Strategy
Deployment Plan doesn’t show this (except indirectly through the link between
internal process excellence and customer perceived value). A Six Sigma project
addressing inventory levels would have an impact on both strategies, but it
wouldn’t be possible to measure the impact from the Strategy Deployment
Plan alone. QFD will help us make this evaluation explicit. A completed Phase
I Strategy Deployment Matrix is shown in Figure 3.12.
The process for developing the Strategy Deployment Matrix is:
1. Create a matrix of the strategies and metrics.
2. Determine the strength of the relationship between each strategy and
3. Calculate a weight indicating the relative importance of the metric.
To begin we create a matrix where the rows are the strategies (what we want
to accomplish) and the columns are the dashboard metrics (how we will operationalize
the strategies and monitor progress). Note that this is the typical
what-how QFD matrix layout, just with a different spin. In each cell (intersection
of a row and a column) we will place a symbol assigning a weight to the relationship
between the row and the column. The weights and symbols used are
shown in Figure 3.13.
The weights are somewhat arbitrary and you can choose others if you desire.
These particular values increase more-or-less exponentially, which places a
high emphasis on strong relationships, the idea being that we are looking for
clear priorities. Weights of 1-2-3 would treat the different relationship strengths
as increasing linearly. Choose the weighting scheme you feel best fits your
After the relationships have been determined for each cell, we are ready to
calculate scores for each row. Remember, the rows represent strategies. For
example, the first row represents our productivity strategy. The Strategy
Deployment Plan indicated that the productivity strategy was operationalized
by the metrics cost-per-unit and asset utilization, and a strong relationship (()
is shown between these metrics and the productivity strategy. However, QFD
Using QFD to Link Six Sigma Projects to Strategies 133
analysis also shows a strong relationship between this strategy and inventory
turns, which affects asset utilization. Critical to quality (CTQ) and profit per
customer are somewhat related to this strategy. To get an overall score for the
productivity strategy just sum the weights across the first row; the answer is
29. These row (strategy) weights provide information on how well the dashboards
measure the strategies. A zero would indicate that the strategy isn’t measured
at all. However, a relatively low score doesn’t necessarily indicate a
problem. For example, the regulatory compliance strategy has a score of 9, but
that comes from a strong relationship between the regulatory compliance
audit and the strategy. Since the audit covers all major compliance issues, it’s
entirely possible that this single metric is sufficient.
The columns represent the metrics on the top-level dashboard, although only
the differentiator metrics will be monitored on an ongoing basis. The metric’s
Figure 3.12. Strategy Deployment Matrix.
Chart produced using QFD Designer software. Qualsoft, LLC.
target is shown at the bottom of each column in the ‘‘how’’ portion of the
matrix. Setting these targets is discussed in Chapter 2. QFD will provide a reality
check on the targets. As you will see, QFD will link the targets to specific
Six Sigma activities designed to achieve them. At the project phase it is far easier
to estimate the impact the projects will have on the metric. If the sum of the project
impacts isn’t enough to reach the target, either more effort is needed or the
target must be revised. Don’t forget, there’s more to achieving goals than Six
Sigma. Don’t hesitate to use QFD to link the organization’s other activities to
the goals.
As discussed in the previous chapter, leadership’s vision for the hypothetical
company is that they be the supplier of choice for customers who want state-ofthe-
art products customized to their demanding requirements. To achieve this
vision they will focus their strategy on four key differentiators: new product
introductions, revenue from new sources, intimate customer relationship, and
R&D deployment time. With our chosen weighting scheme differentiator columns
have a strategic importance score of 5, indicated with a * symbol in the
row labeled Strategic Importance Score. These are the metrics that leadership
will focus on throughout the year, and the goals for them are set very high.
Other metrics must meet less demanding standards and will be brought to the
attention of leadership only on an exception basis. The row labeled Relative
Metric Weight is the product of the criteria score times the strategic importance
score as a percentage for each column. The four differentiator metrics have the
highest relative scores, while product selection (i.e., having a wide variety of standard
products for the customer to choose from) is the lowest.
Using QFD to Link Six Sigma Projects to Strategies 135
Description Weight Symbol
Strong relationship 9 (
Moderate relationship 3 *
Some relationship 1 ~
Di?erentiator metric 5 *
Key requirement metric 1 [
It is vital when using QFD to focus on only the most important columns!
Columns identified with a"in the row labeled Strategic Importance Score
are not part of the organization’s differentiation strategy. This isn’t to say that
they are unimportant. What it does mean is that targets for these metrics will
probably be set at or near their historical levels as indicated by process behavior
charts. The goals will be to maintain these metrics, rather than to drive them to
new heights. An organization has only limited resources to devote to change,
and these resources must be focused if they are to make a difference that will
be noticed by customers and shareholders. This organization’s complete dashboard
has twenty metrics, which can hardly be considered a ‘‘focus.’’ By limiting
attention to the four differentiators, the organization can pursue the strategy
that their leadership believes will make them stand out in the marketplace for
customer and shareholder dollars.*
Deploying differentiators to operations
QFD most often fails because the matrices grow until the analysis becomes
burdensome. As the matrix grows like Topsy and becomes unwieldy, the team
performing QFD begins to sense the lack of focus being documented by the
QFD matrix. Soon, interest begins to wane and eventually the effort grinds to
a halt. This too, is avoided by eliminating"key requirements from the strategy
deployment matrix. We will create a second-level matrix linked only to the differentiators.
This matrix relates the differentiator dashboard metrics to departmental
support strategies and it is shown in Figure 3.14.
To keep things simple, we only show the strategy linkage for three departments:
engineering, manufacturing, and marketing; each department can prepare
its own QFD matrix. Notice that the four differentiator metric columns
now appear as rows in the matrix shown in Figure 3.14. These are the QFD
‘‘whats.’’ The software automatically brings over the criteria performance target,
criteria scores, and relative criteria scores for each row. This information is
used to evaluate the strategy support plans for the three departments.
The support plans for the three departments are shown as columns, theQFD
‘‘hows,’’ or how these three departments plan to implement the strategies. The
relationship between the whats and hows is determined as described above.
For each column the sum of the relationship times the row criteria score is calculated
and shown in the score row near the bottom of the chart. This informa-
*The key requirements probably won’t require explicit support plans. However, if they do QFD can be used to evaluate the
plans. Key requirements QFD should be handled separately.
tion will be used to select and prioritize Six Sigma projects in the next phase of
the QFD.
Figure 3.14 also has a roof, which shows correlations between the whats. This
is useful in identifying related Six Sigma projects, either within the same depart-
Using QFD to Link Six Sigma Projects to Strategies 137
Figure 3.14. Phase II matrix: di?erentiators.
Chart produced using QFD Designer software. Qualsoft, LLC.
ment or in different departments. For example, there is a strong relationship
between the two engineering activities: faster prototype development and
improve concept-to-design cycle time. Perhaps faster prototype development
should be a subproject under the broad heading of improve concept-to-design
cycle time. This also suggests that ‘‘improve concept-to-design cycle time’’ may
have too large a scope. The marketing strategy of ‘‘improve ability to respond
to changing customer needs’’ is correlated with three projects in engineering
and manufacturing. When a strategy support plan involves many cross-functional
projects it may indicate the existence of a core process. This suggests a
need for high-level sponsorship, or the designation of a process owner to coordinate
Deploying operations plans to projects
Figure 3.15 is aQFDmatrix that links the department plans to Six Sigma projects.
(In reality this may require additional flow down steps, but the number
of steps should be kept as small as possible.) The rows are the department
plans. The software also carried over the numeric relative score from the bottom
row of the previous matrix, which is a measure of the relative impact of the
department plan on the overall differentiator strategy. The far right column,
labeled ‘‘Goal Score’’ is the sum of the relationships for the row. For this example
only the top five department plans are deployed to Six Sigma projects. By
summing the numeric relative scores we can determine that these five plans
account for 86% of the impact. In reality you will also only capture the biggest
hitters, although there’s nothing magic about the number five.
There are three Black Belts shown, and eight projects. Each project is shown
as a column in the matrix. The relationship between the project and each departmental
plan is shown in the matrix. The bottom row shows the ‘‘Project
Impact Score,’’ which is the sum of the relationships for the project’s column
times the row’s numeric relative score.
Since the numeric relative scores are linked to department plans, which are
linked to differentiator metrics, which are linked to strategies, the project
impact score measures the project’s impact on the strategy. The validity of
these ‘‘carry-over scores’’ has been questioned (Burke et al., 2002). Through the
Strategy Deployment Plan we can trace the need for the project all the way
back to stakeholders (Figure 3.16). This logical thread provides those engaged
in Six Sigma projects with an anchor to reality and the meaning behind their
The Goal Score column can also be used to determine the support Six Sigma
provides for each department plan. Note that the marketing plan to ‘‘Identify
target markets for new products’’ isn’t receiving any support at all from Six
Sigma projects (assuming that these eight projects are all of the Six Sigma projects).
This may be okay, or it may not be. It all depends on how important the
plan is to the strategic objectives, and what other activities are being pursued
to implement the plan. The Executive Six Sigma Council may wish to examine
project QFD matrices to determine if action is necessary to reallocate Six
Sigma resources.
The Project Impact Score row is useful in much the same way. This row can
be rank-ordered to see which projects have the greatest impact on the strategy.
It is also useful in identifying irrelevant projects. The project Mike Lis pursuing
to improve ‘‘Pin manufacturing capability’’ has no impact on any of the depart-
Using QFD to Link Six Sigma Projects to Strategies 139
Figure 3.15. Phase III matrix: Six Sigma projects.
Chart produced using QFD Designer software. Qualsoft, LLC.
Figure 3.16. Linkage between Six Sigma projects and stakeholders.
mental plans. Unless it impacts some other strategy support plan that isn’t
shown in the QFD matrix, it should be abandoned as a Six Sigma project. The
project may still be something the manufacturing department wants to pursue,
perhaps to meet a goal for a key requirement. However, as a general rule Six
Sigma projects requiring a Black Belt should focus on plans that have a direct
linkage to differentiator strategies.
Oncecustomers havemadetheir demands known, it is important that these be
translated into internal requirements and specifications. The term ‘‘translation’’
is used to describe this process because the activity literally involves interpreting
the words in one language (the customer’s) into those of another (the employee).
For example, regarding the door of her automobile the customer might say ‘‘I
want the door to close completely when I push it, but I don’t want it swinging
closed from just the wind.’’ The engineer working with this requirement must
convert it into engineering terminology such as pounds of force required to
move the door from an open to a closed position, the angle of the door when it’s
opened, and so on. Care must be taken to maintain the customer’s intent
throughout the development of internal requirements. The purpose of specifications
is to transmit the voice of the customer throughout the organization.
In addition to the issue of maintaining the voice of the customer, there is the
related issue of the importance assigned to each demand by the customer.
Design of products and services always involves tradeoffs: gasoline economy
suffers as vehicle weight increases, but safety improves as weight increases.
The importance of each criterion must be determined by the customer. When
different customers assign different importance to criteria, design decisions
are further complicated.
It becomes difficult to choose from competing designs in the face of ambiguity
and customer-to-customer variation. Add to this the differences between
internal personnel and objectivesLdepartment vs. department, designer vs.
designer, cost vs. quality, etc.L and the problem of choosing a design alternative
quickly becomes complex. A rigorous process for deciding which
alternative to settle on is helpful in dealing with the complexity.
Structured decision-making*
The first step in deciding upon a course of action is to identify the goal. For
example, let’s say you’re the owner of the Product Development process for a
* Use of this approach in designing for six sigma is discussed in Chapter 19.
company that sells software to help individuals manage their personal finances.
The product, let’s call it DollarWise, is dominant in its market and your company
is well respected by its customers and competitors, in large part because
of this product’s reputation. The business is profitable and the leadership
naturally wants to maintain this pleasant set of circumstances and to build
on it for the future. The organization has committed itself to a strategy of
keeping DollarWise the leader in its market segment so it can capitalize on
its reputation by launching additional new products directed towards other
financially oriented customer groups, such as small businesses. They have
determined that Product Development is a core process for deploying this
As the process owner, or Business Process Executive, you have control of the
budget for product development, including the resources to upgrade the existing
product. Although it is still considered the best personal financial software
available, DollarWise is getting a little long in the tooth and the competition
has steadily closed the technical gap. You believe that a major product upgrade
is necessary and want to focus your resources on those things that matter most
to customers. Thus, your goal is:
GOAL: Determine where to focus product upgrade resources
Through a variety of ‘‘listening posts’’ (focus groups, user laboratories, internet
forums, trade show interviews, conference hospitality suites, surveys, letters,
technical support feedback, etc.), you have determined that customers ask
questions and make comments like the following:
. Can I link a DollarWise total to a report in my word processor?
. I have a high speed connection and I’d like to be able to download big databases
of stock information to analyze with DollarWise.
. I like shortcut keys so I don’t have to always click around in menus.
. I only have a 56K connection and DollarWise is slow on it.
. I use the Internet to pay bills through my bank. I’d like to do this using
DollarWise instead of going to my bank’s web site.
. I want an interactive tutorial to help me get started.
. I want printed documentation.
. I want the installation to be simple.
. I want the user interface to be intuitive.
. I want to be able to download and reconcile my bank statements.
. I want to be able to upgrade over the Internet.
. I want to manage my stock portfolio and track my ROI.
. I’d like to have the reports I run every month saved and easy to update.
. It’s a pain to set up the di?erent drill downs every time I want to analyze
my spending.
Linking Customer Demands to Budgets 141
. It’s clunky to transfer information between DollarWise and Excel.
. When I have a minor problem, I’d like to have easy-to-use self-help available
on the Internet or in the help ?le.
. When it’s a problem I can’t solve myself, I want reasonably priced, easy to
reach technical support.
. You should be making patches and bug-?xes available free on the Internet.
The first step in using this laundry list of comments is to see if there’s an
underlying structure embedded in them. If these many comments address only
a few issues, it will simplify the understanding of what the customer actually
wants from the product. While there are statistical tools to help accomplish
this task (e.g., structural equation modeling, principal components analysis, factor
analysis), they are quite advanced and require that substantial data be collected
using well-designed survey instruments. An alternative is to create an
‘‘affinity diagram,’’ which is a simple procedure described elsewhere in this
text (see page 314). After creating the affinity diagram, the following structure
was identified:
1. Easy to learn.
1.1. I want the installation to be simple.
1.2. I want an interactive tutorial to help me get started.
1.3. I want printed documentation.
1.4. I want the user interface to be intuitive.
2. Easy to use quickly after I’ve learned it well.
2.1. I like shortcut keys so I don’t have to always click around in menus.
2.2. I’d like to have the reports I run every month saved and easy to
2.3. It’s a pain to set up the di?erent drill downs every time I want to
analyze my spending.
3. Internet connectivity.
3.1. I use the Internet to pay bills through my bank. I’d like to do this
using DollarWise instead of going to my bank’s web site.
3.2. I only have a 56K connection and DollarWise is slow on it.
3.3. I have a high speed connection and I’d like to be able to download
big databases of stock information to analyze with DollarWise.
3.4. I want to be able to download and reconcile my bank statements.
3.5. I want to manage my stock portfolio and track my ROI.
4. Works well with other software I own.
4.1. It’s clunky to transfer information between DollarWise and
4.2. Can I link a DollarWise total to a report in my word processor?
5. Easy to maintain
5.1. I want to be able to upgrade over the Internet.
5.2. You should be making patches and bug-?xes available free on the
5.3. When I have a minor problem, I’d like to have easy-to-use self-help
available on the Internet or in the help ?le.
5.4. When it’s a problem I can’t solve myself, I want reasonably priced,
easy to reach technical support.
The reduced model shows that five key factors are operationalized by the
many different customer comments (Figure 3.17).
Next, we must determine importance placed on each item by customers.
There are a number of ways to do this.
. Have customers assign importance weights using a numerical scale (e.g.,
‘‘How important is ‘Easy self-help’ on a scale between 1 and 10?’’).
. Have customers assign importance using a subjective scale (e.g., unimportant,
important, very important, etc.).
. Have customers ‘‘spend’’ $100 by allocating it among the various items. In
these cases it is generally easier for the customer to ?rst allocate the $100
to the major categories, then allocate another $100 to items within each
. Have customers evaluate a set of hypothetical product o?erings and indicate
their preference for each product by ranking the o?erings, assigning
a ‘‘likely to buy’’ rating, etc. The product o?erings include a carefully
selected mix of items chosen from the list of customer demands. The list
is selected in such a way that the relative value the customer places on
each item in the o?ering can be determined from the preference values.
This is known as conjoint analysis, an advanced technique that is covered
in most texts on marketing statistics.
. Have customers evaluate the items in pairs, assigning a preference rating
to one of the items in each pair, or deciding that both items in a pair
are equally important. This is less tedious if the major categories are
evaluated ?rst, then the items within each category. The evaluation can
use either numeric values or descriptive labels that are converted to
numeric values. The pairwise comparisons can be analyzed to derive
item weights using a method known as the Analytic Hierarchical
Process (AHP) to determine the relative importance assigned to all of
the items.
All of the above methods have their advantages and disadvantages. The simple
methods are easy to use but less powerful (i.e., the assigned weights are less
likely to reflect actual weights). The more advanced conjoint and AHP methods
require special skills to analyze and interpret properly. We will illustrate the
use of AHP for our hypothetical software product. AHP is a powerful technique
that has been proven in a wide variety of applications. In addition to its
Linking Customer Demands to Budgets 143









# '&




# -



# ( 

# .


Figure 3.17. Customer demand model.
use in determining customer importance values, it is useful for decisionmaking
in general.
Category importance weights
We begin our analysis by making pairwise comparisons at the top level. The
affinity diagram analysis identified five categories: easy to learn, easy to use
quickly after I’ve learned it, internet connectivity, works well with other software
I own, and easy to maintain. Arrange these items in a matrix as shown in
Figure 3.18.
For our analysis we will assign verbal labels to our pairwise comparisons;
the verbal responses will be converted into numerical values for analysis.
Customers usually find it easier to assign verbal labels than numeric labels.
All comparisons are made relative to the customer’s goal of determining
which product he will buy, which we assume is synonymous with our goal
of determining where to focus product upgrade efforts. The highlighted cell
in the matrix compares the ‘‘easy to learn’’ attribute and the ‘‘easy to use
quickly after I’ve learned it’’ attribute. The customer must determine which
is more important to him, or if the two attributes are of equal importance.
In Figure 3.18 this customer indicates that ‘‘easy to learn’’ is moderately to
Linking Customer Demands to Budgets 145
Figure 3.18. Matrix of categories for pairwise comparisons.
Created using Expert Choice 2000 Software,*
*Although the analysis is easier with special software, you can obtain a good approximation using a spreadsheet. See the
Appendix for details.
strongly preferred over ‘‘easy to use quickly after I’ve learned it’’ and the software
has placed a +4 in the cell comparing these two attributes. (The scale
goes from ^9 to +9, with ‘‘equal’’ being identified as a +1.) The remaining
attributes are compared one by one, resulting in the matrix shown in Figure
3.19. The shaded bars over the attribute labels provide a visual display of the
relative importance of each major item to the customer. Numerically, the
importance weights are:*
. Easy to learn: 0.264 (26.4%)
. Easy to use quickly after I’ve learned it: 0.054 (5.4%)
. Internet connectivity: 0.358 (35.8%)
. Works well with other software I own: 0.105 (10.5%)
. Easy to maintain: 0.218 (21.8%)
These relative importance weights can be used in QFD as well as in the AHP
process that we are illustrating here. In our allocation of effort, we will want to
emphasize those attributes with high importance weights over those with
lower weights.
Subcategory importance weights
The process used for obtaining category importance weights is repeated for
the items within each category. For example, the items interactive tutorial,
good printed documentation, and intuitive interface are compared pairwise
within the category ‘‘easy to learn.’’ This provides weights that indicate the
importance of each item on the category. For example, within the ‘‘easy to
learn’’ category, the customer weights might be:
. Interactive tutorial: 11.7%
. Good documentation: 20.0%
. Intuitive interface: 68.3%

Figure 3.19. Completed top-level comparison matrix.
*See the Appendix for an example of how to derive approximate importance weights usingMicrosoft Excel.
If there were additional levels below these subcategories, the process would
be repeated for them. For example, the intuitive interface subcategory might
be subdivided into ‘‘number of menus,’’ ‘‘number of submenus,’’ ‘‘menu items
easily understood,’’ etc. The greater the level of detail, the easier the translation
of the customer’s demands into internal specifications. The tradeoff is that the
process quickly becomes tedious and may end up with the customer being
asked for input he isn’t qualified to provide. In the case of this example, we’d
probably stop at the second level.
Global importance weights
The subcategory weights just obtained tell us how much importance the
item has with respect to the category, not with respect to the ultimate goal.
Thus, they are often called local importance weights. However, the subcategory
weights don’t tell us the impact of the item on the overall goal, which
is called its global impact. This is determined by multiplying the subcategory
item weight by the weight of the category in which the item resides.
The global weights for our example are shown in Table 3.4 in descending
The global importance weights are most useful for the purpose of allocating
resources to the overall goal: Determine where to focus product upgrade
efforts. For our example, Internet connectivity obviously has a huge customer
impact. ‘‘Easy to use quickly after I’ve learned it’’ has relatively low impact.
‘‘Easy to learn’’ is dominated by one item: the user interface. These weights
will be used to assess different proposed upgrade plans. Each plan will be
evaluated on each subcategory item and assigned a value depending on how
well it addresses the item. The values will be multiplied by the global weights
to arrive at an overall score for the plan. The scores can be rank-ordered to
provide a list that the process owner can use when making resource allocation
decisions. Or, more proactively, the information can be used to develop
a plan that emphasizes the most important customer demands. Table 3.5
shows part of a table that assesses project plans using the global weights.
The numerical rating used in the table is 0=No Impact, 1=Some Impact,
3=Moderate Impact, 5=High Impact. Since the global weights sum to 1
(100%), the highest possible score is 5. Of the five plans evaluated, Plan C
has the highest score. It can be seen that Plan C has a high impact on the
six most important customer demands. It has at least a moderate impact on
10 of the top 11 items, with the exception of ‘‘Reasonably priced advanced
technical support.’’ These items account for almost 90% of the customer
Linking Customer Demands to Budgets 147
The plan’s customer impact score is, of course, only one input into the decision-
making process. The rigor involved usually makes the score a very valuable
piece of information. It is also possible to use the same approach to incorporate
other information, such as cost, timetable, feasibility, etc. into the final decision.
The process owner would make pairwise comparisons of the different inputs
Table 3.4. Local and global importance weights.
Category Subcategory
Easy to learn Intuitive interface 68.3% 18.0%
Internet connectivity Online billpay 43.4% 15.5%
Internet connectivity Download statements 23.9% 8.6%
Internet connectivity Download investment
23.9% 8.6%
Works well with other
Hotlinks to spreadsheet 75.0% 7.9%
Easy to maintain Free internet patches 35.7% 7.8%
Easy to maintain Great, free self-help technical
assistance on the internet
30.8% 6.7%
Easy to learn Good documentation 20.0% 5.3%
Easy to maintain Reasonably priced advanced
technical support
20.0% 4.4%
Internet connectivity Works well at 56K 8.9% 3.2%
Easy to learn Interactive tutorial 11.7% 3.1%
Easy to maintain Automatic internet upgrades 13.5% 2.9%
Works well with other
Edit reports in word
25.0% 2.6%
Easy to use quickly after I’ve
learned it
Savable frequently used
43.4% 2.3%
Easy to use quickly after I’ve
learned it
Shortcut keys 23.9% 1.3%
Easy to use quickly after I’ve
learned it
Short menus showing only
frequently used
23.9% 1.3%
Easy to use quickly after I’ve
learned it
Macro capability 8.9% 0.5%
(customer impact score, cost, feasibility, etc.) to assign weights to them, and use
the weights to determine an overall plan score. Note that this process is a
mixture of AHP and QFD.
Linking Customer Demands to Budgets 149
Table 3.5. Example of using global weights in assessing alternatives.
GLOBAL WEIGHT 18.0% 15.5% 8.6% 8.6% 7.9% 7.8% 6.7% 5.3% 4.4% 3.2% 3.1%
PLANA 3.57 3 5 1 1 3 3 4 5 5 5 5
PLANB 2.99 1 1 1 3 3 5 5 5 5 5 5
PLANC 4.15 5 5 5 5 5 5 3 3 1 3 3
PLAND 3.36 3 3 3 3 3 3 3 5 5 5 5
PLANE 2.30 5 0 0 0 5 5 1 1 0 1 1
^ ^ ^
Training for Six Sigma
Education (teaching people how to thinkdifferently ) and training (teaching
people how to do things differently) are vital elements in Six Sigma success.
Although education and training are different, for simplicity we will refer to
both as simply ‘‘training.’’
The Six Sigma organization is unlike the traditional organization and the
knowledge, skills and abilities (KSAs) required for success in the new organization
are different than those possessed by most employees. The new KSAs
need to be identified and plans need to be developed to assure that employees
acquire them. The investment required is likely to be significant; careful planning
is required to assure a good ROI.
The first step in the development of the strategic training plan is a training
needs assessment. The training needs assessment provides the background
necessary for designing the training program and preparing the training plan.
The assessment proceeds by performing a task-by-task audit to determine
what the organization is doing, and comparing it to what the organization
should be doing. The assessment process focuses on three major areas:
Process auditLAll work is a process. Processes are designed to add values to
inputs and deliver values to customers as outputs. Are they operating as
designed? Are they operated consistently? Are they measured at key control
points? If so, do the measurements show statistical control? The answers to
Copyright 2003 by The McGraw-Hill Companies, Inc. Click Here for Terms of Use.
these questions, along with detailed observations of how the process is operated,
are input to the development of the training plan.
Assessment of knowledge, skills and abilitiesLIn all probability, there will
be deficiencies (opportunities for improvement) observed during the process
audits. Some of these deficiencies will involve employee KSAs. The first principle
of self-control is that employees must know what they are doing.
Management’s job doesn’t end by simply giving an employee responsibility for
a particular process or task, they must also provide the employee with the
opportunity to acquire the KSAs necessary to successfully perform their new
duties. This means that if the employee is asked to assume new duties as a member
of a Six Sigma improvement team, they are given training in team skills, if
they are to keep a control chart, they receive training in the maintenance and
interpretation of the charts, etc. Since employees are expected to contribute to
the implementation of the organization’s strategic plan, they should be told
what the plan is, and how their job contributes to the plan.
Assessment of employee attitudesLAttitudes are emotions that reflect a
response to something taking place within an organization. A person’s attitude
is, in essence, a judgment about the wisdom of a particular course of events. If
an employee’s attitude is not positive, they will not use their KSAs to help the
organization as effectively as they could. Negative employee attitudes about
the direction being taken by the organization indicate that the employee either
questions the wisdom of the proposed changes, or doubts the sincerity of the
leadership. Regardless, it represents a problem that must be addressed by the
training plan.
Theassessments above can be conducted using audits or the survey techniques
described in Chapter 3. Assessments can be conducted by either internal or external
personnel. In general, employees are more likely to be open and honest
when confidentiality is assured, which is more likely when assessments are conducted
by outside parties. However, internal assessments can reveal valuable
information if proper safeguards are observed to assure the employee’s privacy.
It is important that follow-up assessments be made to determine if the training
conducted closed the gap between the ‘‘is’’ and the ‘‘should be.’’ The follow
up will also provide a basis for additional training. Reassessment should be conducted
first to assure that the desired KSAs were acquired by the target group
of employees, then the process should be reassessed to determine if the new
KSAs improved the process as predicted. It’s common to discover that we
made a mistake in assuming that the root cause of the process ‘‘is/should-be’’
gap is a KSA deficiency. If the reassessments indicate that the KSA gap was
closed but the process gap persists, another approach must be taken to close
the process gap.
Training Needs Analysis 151
The strategic training plan is a project plan describing in detail how the gaps
identified in the training needs analysis will be addressed. As with any project
plan, the strategic training plan includes a charter containing a detailed description
of the training deliverables, major activities to be undertaken, responsibility
for each major activity, resources required, deadlines and timetables, etc. In
most organizations the plan describes several major subprojects, such as the
leadership training project, the Green Belt, Black Belt and Master Black Belt
training project, and so on.
In the traditional organization, the training department is another ‘‘silo,’’
with its own budget and special interests to protect. In this system the training
plans are often not tied to strategic plans. Instead, these plans tend to serve the
needs of the trainers and others in the training department, rather than serving
the needs of the organization as awhole. The effect on the organization’s performance
is the same as when any organizational unit pursues its own best interest
rather than the organization’s interests: negative return on investment,
damaged morale, etc.
In Six Sigma organizations training plans are tied directly to the current and
future needs of external customers. These needs are, in turn, the driver behind
the organization’s strategic plan. The strategic training plan provides the means
of developing the knowledge, skills,andabilities that willbeneededbyemployees
in the organization in the future. The people who develop the strategic plan also
commission the development of the strategic training plan. In many organizations,
Six Sigma training is a subproject of the Six Sigma deployment plan. The
training timetable must be tightly linked to the timetable for overall Six Sigma
deployment. Providing training either too early or too late is a mistake. When
training is provided too early, the recipient will forget much of what he has
learned before it is needed.Whenit is provided too late, the quality of the employee’s
work will suffer. When it comes to training, just-in-time delivery is the goal.
The Six Sigma training plan must include a budget. The Six Sigma training
budget lists those resources that are required to provide the training. The training
budget traditionally includes a brief cost/benefit analysis. Cost/benefit analysis
for training, as for all expenditures for intangibles, is challenging.
Estimating cost is usually simple enough. Examples of training costs include:
. trainer salaries
. consulting fees
. classroom space and materials
. lost time from the job
. sta? salaries
. o?ce space of training sta?
Estimating benefits with the same degree of precision is problematic. It is
usually counterproductive to attempt to get high precision in such estimates.
Instead, most organizations settle for rough figures on the value of the trainee
to the company. Some examples of training benefits include:
. improved e?ciency
. improved quality
. increased customer satisfaction
. improved employee morale
. lower employee turnover
. increased supplier loyalty
It isn’t enough to support training in the abstract. Training budgets are tangible
evidence of management support for the goals expressed in the training
plan. In addition, management support is demonstrated by participating in the
development of the strategic training plan. In most cases, senior management
delegates the development of annual training plans and budgets to the training
department staff.
Training needs of various groups
Leaders should receive guidance in the art of ‘‘visioning.’’ Visioning
involves the ability to develop a mental image of the organization at a future
time. Without a vision, there can be no strategy; how can you develop a strategy
without knowing where it is supposed to lead? The future organization
will more closely approximate the ideal organization, where ‘‘ideal’’ is defined
as that organization which completely achieves the organization’s values.
How will such an organization ‘‘look’’? What will its employees do? Who
will be its customers? How will it behave towards its customers, employees,
and suppliers? Developing a lucid image of this organization will help the leader
see how she should proceed with her primary duty of transforming the
present organization. Without such an image in her mind, the executive will
lead the organization through a maze with a thousand dead ends.
Conversely, with her vision to guide her, the transformation process will proceed
on course. This is not to say that the transformation is ever ‘‘easy.’’ But
when there is a leader with a vision, it’s as if the organization is following
an expert scout through hostile territory. The destination is clear, but the
journey is still difficult.
Leaders need to be masters of communication. Fortunately, most leaders
already possess outstanding communication skills; few rise to the top without
them. However, training in effective communication is still wise, even if it is
The Strategic Training Plan 153
only refresher training for some. Also, when large organizations are involved,
communications training should include mass communication media, such as
video, radio broadcasts, print media, etc. Communicating with customers,
investors, and suppliers differs from communicating with employees and colleagues,
and special training is often required. Communication principles are discussed
in the previous chapter.
When an individual has a vision of where he wants to go himself, he can
pursue this vision directly. However, when dealing with an organization, simply
having a clear vision is not enough. The leader must communicate the
vision to the other members of the organization. Communicating a vision is
a much different task than communicating instructions or concrete ideas.
Visions of organizations that embody abstract values are necessarily abstract
in nature. To effectively convey the vision to others, the leader must convert
the abstractions to concretes. One way to do this is by living the vision. The
leader demonstrates her values in every action she takes, every decision she
makes, which meetings she attends or ignores, when she pays rapt attention
and when she doodles absentmindedly on her notepad. Employees who are
trying to understand the leader’s vision will pay close attention to the behavior
of the leader.
Another way to communicate abstract ideas is to tell stories. In organizations
there is a constant flow of events. Customers encounter the organization
through employees and systems, suppliers meet with engineers, literally thousands
of events take place every day. From time to time an event occurs that captures
the essence of the leader’s vision. A clerk provides exceptional customer
service, an engineer takes a risk and makes a mistake, a supplier keeps the line
running through a mighty effort. These are concrete examples of what the leader
wants the future organization to become. She should repeat these stories to
others and publicly recognize the people who made the stories. She should also
create stories of her own, even if it requires staging an event. There is nothing
dishonest about creating a situation with powerful symbolic meaning and
using it to communicate a vision. For example, Nordstrom has a story about a
sales clerk who accepted a customer return of a defective tire. This story has tremendous
symbolic meaning because Nordstrom doesn’t sell tires! The story
illustrates Nordstrom’s policy of allowing employees to use their own best judgment
in all situations, even if they make ‘‘mistakes,’’ and of going the extra
mile to satisfy customers. However, it is doubtful that the event ever occurred.
This is irrelevant. When employees hear this story during their orientation
training, the message is clear. The story serves its purpose of clearly communicating
an otherwise confusing abstraction.
Leaders need training in conflict resolution. In their role as process owner in
a traditional organization, leaders preside over a report-based hierarchy trying
to deliver value through processes that cut across several functional areas. The
inevitable result is competition for limited resources, which creates conflict.
Of course, the ideal solution is to resolve the conflict by designing organizations
where there is no such destructive competition. Until then, the leader can expect
to find a brisk demand for his conflict-resolution services.
Finally, leaders should demonstrate strict adherence to ethical principles.
Leadership involves trust, and trust isn’t granted to one who violates a moral
code that allows people to live and work together. Honesty, integrity, and
other moral virtues should be second nature to the leader.
Black Belts are expected to deliver tangible results on projects selected to
have a measurable impact on the organization’s bottom line. This is a big order
to fill. The means of accomplishing it is an approach (DMAIC or DFSS) and a
set of techniques that collectively are known as the Six Sigma method. Black
Belts receive from three to six weeks of training in the technical tools of Six
Sigma. Three week curricula are usually given to Black Belts working in service
or transaction-based businesses, administrative areas, or finance. Four week
programs are common for manufacturing environments.* Six weeks of training
are provided for Black Belts working in R&D or similar environments. Figure
4.1 shows the curriculum used for courses in General Electric for personnel
with finance backgrounds who will be applying Six Sigma to financial, general
business, and eCommerce processes. Figure 4.2 shows GE’s curriculum for the
more traditional manufacturing areas.
Some training companies offer highly compressed two week training
courses, but I don’t recommend this. My experience with coaching Black Belts,
as well as student feedback I have received from attendees of four and five
week training programs, indicates that these are compressed plenty already!
Even with the six week courses, in some weeks students receive the equivalent
of two semesters of college-level applied statistics in just a few days. Humans
require a certain ‘‘gestation period’’ to grasp challenging new concepts and
stuffing everything into too short a time period is counterproductive.
In general, Black Belts are hands-on oriented people selected primarily
for their ability to get things done. Tools and techniques are provided to
The Strategic Training Plan 155
*A fifth week of training in LeanManufacturing is often provided for Black Belts working in manufacturing.
help them do this. Thus, the training emphasis is on application, not theory.
In addition, many Black Belts will work on projects in an area where they
possess a high degree of subject-matter expertise. Therefore, Black Belt training
is designed around projects related to their specific work areas. This
requires Master Black Belts or trainers with very broad project experience
to answer application-specific questions. When these personnel aren’t available,
examples are selected to match the Black Belt’s work as close as possible.
For example, if no trainer with human resource experience is available
the examples might be from another service environment; manufacturing
examples would be avoided. Another common alternative is to use consultants
to conduct the training. Consultants with broad experience within
Week 1
The DMAIC and DFSS (design for Six Sigma) improvement strategies
Project selection and ‘‘scoping’’ (Define)
QFD (quality function deployment)
Sampling principles (quality and quantity)
Measurement system analysis (also called ‘‘Gage R&R’’)
Process capability
Basic graphs
Hypothesis testing
Week 2
Design of experiments (DOE) (focus on two-level factorials)
Design for Six Sigma tools
Requirements flowdown
Capability flowup (prediction)
FMEA (failure mode and effects analysis)
Developing control plans
Control charts
Week 3
Power (impact of sample size)
Impact of process instability on capability analysis
Confidence intervals (vs. hypothesis tests)
Implications of the Central Limit Theorem
How to detect ‘‘lying with statistics’’
General linear models
Fractional factorial DOEs
Figure 4.1. Sample curriculum for ?nance Black Belts.
From Hoerl, Roger W. (2001), ‘‘Six Sigma Black Belts: What Do TheyNeed to Know?’’,
Journal of Quality Technology, 33(4), October, p. 395. Reprinted by permission of ASQ.
The Strategic Training Plan 157
–Why Six Sigma
–DMAIC and DFSS processes (sequential case studies)
–Project management fundamentals
–Team effectiveness fundamentals
–Project selection
–Scoping projects
–Developing a project plan
–Multigenerational projects
–Process identification (SIPOC)
–Developing measurable CTQs
–Sampling (data quantity and data quality)
–Measurement system analysis (not just gage R&R)
–SPC Part I
–The concept of statistical control (process stability)
–The implications of instability on capability measures
–Capability analysis
–Basic graphical improvement tools (‘‘Magnificent 7’’)
–Management and planning tools (Affinity, ID, etc.)
–Confidence intervals (emphasized)
–Hypothesis testing (de-emphasized)
–ANOVA (de-emphasized)
–Developing conceptual designs in DFSS
–DOE (focus on two-level factorials, screening designs, and RSM)
–Piloting (of DMAIC improvements)
–DFSS design tools
–CTQ flowdown
–Capability flowup
–Developing control plans
–SPC Part II
–Control charts
–Piloting new designs in DFSS
Figure 4.2. Sample curriculum for manufacturing Black Belts. (The week in which the
material appears is noted as a superscript.)
From Hoerl, Roger W. (2001), ‘‘Six Sigma Black Belts: What Do They Need to Know?’’,
Journal of Quality Technology, 33(4), October, p. 399. Reprinted by permission of ASQ.
the enterprise as well as with other organizations can sometimes offer
Black Belts must work on projects while they are being trained. Typically,
the training classes are conducted at monthly intervals and project work is
pursued between classes. One of the critical differences between Six Sigma
and other initiatives is the emphasis on using the new skills to get tangible
results. It is relatively easy to sit in a classroom and absorb the concepts
well enough to pass an exam. It’s another thing entirely to apply the new
approach to a real-world problem. For one thing, there are other people
involved. The Black Belt has to be able to use her change agent skills to
recruit sponsors and team members and to get these people to work together
on a project with a challenging goal and a tight timetable. The Black Belt is
not yet comfortable with her new skills and she’s reluctant to go in front of
others to promote this new approach. But it won’t get any easier. During
Black Belt training she’ll have the moral support of many others in the same
situation. The instructors can provide coaching and project-specific training
and advice. Because she’s new, people will be more forgiving than at any
future time. In short, there’s no better time to get her feet wet than during
the training.
The curriculum for Green Belts is for a one week course (Figure 4.3). The
same material is sometimes covered in two weeks. The primary difference
between the one and two week courses is the way in which in-class exercises
are handled. In some companies Green Belts are provided with their own
copies of software for statistical analysis, project planning, flowcharting, etc.
and expected to be able to use the software independently. In others, the
Green Belt is not expected to be proficient in the use of the software, but
relies on Black Belts for this service. In the former case, Green Belt training
is extended to provide the necessary hands-on classroom experience using
the software.
‘‘Belts’’ seldom do solitary work. In nearly all cases they work with teams,
sponsors, leadership, etc. Seldom does the Black Belt have any authority to
direct any member of the team. Thus, as a full-time change agent, the Black
Belt needs excellent interpersonal skills. Other change agents also need training
in soft skills. In addition to mastering a body of technical knowledge, Belts and
other change agents need to be able to:
The Strategic Training Plan 159
Opening comments
Red bead demo
Introduction to Six Sigma
The DMAIC and DFSS improvement strategies
Lean manufacturing
Project selection, scope, and charter
Teaming exercise
Process mapping, SIPOC
Define gate criteria (how to close the Define phase)
Data collection, scales, distributions, yields
Measurement systems
–SPC Part I
–The concept of statistical control (process stability)
–The implications of instability on capability measures
Measure gate criteria
Scatter plots
Other 7M tools
Run charts
Box plots
Confidence intervals
Design of experiments (DOE)
Analyze gate criteria
–SPC Part II
–Process behavior charts
Change tools
Force field analysis
Project planning and management (improvement planning)
Improve gate criteria
Process control planning matrix
Process FMEA
Process control plan
Control gate criteria
Figure 4.3. Sample curriculum for Green Belts.
. Communicate e?ectively verbally and in writing
. Communicate e?ectively in both public and private forums
. Work e?ectively in small group settings as both a participant and a leader
. Work e?ectively in one-on-one settings
. Understand and carry out instructions from leaders and sponsors
Too many people believe that so-called soft skills are less important than
technical skills. Others think that, while soft skills are important, they are easier
to master. Neither of these beliefs are correct. Soft skills are neither less important
nor easier to master, they are just different. In my experience, a change
agent deficient in soft skills will nearly always be ineffective. They are usually
frustrated and unhappy souls who don’t understand why their technically brilliant
case for change doesn’t cause instantaneous compliance by all parties. The
good news is that if the person is willing to apply as much time and effort to
soft-skill acquisition and mastery as they applied to honing their technical skills,
they will be able to develop proficiency.
Soft skills are employed in a variety of ways, such as:
CoachingLAcoach doesn’t simply tell the players what to do, he or she clearly
explains how to do it. A baseball pitching coach studies the theory of pitching
and the individual pitcher and provides guidance on how to hold the ball, the
windup, the delivery, etc. The coach is constantly trying to ?nd ways to help
the pitcher do a better job. In Six Sigma work there is a coaching chain: leaders
coach champions and sponsors, champions and sponsors coach Master Black
Belts, Master Black Belts coach Black Belts, Black Belts coach Green Belts, and
Green Belts coach team members. Each link in the chain helps the next link
learn more about doing their job right.
MentoringLThementor understands the organization to such a degree that he
has acquired deep wisdom regarding the way it works. This allows him to see
relationships that are not apparent to others. The mentor helps the change
agent avoid organizational obstacles and to negotiate the barriers. Mentoring
isn’t so much about the mentor blazing a trail for the change agent as it is
about providing the change agent with amap for getting things done e?ectively.
NegotiationLChange agents must negotiate with others in the organization,
as well as with suppliers and customers, to acquire the resources necessary to
accomplish his department’s goals. Obtaining these resources without engendering
ill will requires negotiating skill and diplomacy.
Con?ict resolutionLThe change agent must coordinate the activities of many
people, and do so without line authority over any of them. When these people
cannot resolve their own di?erences, the change agent must provide guidance.
Change agents should also receive training in the fundamentals of accounting
and finance. Such information is essential to such activities as cost/benefit
analysis, budgeting, and quality costs. The goal isn’t to make them accountants,
but to familiarize them with basic concepts.
Finally, change agents should possess certain technical skills that are crucial
to their ability to carry out Six Sigma projects. Change agents must understand
enough about measurement issues in the social sciences to be able to measure
the effectiveness of their employee and customer projects. Deming lists an
understanding of theory of variation as one of the cornerstones of his system
of profound knowledge. This requires rudimentary statistical skills. Change
agents without this training will misdiagnose problems, see trends where none
exist, overreact to random variation, and in general make poor decisions.
Facilitating group activities requires that the change agent possess certain
unique skills. It is unlikely that an individual who is not already a facilitator
will already possess the needed skills. Thus, it is likely that facilitator training
will be needed for change agents. A good part of the facilitator’s job involves
communicating with people who are working on teams. This role involves the
following skills:
Communication skillsLQuite simply, the change agent who cannot communicate
well is of little use to the modern organization.
Meeting management skillsLSchedule the meeting well ahead of time. Be
sure that key people are invited and that they plan to attend. Prepare an agenda
and stick to it! Start on time. State the purpose of the meeting clearly at the outset.
Take minutes. Summarize from time to time. Actively solicit input from
those less talkative. Curtail the overly talkative members. Manage con?icts.
Make assignments and responsibilities explicit and speci?c. End on time.
Presentation skillsLKnow why you are speaking to this audience (inform/
educate or convince/persuade); perform the task; solicit the desired audience
Presentation preparationLPrepare a list of every topic you want to cover.
Cull the list to those select few ideas that are most important. Number your
points. Analyze each major point. Summarize.
Use of visual aidsLA visual aid in a speech is a pictorial used by a speaker to
convey an idea. Well-designed visual aids add power to a presentation by
showing the idea more clearly and easily than words alone. Whereas only
10% of presented material is retained from a verbal presentation after 3 days,
65% is retained when the verbal presentation is accompanied by a visual aid.
However, if the visual aids are not properly designed, they can be distracting
and even counterproductive. ASQ reports that poor visuals generate more
The Strategic Training Plan 161
negative comment from conference attendees than any other item. Change
agents must be sensitive to non-verbal communication. There is much more
to communication than mere words. Facilitators should carefully observe posture
and body movements, facial expressions, tone of voice, ?dgeting, etc. If
the facilitator sees these non-verbal signals he should use them to determine
whether or not to intervene. For example, a participant who shakes his head
when hearing a particular message should be asked to verbalize the reasons
why he disagrees with the speaker. A person whose voice tone indicates sarcasm
should be asked to explain the rationale behind his attitude. A wall-
?ower who is squirming during a presentation should be asked to tell the
group her thoughts. Facilitators should be active listeners. Active listening
involves certain key behaviors:
. look at the speaker
. concentrate on what the speaker is saying, not on how you will respond to
. wait until the speaker is ?nished before responding
. focus on the main idea, rather than on insigni?cant details
. keep emotional reactions under control
Because all of the work of facilitators involves groups, facilitators should
have an in-depth understanding of group dynamics and the team process (see
Chapter 5). Also, because the groups and teams involved are usually working
on Six Sigma improvement projects, the facilitator should be well versed in
project management principles and techniques (see Chapters 6 and 15).
Post-training evaluation and reinforcement
Training is said to have ‘‘worked’’ if it accomplishes its objectives. Since the
training objectives are (or should be) derived from the strategic plan, the ultimate
test is whether or not the organization has accomplished its strategic
objectives. However, training is only one of dozens of factors that determine if
an organization accomplishes its strategic objectives, and one that is often far
removed in time from the final result. To assess training effectiveness we need
more direct measures of success, and we need to measure near the time the training
has been completed.
Except in academic settings, imparting knowledge or wisdom is seldom the
ultimate goal of training. Instead, it is assumed that the knowledge or wisdom
will result in improved judgments, lower costs, better quality, higher levels of
customer satisfaction, etc. In other words, the training will produce observable
results. These results were the focus of the training plan development and training
needs analysis described earlier in this chapter. Training evaluation requires
that they be converted to training measurables or objectives.
Regardless of the format of the presentation, the basic unit of training is the
lesson. A lesson is a discrete ‘‘chunk’’ of information to be conveyed to a learner.
The training objectives form the basis of each lesson, and the lessons provide
guidance for development of measurements of success.
Lesson plans provide the basis for measurement at the lowest level. The
objectives in the lesson plan are specific and the lesson is designed to accomplish
these specific objectives. The assumption is that by accomplishing the set of
objectives for each lesson, the objectives of the seminar or other training activity
will be met. A further assumption is that by meeting the objectives of all of the
training activities, the objectives of the training plan will be met. Finally, it is
assumed that by meeting the objectives of the training plan, the objectives of
the strategic plan (or strategic quality plan) will be met, or at least will not be
compromised due to training inadequacies. All of these assumptions should be
subjected to evaluation.
The evaluation process involves four elements (Kirkpatrick, 1996):
1. ReactionLHow well did the conferees like the program? This is
essentially customer satisfaction measurement. Reaction is usually
measured using comment sheets, surveys, focus groups and other
customer communication techniques. See Chapter 3 for additional
information on these topics.
2. LearningLWhat principles, facts, and techniques were learned? What
attitudes were changed? It is entirely possible that conferees react
favorably to training, even if learning does not occur. The learning of
each conferee should be quanti?ed using pre- and post-tests to identify
learning imparted by the training. Results should be analyzed using
proper statistical methods. In exceptional cases, e.g., evaluating a consulting
company for a large training contract, a formal designed experiment
may be justi?ed.
3. BehaviorLWhat changes in behavior on-the-job occurred? If the conferee
leaves the SPC presentation and immediately begins to e?ectively
apply control charts where none were used before, then the training
had the desired e?ect on behavior. However, if the conferee’s tests indicate
that she gained competence in the subject matter from the training,
but no change in behavior took place, the training investment was
wasted. Note that behavior change is dependent on a great number of
factors besides the training, e.g., management must create systems
where the newly learned behaviors are encouraged.
The Strategic Training Plan 163
4. ResultsLWhat were the tangible results of the program in terms of
reduced cost, improved quality, improved quantity, etc.? This is the real
payback on the training investment. The metrics used for measuring
results are typically built into the action plans, project plans, budgets,
etc. Again, as with behavior change, there are many factors other than
training that produce the desired results.
Phillips adds a fifth item to the above list (Phillips, 1996, p. 316):
5. Return on investment (ROI)LDid the monetary value of the results
exceed the cost for the program?
Phillips considers these five items to be different levels of evaluation. Each
evaluation level has a different value, as shown in Figure 4.4.
Due to the difficulty and cost involved, it is impractical and uneconomical to
insist that every program be evaluated at all five levels. Sampling can be used to
obtain evaluations at the higher levels. As an example, one large electric utility
set the sampling targets in Table 4.1.
Where sampling is used, programs should be selected using a randomization
procedure such as random numbers tables. ROI calculations are not difficult
and they are described in Chapter 4. However, to make the results credible,
finance and accounting personnel should be involved in calculating financial
ratios of this type.
Figure 4.4. Characteristics of evaluation levels.
From Phillips, J.J. (1996), ‘‘Measuring the Results of Training,’’ in Craig, R.L., editor in chief,
The ASTD Training & Development Handbook: A Guide to Human Resources Development.
New York: McGraw-Hill, p. 317.
When the subject of reinforcement is raised, monetary remuneration usually
comes to mind first. Skill-based pay is gaining favor in some quarters for a variety
of reasons:
. encourage employees to acquire additional skills
. reward people for training and education
. as a reaction to the negative aspects of performance-based pay
While skill-based pay may have merit, cash awards and other such ‘‘rewards’’
are of dubious value and should probably not be used.
Rather than assuming that employees will only engage in training if they
receive an immediate tangible reward, research and experience indicate that
most employees find value in training that helps them better achieve their personal,
job, and career goals. Thus, reinforcement is accomplished by providing
the trainee with the opportunity to use the skills they learned. Proficiency is
gained with practice soon after the learning has taken place. Management should
provide an environment where the new skills can be honed without pressure
and distraction. The ‘‘just-in-time’’ (JIT) principle applies here. Don’t provide
training for skills that won’t be used in the near future.
People who have just learned something new, be it a job skill or a new philosophy
such as quality focus, often have questions arise as they attempt to integrate
their new knowledge into their daily thoughts and routine. User groups
are very helpful. A user group consists of a number of people who have received
similar training. User groups meet from time to time to discuss their under-
The Strategic Training Plan 165
Table 4.1. Targets for percentages of programs to be evaluated.
From Phillips, J.J. (1996), ‘‘Measuring the Results of Training,’’ in Craig,
R.L., editor in chief, The ASTDTraining & Development Handbook : A
Guide to Human Resources Development. New York: McGraw-Hill, p. 317.
Participant’s satisfaction 100
Learning 70
On-the-job-applications (behavior) 50
Results 10
standing of the material with others. The focus of the group varies from ‘‘How
are you using this?’’ to ‘‘I don’t understand this’’ to ‘‘Here’s what I am doing!’’
At times, well-versed speakers will be invited to clarify particular aspects of the
subject. Presentations of successful applications may be made. Management
can encourage user groups by providing facilities, helping with administrative
details, and, especially, by attending their meetings on occasion.
Electronic forums are gaining in popularity. Trainers will often make themselves
available to answer questions via email. Forum subscribers will send
their questions or comments to a ‘‘list server.’’ The list server then automatically
broadcasts the question or comment to other subscribers on the list. Every subscriber
receives every user’s message, along with the responses to the message.
This often produces ‘‘threads.’’ A thread is an exchange of information on a particular
topic. E.g., subscriber A has a question about using control charts on
financial data, subscriber B responds, then C responds to B and so on. These
threads look remarkably like a free-wheeling face-to-face discussion. The result
is that a great deal of learning takes place in a format that everyone finds to be
more natural (and more interesting) than a traditional classroom environment.
Learning that isn’t used right away tends to fade. Even when just-in-time
training (JITT) is used, it is unlikely that every skill will be put to immediate
and routine use. It is wise to plan for periodic refresher courses to hone the skills
acquired during prior training. A refresher course is usually shorter, faster, and
more intense than new learning. Instructors need not have the same subject matter
mastery as those teaching new material. In fact, it may not even be necessary
to have an instructor available. Media such as video and audio tape programs,
CD ROM, slide presentations, etc., may be sufficient. These self-study media
offer a number of cost and scheduling advantages.
If in-house instructors are used, they may be available to answer occasional
questions from previous trainees. Of course, when the instructor is not a fulltime
trainer, this must be strictly limited. There are a number of ways to reduce
the demands on the trainer’s time while still answering most of the questions
from participants. If the need is not urgent, the question may be asked using
mail or an online forum. House newsletters or bulletins can be used to provide
answers to frequently asked questions. More companies now have ‘‘Intranets’’
where such information is made available. The trainee may be able to find
an Internet news group devoted to the subject of their concern. There are
thousands of news groups covering a huge variety of subjects, many relating to
quality and training.
^ ^ ^
Six Sigma Teams
Six Sigma teams working on projects are the primary means of deploying Six
Sigma and accomplishing the goals of the enterprise. Six Sigma teams are sometimes
lead by the Black Belt, but the team leader is often the Green Belt or a
Six Sigma champion who has a passion for the project. Six Sigma teams are composed
of groups of individuals who bring authority, knowledge, skills, abilities
and personal attributes to the project. There is nothing particularly special
about Six Sigma teams compared with other work teams. They are people with
different backgrounds and talents pursuing a common short-term goal. Like all
groups of people, there are dynamics involved that must be understood if the
mission of the team is to be accomplished. This chapter addresses the subject
of what members, Black Belts, Green Belts, sponsors, champions, facilitators,
and leaders can do to assure that Six Sigma teams are successful. It is not a
discussion of project management techniques; these are covered elsewhere in
this book. Instead, this chapter focuses on:
. Stages in learning to work as a team
. The di?erence between group maintenance roles and group task roles
. Identifying and encouraging productive roles essential to team success
. Identifying and discouraging counterproductive behavior on teams
. Facilitating team meetings
. Dealing constructively with con?icts
. Evaluating, recognizing and rewarding teams
The structure of modern organizations is based on the principle of division of
labor. Most organizations today consist of a number of departments, each
Copyright 2003 by The McGraw-Hill Companies, Inc. Click Here for Terms of Use.
devoted to their own specialty.Afundamental problem is that the separate functional
departments tend to optimize their own operations, often to the detriment
of the organization as a whole.
Traditional organizations, in effect, create barriers between departments.
Departmental managers are often forced to compete for shares of limited
budgets; in other words, they are playing a ‘‘zero sum game’’ where another
manager’s gain is viewed as their department’s loss. Behavioral research has
shown that people engaged in zero sum games think in terms of win-lose. This
leads to self-destructive, cut-throat behavior. Overcoming this tendency
requires improved communication and cooperation between departments.
Interdepartmental teams are groups of people with the skills needed to deliver
the value desired. Processes are designed by the team to create the value in
an effective and efficient manner. Management must see to it that the needed
skills exist in the organization. It is also management’s job to see that they
remove barriers to cooperation.
There are two ways to make improvements: improve performance given the
current system, or improve the system itself. Much of the time improving performance
given the current system can be accomplished by individuals working
alone. For example, an operator might make certain adjustments to the
machine. Studies indicate that this sort of action will be responsible for about
5%^15% of the improvements. The remaining 85%^95% of all improvements
will require changing the system itself. This is seldom accomplished by individuals
working alone. It requires group action. Thus, the vast majority of Six
Sigma improvement activity will take place in a group setting. As with nearly
everything, the group process can be made more effective by acquiring a better
understanding of the way it works.
Management of cross-functional projects is discussed in Chapter 15. In this
section we will focus on the team aspect of process improvement activity.
Process improvement teams focus on improving one or more important
characteristics of a process, e.g., quality, cost, cycle time, etc. The focus is on an
entire process, rather than on a particular aspect of the process. A process is an
integrated chain of activities that add value. A process can be identified by its
beginning and ending states, e.g., manufacturing’s beginning state is procurement,
its ending state is shipment. Methods of analyzing and characterizing
process are discussed throughout this book. Usually several departments are
involved in any given value-added process.
Process improvement teams work on both incremental improvement
(KAIZEN) and radical change (breakthrough). The team is composed of mem-
bers who work with the process on a routine basis. Team members typically
report to different bosses, and their positions can be on different levels of the
organization’s hierarchy.
Process improvement projects must be approved by the process owner,
usually a senior leader in the organization. Process improvement teams must
be chartered and authorized to pursue process improvement. All of this falls
in the area of project management. Project management is discussed in
Chapter 15.
Work groups focus on improvement within a particular work area. The work
area is usually contained within a single department or unit. The process
owner is usually the department manager. Team members are usually at the
same level within the organization’s hierarchy and they usually report to one
Work group members are trained in the use of quality control techniques and
supported by management. The idea is that all workers have an important contribution
to make to the quality effort and the work group is one mechanism
for allowing them the opportunity to make their contribution.
Quality circles
An example of a work group is the quality circle. Quality circles originated
in Japan and Japanese companies continue to use quality circles on a massive
scale. Quality circles were tried on a massive scale in America, with only limited
success. However, the quality circle is the historical forerunner of the
modern quality improvement work team; a study of them reveals a great
deal about the success factors needed for successful use of other types of
work groups.
Quality circles (circles) are local groups of employees who work to continuously
improve those processes under their direct control. Here are some
necessary steps that must be completed before circles can succeed:
. Management from the top level to the supervisory level must have a clear
idea of their organization’s purpose. Everyone in the organization must
be committed to helping the organization achieve its purpose.
. Senior leadership must have an e?ective organization for dealing with
company-wide issues such as quality, cost, cycle time, etc. (e.g., the crossfunctional
form discussed earlier).
. Attention must be focused on processes rather than on internal politics
and reporting relationships.
Work Groups 169
. Personnel involved must be trained in cooperation skills (e.g., team work,
group dynamics, communication and presentation skills). This applies to
area supervisors and managers, not just circle members.
. Personnel involved must be trained in problem-solving skills (e.g., the
traditional QC tools, the 7M tools, brainstorming, etc.).
. Circle participation must be encouraged by local management.
This author believes that circles have an important place and that they can
succeed anywhere providing the proper corporate environment exists. This
environment did not exist in Western business organizations in the 1970s, and
for the most part still does not exist. Merely grafting quality circles onto a traditional
command-and-control hierarchy won’t work. There were many reasons
why quality circles failed in America; they are the same reasons why work
groups fail to this day.
1. The quality circle in an American ?rm was isolated, not part of a
company-wide quality control e?ort. As a result, circles were usually
unable to deal successfully with problems involving other areas of the
company. There were no resources in other areas to draw upon.
2. Key management personnel moved about too frequently and circles
were not provided with consistent leadership and management
3. Employees transferred in and out of circle work areas too frequently.
Without stability in the membership, circles never developed into e?ective
groups. Building e?ective teams takes time.
In addition to process-improvement teams and work groups, there are many
other types of teams and groups involved to some extent in Six Sigma. Self-managed
teams are a way to reintegrate work and flatten the management hierarchy.
If properly implemented and managed, the result can be improved quality and
productivity. If poorly implemented and managed, the result can be added
Self-managed teams are often given some of the responsibilities that, in traditional
organizations, are reserved to management. This includes the authority
to plan and schedule work, hiring, performance assessment, etc. This is essentially
a reversal of over 90 years of scientific management. While difficult to
implement successfully, the result is a leaner, more efficient organization,
higher employee morale, and better quality. Several preconditions are necessary
to assure success:
1. Communicate and listenLEncourage two-way, honest, open, frequent
communication. The more informed employees are, the more secure
and motivated they will be.
2. Train employeesLAn empowering culture is built on the bedrock of
continuing education in every form imaginable. If an employee doesn’t
know what to do, how to do it right, or most important, why it is done
a certain way and what di?erence it makes, don’t expect him to feel or
act empowered.
3. Team employeesLNo one has found a technological alternative to
cooperation when it comes to building a positive work climate. Teams
make it possible for people to participate in decision-making and implementation
that directly a?ects them.
4. Trust employeesLSupport team decisions even if they aren’t the outcomes
you had in mind. Trust teams with information and allow them
to fail.
5. FeedbackLFind people doing things right. Recognize e?orts as well as
results by ?nding ways to frequently and creatively say thank you.
Share the glory in every way possible. Give frequent speci?c performance
feedback (good news as well as bad).
Conflict management is a duty shared by the facilitator and the team leader.
The facilitator can assist the leader by assuring that creative conflict is not repressed,
but encouraged. Explore the underlying reasons for the conflict. If
‘‘personality disputes’’ are involved that threaten to disrupt the team meeting,
arrange one-on-one meetings between the parties and attend the meetings to
help mediate.
The first step in establishing an effective group is to create a consensus
decision rule for the group, namely:
No judgment may be incorporated into the group decision until it meets at
least tacit approval of every member of the group.
This minimum condition for group movement can be facilitated by adopting
the following behaviors:
. Avoid arguing for your own position. Present it as lucidly and logically as
possible, but be sensitive to and consider seriously the reactions of the
group in any subsequent presentations of the same point.
Team Dynamics Management, Including Conflict Resolution 171
. Avoid ‘‘win-lose’’ stalemates in the discussion of opinions. Discard the
notion that someone must win and someone must lose in the discussion;
when impasses occur, look for the next most acceptable alternative for
all the parties involved.
. Avoid changing your mind only to avoid con?ict and to reach agreement
and harmony. Withstand pressures to yield which have no objective or
logically sound foundation. Strive for enlightened ?exibility; but avoid
outright capitulation.
. Avoid con?ict-reducing techniques such as the majority vote, averaging,
bargaining, coin-?ipping, trading out, and the like. Treat di?erences of
opinion as indicative of an incomplete sharing of relevant information
on someone’s part, either about task issues, emotional data, or gut level
. View di?erences of opinion as both natural and helpful rather than as a hindrance
in decision-making. Generally, the more ideas expressed, the
greater the likelihood of con?ict will be; but the richer the array of
resources will be as well.
. View initial agreement as suspect. Explore the reasons underlying apparent
agreements; make sure people have arrived at the same conclusions
for either the same basic reasons or for complementary reasons before
incorporating such opinions into the group decision.
. Avoid subtle forms of in?uence and decision modi?cation. E.g., when a
dissenting member ?nally agrees, don’t feel that he must be rewarded by
having his own way on some subsequent point.
. Be willing to entertain the possibility that your group can achieve all the
foregoing and actually excel at its task. Avoid doomsaying and negative
predictions for group potential.
Collectively, the above steps are sometimes known as the ‘‘consensus
technique.’’ In tests it was found that 75% of the groups who were
instructed in this approach significantly outperformed their best individual
Stages in group development
Groups of many different types tend to evolve in similar ways. It often helps
to know that the process of building an effective group is proceeding normally.
Bruce W. Tuckman (1965) identified four stages in the development of a group:
forming, storming, norming, and performing.
During the forming stage a group tends to emphasize procedural matters.
Group interaction is very tentative and polite. The leader dominates the
decision-making process and plays a very important role in moving the group
The storming stage follows forming. Conflict between members, and
between members and the leader, are characteristic of this stage. Members question
authority as it relates to the group objectives, structure, or procedures. It
is common for the group to resist the attempts of its leader to move them toward
independence. Members are trying to define their role in the group.
It is important that the leader deal with the conflict constructively. There are
several ways in which this may be done:
. Do not tighten control or try to force members to conform to the
procedures or rules established during the forming stage. If disputes over
procedures arise, guide the group toward new procedures based on a
group consensus.
. Probe for the true reasons behind the con?ict and negotiate amore acceptable
. Serve as a mediator between group members.
. Directly confront counterproductive behavior.
. Continue moving the group toward independence from its leader.
During the norming stage the group begins taking responsibility, or ownership,
of its goals, procedures, and behavior. The focus is on working together
efficiently. Group norms are enforced on the group by the group itself.
The final stage is performing. Members have developed a sense of pride in the
group, its accomplishments, and their role in the group. Members are confident
in their ability to contribute to the group and feel free to ask for or give assistance.
Common problems
Table 5.1 lists some common problems with teams, along with recommended
remedial action (Scholtes, 1988).
Member roles and responsibilities
There are two basic types of roles assumed by members of a group: task roles
and group maintenance roles. Group task roles are those functions concerned
with facilitating and coordinating the group’s efforts to select, define, and
solve a particular problem. The group task roles shown in Table 5.2 are generally
Team Dynamics Management, Including Conflict Resolution 173
Table 5.1. Common team problems and remedial action.
Floundering  Review the plan
 Develop a plan for movement
The expert  Talk to o?ending party in private
 Let the data do the talking
 Insist on consensus decisions
Dominating participants  Structure participation
 Balance participation
 Act as gate-keeper
Reluctant participants  Structure participation
 Balance participation
 Act as gate-keeper
Using opinions instead of
facts  Insist on data
 Use scienti?c method
Rushing things  Provide constructive feedback
 Insist on data
 Use scienti?c method
Attribution (i.e., attributing
motives to people with
whom we disagree)
 Don’t guess at motives
 Use scienti?c method
 Provide constructive feedback
Ignoring some comments  Listen actively
 Train team in listening techniques
 Speak to o?ending party in private
Wanderlust  Follow a written agenda
 Restate the topic being discussed
Feuds  Talk to o?ending parties in private
 Develop or restate groundrules
Another type of role played in small groups are the group maintenance roles.
Group maintenance roles are aimed at building group cohesiveness and groupcentered
behavior. They include those behaviors shown in Table 5.3.
Team Dynamics Management, Including Conflict Resolution 175
Table 5.2. Group task roles.
Initiator Proposes new ideas, tasks, or goals; suggests procedures or
ideas for solving a problem or for organizing the group.
Information seeker Asks for relevant facts related to the problem being
Opinion seeker Seeks clari?cation of values related to problem or
Information giver Provides useful information about subject under discussion.
Opinion giver O?ers his/her opinion of suggestions made. Emphasis is on
values rather than facts.
Elaborator Gives examples.
Coordinator Shows relationship among suggestions; points out issues and
Orientor Relates direction of group to agreed-upon goals.
Evaluator Questions logic behind ideas, usefulness of ideas, or
Energizer Attempts to keep the group moving toward an action.
Procedure technician Keeps group from becoming distracted by performing such
tasks as distributing materials, checking seating, etc.
Recorder Serves as the group memory.
The development of task and maintenance roles is a vital part of the teambuilding
process. Team building is defined as the process by which a group
learns to function as a unit, rather than as a collection of individuals.
In addition to developing productive group-oriented behavior, it is also
important to recognize and deal with individual roles which may block the
building of a cohesive and effective team. These roles are shown in Table 5.4.
The leader’s role includes that of process observer. In this capacity, the leader
monitors the atmosphere during group meetings and the behavior of indivi-
Table 5.3. Group maintenance roles.
Encourager O?ers praise to other members; accepts the
contributions of others.
Harmonizer Reduces tension by providing humor or by promoting
reconciliation; gets people to explore their di?erences
in a manner that bene?ts the entire group.
Compromiser This role may be assumed when a group member’s
idea is challenged; admits errors, o?ers to modify his/
her position.
Gate-keeper Encourages participation, suggests procedures for
keeping communication channels open.
Standard setter Expresses standards for group to achieve, evaluates
group progress in terms of these standards.
Observer/commentator Records aspects of group process; helps group
evaluate its functioning.
Follower Passively accepts ideas of others; serves as audience in
group discussions.
duals. The purpose is to identify counterproductive behavior. Of course, once
identified, the leader must tactfully and diplomatically provide feedback to the
group and its members. The success of Six Sigma is, to a great extent, dependent
on the performance of groups.
Perhaps the most important thing management can do for a group is to give it
time to become effective. This requires, among other things, that management
Team Dynamics Management, Including Conflict Resolution 177
Table 5.4. Counterproductive group roles.
Aggressor Expresses disapproval by attacking the values, ideas, or
feelings of other. Shows jealousy or envy.
Blocker Prevents progress by persisting on issues that have been
resolved; resists attempts at consensus; opposes without
Recognition-seeker Calls attention to himself/herself by boasting, relating
personal achievements, etc.
Confessor Uses group setting as a forum to air personal ideologies
that have little to do with group values or goals.
Playboy Displays lack of commitment to group’s work by
cynicism, horseplay, etc.
Dominator Asserts authority by interrupting others, using ?attery to
manipulate, claiming superior status.
Help-seeker Attempts to evoke sympathy and/or assistance from
other members through ‘‘poor me’’ attitude.
Special-interest pleader Asserts the interests of a particular group. This group’s
interest matches his/her self-interest.
work to maintain consistent group membership. Group members must not be
moved out of the group without very good reason. Nor should there be a constant
stream of new people temporarily assigned to the group. If a group is to
progress through the four stages described earlier in this chapter, to the crucial
performing stage, it will require a great deal of discipline from both the group
and management.
Another area where management must help is creating an atmosphere within
the company where groups can be effective.
When to use an outside facilitator
It is not always necessary to have an outside party facilitate a group or team.
While facilitators can often be of benefit, they may also add cost and the use of
facilitators should, therefore, be carefully considered. The following guidelines
can be used to determine if outside facilitation is needed (Schuman, 1996):
1. Distrust or biasLIn situations where distrust or bias is apparent or suspected,
groups should make use of an unbiased outsider to facilitate
(and perhaps convene) the group.
2. IntimidationLThe presence of an outside facilitator can encourage the
participation of individuals who might otherwise feel intimidated.
3. RivalryLRivalries between individuals and organizations can be
mitigated by the presence of an outside facilitator.
4. Problem de?nitionLIf the problem is poorly de?ned, or is de?ned
di?erently by multiple parties, an unbiased listener and analyst can
help construct an integrated, shared understanding of the problem.
5. Human limitsLBringing in a facilitator to lead the group process lets
members focus on the problem at hand, which can lead to better results.
6. Complexity or noveltyLIn a complex or novel situation, a process
expert can help the group do a better job of working together intellectually
to solve the problem.
7. TimelinesLIf a timely decision is required, as in a crisis situation, the
use of a facilitator can speed the group’s work.
8. CostLA facilitator can help the group reduce the cost of meetingL
a signi?cant barrier to collaboration.
Selecting a facilitator
Facilitators should possess four basic capabilities (Schuman, 1996):
1. He or she should be able to anticipate the complete problem-solving and
decision-making processes.
2. He or she should use procedures that support both the group’s social and
cognitive process.
3. He or she should remain neutral regarding content issues and values.
4. He or she should respect the group’s need to understand and learn from
the problem solving process.
Facilitation works best when the facilitator:
. Takes a strategic and comprehensive view of the problem-solving and
decision-making processes and selects, from a broad array, the speci?c
methods that match the group’s needs and the tasks at hand.
. Supports the group’s social and cognitive processes, freeing the group
members to focus their attention on substantive issues.
. Is trusted by all group members as a neutral party who has no biases or
vested interest in the outcome.
. Helps the group understand the techniques being used and helps the
group improve its own problem-solving processes.
Principles of team leadership and facilitation
Human beings are social by nature. People tend to seek out the company of
other people. This is a great strength of our species, one that enabled us to rise
above and dominate beasts much larger and stronger than ourselves. It is this
ability that allowed men to control herds of livestock to hunt swift antelope,
and to protect themselves against predators. However, as natural as it is to
belong to a group, there are certain behaviors that can make the group function
more (or less) effectively than their members acting as individuals.
We will define a group as a collection of individuals who share one or more
common characteristics. The characteristic shared may be simple geography,
i.e., the individuals are gathered together in the same place at the same time.
Perhaps the group shares a common ancestry, like a family. Modern society consists
of many different types of groups. The first group we join is, of course,
our family. We also belong to groups of friends, sporting teams, churches,
PTAs, and so on. The groups differ in many ways. They have different purposes,
different time frames, and involve varying numbers of people. However, all
effective groups share certain common features. In their work, Joining
Together, Johnson and Johnson (1999) list the following characteristics of an
effective group:
. Group goals must be clearly understood, be relevant to the needs of group
members, and evoke from every member a high level of commitment to
their accomplishment.
Facilitation Techniques 179
. Group members must communicate their ideas and feelings accurately
and clearly. E?ective, two-way communication is the basis of all group
functioning and interaction among group members.
. Participation and leadership must be distributed among members. All
should participate, and all should be listened to. As leadership needs
arise, members should all feel responsibility for meeting them. The equalization
of participation and leadership makes certain that all members
will be involved in the group’s work, committed to implementing the
group’s decisions, and satis?ed with their membership. It also assures
that the resources of every member will be fully utilized, and increases
the cohesiveness of the group.
. Appropriate decision-making procedures must be used ?exibly if they are
to be matched with the needs of the situation. There must be a balance
between the availability of time and resources (such as member’s skills)
and the method of decision-making used for making the decision. The
most e?ective way of making a decision is usually by consensus (see
below). Consensus promotes distributed participation, the equalization
of power, productive controversy, cohesion, involvement, and commitment.
. Power and in?uence need to be approximately equal throughout the
group. They should be based on expertise, ability, and access to information,
not on authority. Coalitions that help ful?ll personal goals should
be formed among group members on the basis of mutual in?uence and
. Con?icts arising from opposing ideas and opinions (controversy) are to be
encouraged . Controversies promote involvement in the group’s work,
quality, creativity in decision-making, and commitment to implementing
the group’s decisions. Minority opinions should be accepted and used.
Con?icts prompted by incompatible needs or goals, by the scarcity of a
resource (money, power), and by competitiveness must be negotiated in
a manner that is mutually satisfying and does not weaken the cooperative
interdependence of group members.
. Group cohesion needs to be high. Cohesion is based on members liking
each other, each member’s desire to continue as part of the group, the
satisfaction of members with their group membership, and the level of
acceptance, support, and trust among the members. Group norms supporting
psychological safety, individuality, creativeness, con?icts of
ideas, growth, and change need to be encouraged.
. Problem-solving adequacy should be high. Problems must be resolved
with minimal energy and in a way that eliminates them permanently.
Procedures should exist for sensing the existence of problems, inventing
and implementing solutions, and evaluating the e?ectiveness of the solutions.
When problems are dealt with adequately, the problem-solving ability
of the group is increased, innovation is encouraged, and group
e?ectiveness is improved.
. The interpersonal e?ectiveness of members needs to be high.
Interpersonal e?ectiveness is a measure of how well the consequences of
your behavior match intentions.
These attributes of effective groups apply regardless of the activity in which
the group is engaged. It really doesn’t matter if the group is involved in a study
of air defense, or planning a prom dance. The common element is that there is
a group of human beings engaged in pursuit of group goals.
Facilitating the group task process
Team activities can be divided into two subjects: task-related and maintenance-
related. Task activities involve the reason the team was formed, its
charter, and its explicit goals.
The facilitator should be selected before the team is formed and he or she
should assist in identifying potential team members and leaders, and in developing
the team’s charter. The subject of team formation and project chartering
is discussed in detail in Chapter 15.
The facilitator also plays an important role in helping the team develop specific
goals based on their charter. Goal-setting is an art and it is not unusual to
find that team goals bear little relationship to what management actually had
in mind when the team was formed. Common problems are goals that are too
ambitious, goals that are too limited and goals that assume a cause and effect
relationship without proof. An example of the latter would be a team chartered
to reduce scrap assuming that Part X had the highest scrap loss (perhaps based
on a week’s worth of data) and setting as its goal the reduction of scrap for that
part. The facilitator can provide a channel of communication between the
team and management.
Facilitators can assist the team leader in creating a realistic schedule for the
team to accomplish its goals. The issue of scheduling projects is covered in
Chapter 15.
Facilitators should assure that adequate records are kept on the team’s projects.
Records should provide information on the current status of the project.
Records should be designed to make it easy to prepare periodic status reports
for management. The facilitator should arrange for clerical support with such
tasks as designing forms, scheduling meetings, obtaining meeting rooms, securing
audio visual equipment and office supplies, etc.
Facilitation Techniques 181
Other activities where the facilitator’s assistance is needed include:
Meeting managementLSchedule the meeting well ahead of time. Be sure
that key people are invited and that they plan to attend. Prepare an agenda
and stick to it! Start on time. State the purpose of the meeting clearly at the
outset. Take minutes. Summarize from time-to-time. Actively solicit input
from those less talkative. Curtail the overly talkative members. Manage conflicts.
Make assignments and responsibilities explicit and specific. End on
CommunicationLThe idea that ‘‘the quality department’’ can ‘‘assure’’ or
‘‘control’’ quality is now recognized as an impossibility. To achieve quality the
facilitator must enlist the support and cooperation of a large number of people
outside of the team. The facilitator can relay written and verbal communication
between the team and others in the organization. Verbal communication is valuable
even in the era of instantaneous electronic communication. A five minute
phone call can provide an opportunity to ask questions and receive answers
that would take a week exchanging email and faxes. Also, the team meeting is
just one communication forum, the facilitator can assist team members in communicating
with one another between meetings by arranging one-on-one meetings,
acting as a go-between, etc.
Facilitating the group maintenance process
Study the group process. The facilitator is in a unique position to stand
back and observe the group at work. Are some members dominating the
group? Do facial expressions and body language suggest unspoken disagreement
with the team’s direction? Are quiet members being excluded from the
When these problems are observed, the facilitator should provide feedback
and guidance to the team. Ask the quiet members for their ideas and input. Ask
if anyone has a problem with the team’s direction. Play devil’s advocate to
draw out those with unspoken concerns.
Evaluating team performance involves the same principles as evaluating
performance in general. Before one can determine how well the team’s task
has been done, a baseline must be established and goals must be identified.
Setting goals using benchmarking and other means is discussed elsewhere in
this book (see Chapter 2). Records of progress should be kept as the team
pursues its goals.
Performance measures generally focus on group tasks, rather than on internal
group issues. Typically, financial performance measures show a payback
ratio of between 2:1 and 8:1 on team projects. Some examples of tangible
performance measures are:
. productivity
. quality
. cycle time
. grievances
. medical usage (e.g., sick days)
. absenteeism
. service
. turnover
. dismissals
. counseling usage
Many intangibles can also be measured. Some examples of intangibles
effected by teams are:
. employee attitudes
. customer attitudes
. customer compliments
. customer complaints
The performance of the team process should also be measured. Project
failure rates should be carefully monitored. A p chart can be used to evaluate
the causes of variation in the proportion of team projects that succeed.
Failure analysis should be rigorously conducted.
Aubrey and Felkins (1988) list the effectiveness measures shown below:
. leaders trained
. number of potential volunteers
. number of actual volunteers
. percent volunteering
. projects started
. projects dropped
. projects completed/approved
. projects completed/rejected
. improved productivity
. improved work environment
. number of teams
. inactive teams
. improved work quality
. improved service
. net annual savings
Team Performance Evaluation 183
Recognition is a form of employee motivation in which the company identifies
and thanks employees who have made positive contributions to the company’s
success. In an ideal company, motivation flows from the employees’
pride of workmanship. When employees are enabled by management to do
their jobs and produce a product or service of excellent quality, they will be
The reason recognition systems are important is not that they improve work
by providing incentives for achievement. Rather, they make a statement about
what is important to the company. Analyzing a company’s employee recognition
system provides a powerful insight into the company’s values in action.
These are the values that are actually driving employee behavior. They are not
necessarily the same as management’s stated values. For example, a company
that claims to value customer satisfaction but recognizes only sales achievements
probably does not have customer satisfaction as one of its values in
Public recognition is often better for two reasons:
1. Some (but not all) people enjoy being recognized in front of their colleagues.
2. Public recognition communicates a message to all employees about the
priorities and function of the organization.
The form of recognition can range from a pat on the back to a small gift to a
substantial amount of cash. When substantial cash awards become an established
pattern, however, it signals two potential problems:
1. It suggests that several top priorities are competing for the employee’s
attention, so that a large cash award is required to control the employee’s
2. Regular, large cash awards tend to be viewed by the recipients as part of
the compensation structure, rather than as a mechanism for recognizing
support of key corporate values.
Carder and Clark (1992) list the following guidelines and observations
regarding recognition:
Recognition is not a method by which management can manipulate employees.
If workers are not performing certain kinds of tasks, establishing a recognition
program to raise the priority of those tasks might be inappropriate.
Recognition should not be used to get workers to do something they are not currently
doing because of conflicting messages from management. A more effective
approach is for management to first examine the current system of
priorities. Only by working on the system can management help resolve the
Recognition is not compensation. In this case, the award must represent
a significant portion of the employee’s regular compensation to have significant
impact. Recognition and compensation differ in a variety of
. Compensation levels should be based on long-term considerations such as
the employee’s tenure of service, education, skills, and level of responsibility.
Recognition is based on the speci?c accomplishments of individuals
or groups.
. Recognition is ?exible. It is virtually impossible to reduce pay levels once
they are set, and it is di?cult and expensive to change compensation plans.
. Recognition is more immediate. It can be given in timely fashion and
therefore relate to speci?c accomplishments.
. Recognition is personal. It represents a direct and personal contact
between employee and manager. Recognition should not be carried out
in such a manner that implies that people of more importance (managers)
are giving something to people of less importance (workers).
Positive reinforcement is not always a good model for recognition. Just because
the manager is using a certain behavioral criterion for providing recognition, it
doesn’t mean that the recipient will perceive the same relationship between
behavior and recognition.
Employees should not believe that recognition is based primarily on luck. An
early sign of this is cynicism. Employees will tell you that management says
one thing but does another.
Recognition meets a basic human need. Recognition, especially public recognition,
meets the needs for belonging and self-esteem. In this way, recognition
can play an important function in the workplace. According to Abraham
Maslow’s theory, until these needs for belonging and self-esteem are satisfied,
self-actualizing needs such as pride in work, feelings of accomplishment,
personal growth, and learning new skills will not come into play.
Recognition programs should not create winners and losers. Recognition programs
should not recognize one group of individuals time after time while
never recognizing another group. This creates a static ranking system, with all
of the problems discussed earlier.
Recognition should be given for efforts, not just for goal attainment.
According to Imai (1986), a manager who understands that a wide variety
of behaviors are essential to the company will be interested in criteria
of discipline, time management, skill development, participation, morale,
and communication, as well as direct revenue production. To be able to
effectively use recognition to achieve business goals, managers must
develop the ability to measure and recognize such process accomplishments.
Team Recognition and Reward 185
Employee involvement is essential in planning and executing a recognition
program. It is essential to engage in extensive planning before instituting a
recognition program or before changing a bad one. The perceptions and expectations
of employees must be surveyed.
^ ^ ^
Selecting and Tracking Six
Sigma Projects*
The best Six Sigma projects begin not inside the business but outside it,
focused on answering the question: How can we make the customer more
competitive? What is critical to the customer’s success? Learning the answer
to that question and learning how to provide the solution is the only focus
we need.
JackWelch, CEO, General Electric
This chapter covers the subject of Six Sigma project selection. Project management,
monitoring, results capture, and lessons learned capture and dissemination
are discussed in Chapter 15. Projects are the core activity driving
change in the Six Sigma organization. Although change also takes place due to
other efforts, such as Kaizen, project-based change is the force that drives breakthrough
and cultural transformation. In a typical Six Sigma organization
about one percent of the workforce is engaged full time in change activities,
and each of these change agents will complete between three and seven projects
in a year. In addition there are another five percent or so part-time change
agents, each of whom will complete about two smaller projects per year. The
mathematics translate to about 500 major projects and 1,000 smaller projects
*Some of the material in this chapter is from The Six Sigma Project Planner, by Thomas Pyzdek.#2003 by McGraw-Hill.
Copyright 2003 by The McGraw-Hill Companies, Inc. Click Here for Terms of Use.
in an organization with 10,000 employees in any given year. Clearly, learning
how to effectively deal with projects is critical to Six Sigma success.
Projects must be focused on the right goals. This is the responsibility of the
senior leadership, e.g., the project sponsor, Executive Six Sigma Council or
equivalent group. Senior leadership is the only group with the necessary authority
to designate cross-functional responsibilities and allow access to interdepartmental
resources. Six Sigma projects will impact one of the major stakeholder
groups: customers, shareholders, or employees. Although it is possible to calculate
the impact of any given project on all three groups, I recommend that initially
projects be evaluated separately for each group. This keeps the analysis
relatively simple and assures that a good stakeholder mix is represented in the
project portfolio.
Customer value projects
Many, if not most Six Sigma projects are selected because they have a positive
impact on customers. To evaluate such projects one must be able to determine
the linkage between business processes and customer-perceived value. Chapter
3 discussed how to create organizations that are customer-driven, which is
essential. Customer-driven organizations, especially process enterprises, focus
on customer value as a matter of routine. This focus will generate many Six
Sigma customer value projects in the course of strategy deployment. However,
in addition to the strategy-based linkage of Six Sigma projects described in
Chapter 3, there is also a need to use customer demands directly to generate
focused Six Sigma projects. The techniques for obtaining this linkage are the
same as those used in Chapter 3. The difference is that the focus here is not on
strategy deployment or budgeting, but on Six Sigma improvement projects
focused on specific customer demands.
Learning what customers value is primarily determined by firsthand contact
with customers through customer focus groups, interviews, surveys, etc. The
connection between customer-perceived value and business processes, or ‘‘customer
value streams,’’ is established through business process mapping (see
Chapter 8) and quality function deployment (QFD). The Executive Six Sigma
Council and project sponsors should carefully review the results of these efforts
to locate the ‘‘lever points’’ where Six Sigma projects will have the greatest
impact on customer value.
Shareholder value projects
Six Sigma provides a ‘‘double-whammy’’ by addressing both efficiency and
revenues. Revenue is impacted by improving the customer value proposition,
which allows organizations to charge premium prices for superior quality, or to
keep prices competitive and increase sales volume and market share due to superior
quality. Improved efficiency is achieved by reducing the cost of poor quality,
reducing cycle time, or eliminating waste in business processes. To determine
which Six Sigma projects address the issue of business process efficiency evaluate
the high-level business process maps (including SIPOC) and flow charts.
Other Six Sigma projects
Some Six Sigma projects address intangibles, such as employee morale, regulatory
concerns, or environmental issues. These projects can be just as important
as those which address customer or shareholder value.
You now have a list of candidate Six Sigma projects. Assuming that the organization
has limited resources, the next task is to select a subset of these projects
to fund and staff.
Projects cost money, take time, and disrupt normal operations and standard
routines. For these reasons projects designed to improve processes should be
limited to processes that are important to the enterprise. Furthermore, projects
should be undertaken only when success is highly likely. Feasibility is determined
by considering the scope and cost of a project and the support it receives
from the process owner. In this section a number of techniques and approaches
are presented to help identify those projects that will be chosen for Six Sigma.
Benefit-cost analysis
Benefit-cost analysis can be as elaborate or as simple as the magnitude of the
project expenditures demands. The Six Sigma manager is advised that most
such analyses are easier to ‘‘sell’’ to senior management if done by (or reviewed
and approved by) experts in the finance and accounting department. The plain
fact is that the finance department has credibility in estimating cost and benefit
that the Six Sigma department, and any other department, lacks. The best
approach is to get the finance department to conduct the benefit-cost analysis
with support from the other departments involved in the project. We will provide
an overview of some principles and techniques that are useful in benefitcost
Analyzing Project Candidates 189
A fundamental problem with performing benefit-cost analysis is that, in general,
it is easier to accurately estimate costs than benefits. Costs can usually be
quantified in fairly precise terms in a budget. Costs are claims on resources the
firm already has. In contrast, benefits are merely predictions of future events,
which may or may not actually occur. Also, benefits are often stated in units
other than dollars, making the comparison of cost and benefit problematic. The
problem is especially acute where quality improvement projects are concerned.
For example, a proposed project may involve placing additional staff on a customer
‘‘hot line.’’ The cost is easy to compute: X employees at a salary of $Y each,
equipment, office space, supervision, etc. The benefit is much more difficult to
determine. Perhaps data indicate that average time on hold will be improved,
but the amount of the improvement and the probability that it will occur are
speculations. Even if the time-on-hold improvement were precise, the impact
on customer satisfaction would be an estimate. And the association between customer
satisfaction and revenues is yet another estimate. Still, the intelligent manager
will realize that despite these difficulties, reasonable cause-and-effect
linkages can be established to form the basis for benefit-cost analysis. Such is
often the best one can expect. To compensate for the uncertainties associated
with estimates of benefits, it makes sense to demand a relatively high ratio of benefit
to cost. For example, it is not unusual to have senior leadership demand a
ROI of 100% in the first year on a Six Sigma project. Rather than becoming distressed
at this ‘‘injustice,’’ the Black Belt should realize that such demands are a
response to the inherent difficulties in quantifying benefits.
A system for assessing Six Sigma projects
Assessing Six Sigma projects is an art as well as a science. It is also critical to
the success of Six Sigma, and to the individual Black Belt. Far too many Black
Belts fail because they are not discriminating enough in their selection of projects.
If project selection is systematically sloppy, the entire Six Sigma effort
can fail.
The approach offered here is quantitative in the sense that numbers are determined
and an overall project score calculated. It is subjective to a degree because
it requires interpretation of the situation, estimating probabilities, costs, and
commitments, etc. However, the rigor that goes with completing this assessment
process will help you make better judgments regarding projects. The numbers
(weights, scores, acceptable length of projects, dollar cutoffs, etc.) are
strictly my own personal judgments; feel free to assign your own values or
those of your leadership. The scale ranges from 0 to 9 for each criterion, and
the weights sum to 1.00, so the highest possible weighted score for a project is 9.
The Six Sigma department or Process Excellence function can compile summary
listings of project candidates from these assessments. Sorting the list in
descending order provides a guide to the final decision as to which projects to
pursue. Each Black Belt or Green Belt will probably have their own list, which
can also be sorted and used to guide their choices.
Analyzing Project Candidates 191
Worksheet 1. Six Sigma project evaluation.
Project Name: Date of Assessment:
Black Belt: Master Black Belt:
Weighted Overall Project Score: ProjectNumber:
Criteria Score Weight
1. Sponsorship 0.23
2. Benefits (specify main beneficiary)
& 2.1 External Customer:
& 2.2 Shareholder:
& 2.3 Employee or internal customer:
& 2.4 Other (e.g., supplier, environment):
Benefit Score
3. Availability of resources other than team 0.16
4. Scope in terms of Black Belt Effort 0.12
5. Deliverable 0.09
6. Time to complete 0.09
7. Team membership 0.07
8. Project Charter 0.03
9. Value of Six Sigma Approach 0.02
TOTAL(sum of weighted score column) 1.00
Note: Any criterion scores of zero must be addressed before project is approved.
*Weighted score ?project’s score for each criterion times the weight.
Worksheet 2. Six Sigma project evaluation guidelines.
1.0 Sponsorship
Score Interpretation
9 Director-level sponsor identified, duties specified and sufficient
time committed and scheduled
3 Director-level sponsor identified, duties specified and sufficient
time committed but not scheduled
1 Willing Director-level sponsor who has accepted charter
0 Director-level sponsor not identified, or sponsor has not
accepted the charter
2.0 Stakeholder Benefits*
‘‘Tangible and verifiable benefits for a major stakeholder’’
2.1 Stakeholder: External Customer
2.1.1 Customer Satisfaction
Score Interpretation
9 Substantial and statistically significant increase in overall
customer satisfaction or loyalty
3 Substantial and statistically significant increase in a major
subcategory of customer satisfaction
1 Substantial and statistically significant increase in a focused
area of customer satisfaction
0 Unclear or non-existent customer satisfaction impact
*Note: Several stakeholder benefit categories are shown in section 2. At least one stakeholder category is required. Show benefit
scores for each category, then use your judgment to determine an overall benefit score for the project.
Analyzing Project Candidates 193
2.1.2 Quality Improvement (CTQ)
Score Interpretation
9 10or greater improvement in critical to quality (CTQ) metric
5 5to 10improvement in CTQ metric
3 2to 5improvement in CTQ metric
1 Statistically significant improvement in CTQ metric, but less than
0 Project’s impact on CTQ metrics undefined or unclear
2.2 Stakeholder: Shareholder
2.2.1 Financial Benefits
Score Interpretation
9 Hard net savings (Budget or Bid Model change) greater than
$500K. Excellent ROI
5 Hard net savings between $150K and $500K. Excellent ROI
3 Hard net savings between $50K and $150K, or cost avoidance
greater than $500K. Good ROI
1 Hard savings of at least $50K, or cost avoidance of between $150K
and $500K. Acceptable ROI
0 Project claims a financial benefit but has hard savings less than
$50K, cost avoidance less than $150K, or unclear financial benefit
2.2.2 Cycle Time Reduction
Score Interpretation
9 Cycle time reduction that improves revenue, Bid Model or Budget
by more than $500K. Excellent ROI
5 Cycle time reduction that improves revenue, Bid Model or Budget
by $150K to $500K. Excellent ROI
Continued on next page . . .
3 Cycle time reduction that improves revenue, Bid Model or Budget
by $50K to $150K, or creates a cost avoidance of more than $500K.
Good ROI
1 Cycle time reduction that results in cost avoidance between $150K
and $500K. Acceptable ROI
0 Project claims a cycle time improvement but has hard savings less
than $50K, cost avoidance less than $150K, or unclear financial
benefit from the improvement in cycle time
2.2.3 Revenue Enhancement
Score Interpretation
9 Significant increase in revenues, excellent ROI
3 Moderate increase in revenues, good ROI
1 Increase in revenues with acceptable ROI
0 Unclear or non-existent revenue impact
2.3 Stakeholder: Employee or Internal Customer
2.3.1 Employee Satisfaction
Score Interpretation
9 Substantial and statistically significant increase in overall
employee satisfaction
3 Substantial and statistically significant increase in a major element
of employee satisfaction
1 Substantial and statistically significant increase in a focused area
of employee satisfaction
0 Unclear or non-existent employee satisfaction impact
2.2.2 (cont.)
Analyzing Project Candidates 195
2.4 Stakeholder: Other
2.4.1 Specify Stakeholder: ____________________________________________________________________________________________________________________________________________
Score Interpretation
0 Unclear or non-existent benefit
3.0 Availability of Resources Other Than Team
Score Interpretation
9 Needed resources available when needed
3 Limited or low priority access to needed to resources
0 Resources not available, or excessive restrictions on access to
4.0 Scope in Terms of Black Belt Effort
Score Interpretation
9 Projected return substantially exceeds required return
3 Projected return exceeds required return
1 Projected return approximately equals required return
0 Projected return not commensurate with required return
Required return can be calculated as follows:*
(1) Length of project (months) = ____________________________________________________________________________
(2) Proportion of Black Belt’s time required (between 0 and 1) = ________________________________
(3) Probability of success (between 0 and 1) = ___________________________________
Required return** = $83,333(1) (2) 	(3) = $ ______________________________________________
Projected return: $________________________________________
5.0 Deliverable (Scope)
Score Interpretation
9 New or improved process, product or service to be created is
clearly and completely defined
3 New or improved process, product or service to be created is
0 Deliverable is poorly or incorrectly defined. For example, a
‘‘deliverable’’ that is really a tool such as a process map
*Thanks toTony L in of Boeing Satellite Systems for this algorithm.
**Based on expected Black Belt results of $1million/year.
Analyzing Project Candidates 197
6.0 Time to Complete
Score Interpretation
9 Results realized in less than 3 months
3 Results realized in between 3 and 6 months
1 Results realized in 7 to 12 months
0 Results will take more than 12 months to be realized
7.0 Team Membership
Score Interpretation
9 Correct team members recruited and time commitments scheduled
3 Correct team members recruited, time committed but not
1 Correct team members recruited
0 Team members not recruited or not available
8.0 Project Charter
Score Interpretation
9 All elements of the project charter are complete and acceptable.
Linkage between project activities and deliverable is clear
3 Project charter acceptable with minor modifications
0 Project charter requires major revisions
9.0 Value of Six Sigma Approach (DMAIC or equivalent)
Score Interpretation
9 Six Sigma approach essential to the success of the project. Black
Belt/Green Belt skill set required for success
3 Six Sigma approach helpful but not essential. Black Belt/Green
Belt skill set can be applied
0 Usefulness of Six Sigma approach not apparent. Specific Black Belt
or Green Belt skills are not necessary
Other methods of identifying promising
Projects should be selected to support the organization’s overall strategy
and mission. Because of this global perspective most projects involve the
efforts of several different functional areas. Not only do individual projects
tend to cut across organizational boundaries, different projects are often
related to one another. To effectively manage this complexity it is necessary
to integrate the planning and execution of projects across the entire enterprise.
One way to accomplish this is QFD, which is discussed in detail elsewhere
in this book (see Chapter 3, ‘‘Using QFD to link Six Sigma projects
to strategies’’). In addition to QFD and the scoring method described above,
a number of other procedures are presented here to help identify a project’s
potential worth.
Pareto principle refers to the fact that a small percentage of processes
cause a large percentage of the problems. The Pareto principle is useful in
narrowing a list of choices to those few projects that offer the greatest
potential (see Chapter 8). When using Pareto analysis keep in mind that
there may be hidden ‘‘pain signals.’’ Initially problems create pain signals
such as schedule disruptions and customer complaints. Often these symptoms
are treated rather than their underlying ‘‘diseases’’; for example, if
quality problems cause schedule slippages which lead to customer complaints,
the ‘‘solution’’ might be to keep a large inventory and sort the
good from the bad. The result is that the schedule is met and customers
stop complaining, but at huge cost. These opportunities are often greater
than those currently causing ‘‘pain,’’ but they are now built into business
systems and therefore very difficult to see. One solution to the hidden problem
phenomenon is to focus on processes rather than symptoms. Some
guidelines for identifying dysfunctional processes for potential improvement
are shown in Table 6.1.
The ‘‘symptom’’ column is useful in identifying problems and setting priorities.
The ‘‘disease’’ column focuses attention on the underlying causes of the
problem, and the ‘‘cure’’ column is helpful in chartering quality improvement
project teams and preparing mission statements.
After a serious search for improvement opportunities the organization’s leaders
will probably find themselves with more projects to pursue than they
have resources. The Pareto Priority Index (PPI) is a simple way of prioritizing
these opportunities. The PPI is calculated as follows (Juran and Gryna, 1993,
p. 49):
Savings  probability of success
Cost  time to completion (years) ?6:1?
Analyzing Project Candidates 199
Table 6.1. Dysfunctional process symptoms and underlying diseases.
Symptom Disease Cure
Extensive information
exchange, data redundancy,
Arbitrary fragmentation of
a natural process
Discover why people need
to communicate with each
other so often; integrate
the process
Inventory, buffers, and
other assets stockpiled
Systemslack to cope with
Remove the uncertainty
High ratio of checking and
control to value-added
work (excessive test and
inspection, internal
controls, audits, etc.)
Fragmentation Eliminate the
fragmentation, integrate
Rework and iteration Inadequate feedback in a
long work process
Process control
Complexity, exceptions and
special causes
Accretion onto a simple
Uncover original ‘‘clean’’
process and create new
process(es) for special
situations; eliminate
excessive standardization
of processes
A close examination of the PPI equation shows that it is related to return
on investment adjusted for probability of success. The inputs are, of course,
estimates and the result is totally dependent on the accuracy of the inputs.
The resulting number is an index value for a given project. The PPI values
allow comparison of various projects. If there are clear standouts the PPI
can make it easier to select a project. Table 6.2 shows the PPIs for several
hypothetical projects.
The PPI indicates that resources be allocated first to reducing wave solder
defects, then to improving NC machine capability, and so on. The PPI may
not always give such a clear ordering of priorities. When two or more projects
have similar PPIs a judgment must be made on other criteria.
Table 6.2. Illustration of the Pareto Priority Index (PPI).
Savings $
thousands Probability
Cost, $
thousands Time, years PPI
Reduce wave
solder defects
$70 0.7 $25 0.75 2.61
NC machine
$50 0.9 $20 1.00 2.25
ISO 9001
$150 0.9 $75 2.00 0.90
$250 0.5 $75 1.50 1.11
defects 50%
$90 0.7 $30 1.50 1.40
Throughput-based project selection
While careful planning and management of projects is undeniably important,
they matter little if the projects being pursued have no impact on the bottom
line (throughput). As you will see below, if you choose the wrong projects
it is possible to make big ‘‘improvements’’ in quality and productivity that
have no impact whatever on the organization’s net profit. Selecting which projects
to pursue is of critical importance. In this section we will use the theory
of constraints (TOC) to determine which project(s) to pursue.
Every organization has constraints. Constraints come in many forms. When
a production or service process has a resource constraint (i.e., it lacks a sufficient
quantity of some resource to meet the market demand), then the sequence of
improvement projects should be identified using very specific rules. According
to Eliyahu M. Goldratt (1990), the rules are:
1. Identify the system’s constraint(s). Consider a ?ctitious company that
produces only two products, P and Q (Figure 6.1). The market demand
for P is 100 units per week and P sells for $90 per unit. The market
demand for Q is 50 units per week and Q sells for $100 per unit.
Assume that A, B, C and D are workers who have di?erent non-inter-
Analyzing Project Candidates 201
Figure 6.1. A simple process with a constraint.
changeable skills and that each worker is available for only 2,400 minutes
per week (8 hours per day, 5 days per week). For simplicity, assume that
there is no variation, waste, etc. in the process. This process has a constraint,
Worker B. This fact has profound implications for selecting Six
Sigma projects.
2. Decide how to exploit the system’s constraint(s). Look for Six Sigma projects
that minimize waste of the constraint. For example, if the constraint
is the market demand, then we look for Six Sigma projects that
provide 100% on time delivery. Let’s not waste anything! If the constraint
is a machine, focus on reducing setup time, eliminating scrap, and keeping
the machine running as much as possible.
3. Subordinate everything else to the above decision. Choose Six Sigma projects
to maximize throughput of the constraint. After completing step
2, choose projects to eliminate waste from downstream processes; once
the constraint has been utilized to create something we don’t want to
lose it due to some downstream blunder. Then choose projects to assure
that the constraint is always supplied with adequate non-defective
resources from upstream processes. We pursue upstream processes last
because by de?nition they have slack resources, so small amounts of
waste upstream that are detected before reaching the constraint are less
damaging to throughput.
4. Elevate the system’s constraint(s). Elevate means ‘‘Lift the restriction.’’
This is step #4, not step #2! Often the projects pursued in steps 2 and
3 will eliminate the constraint. If the constraint continues to exist after
performing steps 2 and 3, look for Six Sigma projects that provide additional
resources to the constraint. These might involve, for example, purchasing
additional equipment or hiring additional workers with a
particular skill.
5. If, in the previous steps, a constraint has been broken, go back to step 1.
There is a tendency for thinking to become conditioned to the existence
of the constraint. A kind of mental inertia sets in. If the constraint has
been lifted, then you must rethink the entire process from scratch.
Returning to step 1 takes you back to the beginning of the cycle.
It can be shown that the TOC approach is superior to the traditional TQM
approaches to project selection. For example, consider the data in the table
below. If you were to apply Pareto analysis to scrap rates you would begin
with Six Sigma projects that reduced the scrap produced by Worker A. In fact,
assuming the optimum product mix, Worker A has about 25% slack time, so
the scrap loss can be made up without shutting down Worker B, who is the constraint.
The TOC would suggest that the scrap loss of Worker B and the downstream
processes C and D be addressed first, the precise opposite of what
Pareto analysis recommends.
Process Scrap Rates.
Process A B C D
Scrap Rate 8% 3% 5% 7%
Of course, before making a decision as to which projects to finance cost/benefit
analyses are still necessary, and the probability of the project succeeding
must be estimated. But by using the TOC you will at least know where to look
first for opportunities.
Applying the TOC strategy described above tells us where in the process to
focus. Adding CTx information (see Table 6.3) can help tell us which type of
project to focus on, i.e., should we focus on quality, cost or schedule projects?
Assume that you have three Six Sigma candidate projects, all focusing on process
step B, the constraint. The area addressed is correct, but which project
should you pursue first? Let’s assume that we learn that one project will primarily
improve quality, another cost, and another schedule. Does this new information
help? Definitely! Take a look at Table 6.3 to see how this information can
be used. Projects in the same priority group are ranked according to their impact
on throughput.
The same thought process can be applied to process steps before and after the
constraint. The results are shown in Table 6.4.
Note that Table 6.4 assumes that projects before the constraint do not
result in problems at the constraint. Remember, impact should always be
measured in terms of throughput. If a process upstream from the constraint
has an adverse impact on throughput, then it can be considered to be a constraint.
If an upstream process average yield is enough to feed the constraint
on the average, it may still present a problem. For example, an upstream process
producing 20 units per day with an average yield of 90% will produce,
on average, 18 good units. If the constraint requires 18 units, things will be
Analyzing Project Candidates 203
Table 6.3. Throughput priority of CTx projects that a?ect the constraint.
Type Discussion
CTQ Any unit produced by the constraint is especially valuable because if it is lost
as scrap additional constraint time must be used to replace it or rework it.
Since constraint time determines throughput (net profit of the entire system),
the loss far exceeds what appears on scrap and rework reports. CTQ projects
at the constraint are the highest priority.
CTS CTS projects can reduce the time it takes the constraint to produce a unit,
which means that the constraint can produce more units. This directly
impacts throughput. CTS projects at the constraint are the highest priority.
CTC Since the constraint determines throughput, the cost of the constraint going
down is the lost throughput of the entire system. This makes the cost of
constraint down time extremely high. The cost of operating the constraint is
usually miniscule by comparison. Also, CTC projects often have an adverse
impact on quality or schedule. Thus, CTC projects at the constraint are low
Table 6.4. Project throughput priority versus project focus.
Focus of Six Sigma Project
addressed is
critical to . . .
Before the
At the
After the
Quality (CTQ) ~ 8 8
Cost (CTC) * ~ *
Schedule (CTS) ~ 8 *
~Low throughput priority.
*Moderate throughput priority.
8High throughput priority.
okay about 50% of the time, but the other 50% of the time things won’t be
okay. One solution to this problem is to place a work-in-process (WIP) inventory
between the process and the constraint as a safety buffer. Then on those
days when the process yield is below 18 units, the inventory can be used to
keep the constraint running. However, there is a cost associated with carrying
a WIP inventory. A Six Sigma project that can improve the yield will reduce
or eliminate the need for the inventory and should be considered even if it
doesn’t impact the constraint directly, assuming the benefit-cost analysis justifies
the project. On the other hand, if an upstream process can easily make
up any deficit before the constraint needs it, then a project for the process
will have a low priority.
Knowing the project’s throughput priority will help you make better
project selection decisions by helping you select from among project candidates.
Of course, the throughput priority is just one input into the project
selection process, other factors may lead to a different decision. For example,
impact on other projects, a regulatory requirement, a better payoff in the
long-term, etc.
Multi-tasking and project scheduling
A Six Sigma enterprise will always have more projects to pursue than it has
resources to do them. The fact that resources (usually Black Belts or Green
Belts) are scarce means that projects must be scheduled, i.e., some projects
must be undertaken earlier than others. In such situations it is tempting to use
multi-tasking of the scarce resource. Multi-tasking is defined as the assignment
of a resource to several priorities during the same period of time. The logic is
that by working on several projects or assignments simultaneously, the entire
portfolio of work will be done more quickly. However, while this is true for
independent resources working independent projects or subprojects in parallel,
it is not true when applied to a single resource assigned to multiple projects or
interdependent tasks within a project.
Consider the following situation. You have three Six Sigma projects, A, B,
and C. A single-tasking solution is to first do A, then B, and then C. Here’s the
single-activity project schedule.
(Complete in wk. 10)
(Complete in wk. 20)
(Complete in wk. 30)
If each project takes 10 weeks to complete, then A will be completed in 10
weeks, B in 20 weeks, and C in 30 weeks. The average time to complete the
Analyzing Project Candidates 205
three projects is ?10 ? 20 ? 30?=3 ? 60=3 ? 20 weeks. The average doesn’t tell
the whole story, either. The benefits will begin as soon as the project is completed
and by the end of the 30 week period project A will have been completed
for 20 weeks, and project B for 10 weeks.
Now let’s consider a multi-tasking strategy. Here we split our time equally
between the three projects in a given 10 week period. That way the sponsor of
projects B and C will see activity on their projects much sooner than if we used
a single-task approach to scheduling. The new schedule looks like this:
With this multi-tasking schedule project A will be completed in 23.3 weeks,
project B in 26.7 weeks, and project C will still take 30 weeks. The completion
time for project A went from 10 weeks to 23.3 weeks, for project B it went
from 20 weeks to 26.7 weeks, and for project C it remained the same, 30
weeks. The overall average completion time went from 20 weeks to 26.67
weeks, a 33% deterioration in average time to complete. And this is a bestcase
scenario. In real life there is always some lost time when making the transition
from one project to another. The Black Belt has to clear her head of
what she was doing, review the next project, get the proper files ready, reawaken
sponsors and team members, and so on. This can often take a considerable
amount of time, which is added to the time needed to complete the
Critical chain project management avoids the multi-tasking problem by
changing the way the organization manages groups of projects, and the way
the individual projects are managed.
Managing the organization’s projects
First, at the organizational level, multi-tasking of key resources is
stopped. People and other resources are allowed to focus on projects one at
a time. This means that management must accept responsibility for prioritizing
projects, and policies must be developed which mandate single-project
focus and discourage multi-tasking. To be successful the organization must
determine its capacity to complete projects. Every organization finds itself
with more opportunities than it can successfully pursue with finite
resources. This means that only a select portfolio of projects should be
undertaken in any time interval. The constraining resource is usually a key
position in the organization, say the time available by project sponsors, engineers,
programmers, etc. This information can be used to determine organizational
capacity and to schedule project start dates according to the
availability of the key resource. This is called project launch synchronization
and the scarce resource that paces the project system is called a synchronizer
Synchronizer resource usage
Critical chain project management does not permit multi-tasking of scarce
resources. People and equipment that are fully utilized on projects, synchronizer
resources, are assigned to a sequence of single projects. The sequence of
projects is based on enterprise priorities. If a project requires one or more synchronizer
resources it is vital that your project start dates integrate the schedules
of these resources. In particular, this will require that those activities
that require time from a synchronizer resource (and the project as a whole)
stipulate ‘‘Start no earlier than’’ dates. Although synchronizer resources are
protected by capacity buffers and might hypothetically start at a date earlier
than specified, the usual practice is to utilize any unplanned excess capacity
to allow the organization to pursue additional opportunities, thereby increasing
the organization’s capacity to complete projects. Note that human
resources are defined in terms of the skills required for the activity, not in
terms of individual people. In fact, the resource manager should refrain from
assigning an activity to an individual until all predecessors have been completed
and the activity is ready to begin. This precludes the temptation to
multi-task as the individual looks ahead and sees the activity start date drawing
Project start dates are determined by beginning with the highest priority
project and calculating the end date for the synchronizing resource based
on the estimated duration of all activities that require the synchronizing
resource. The second highest priority project’s start date will be calculated
by adding a capacity buffer to the expected end date of the first project.
The third highest priority project’s start date is based on the completion
date of the second, and so on. If, by chance, the synchronizing resource is
available before the scheduled start date, the time can be used to increase
the organization’s capacity to complete more projects. Figure 6.2 illustrates
this strategy.
Analyzing Project Candidates 207
Summary and preliminary project selection
At this point you have evaluated project candidates using a number of
different criteria. You must now rank the projects, and make your preliminary
selections. You may use Worksheet 3 to assist you with this. The reason your
selections are preliminary is that you lack complete data. As they work the project,
Six Sigma project teams will continuously reevaluate it and they may
uncover data which will lower or raise the project’s priority. The project sponsor
is responsible for coordinating changes in priority with the process owners.
It is vital that information regarding results be accumulated and reported.
This is useful for a variety of purposes:
& Evaluating the e?ectiveness of the Six Sigma project selection system
& Determining the overall return on investment
& Setting budgets
Figure 6.2. Critical chain scheduling illustration.
& Appraising individual and group performance
& Setting goals and targets
& Identifying areas where more (or less) emphasis on Six Sigma is indicated
& Helping educate newcomers on the value of Six Sigma
& Answering skeptics
& Quieting cynics
A major difference between Six Sigma and failed programs of the past is the
emphasis on tangible, measurable results. Six Sigma advocates make a strong
point of the fact that projects are selected to provide a mixture of short- and
long-term paybacks that justify the investment and the effort. Unless proof is
provided any statements regarding paybacks are nothing more than empty
Data storage is becoming so inexpensive that the typical organization can
afford to keep fairly massive amounts of data in databases. The limiting factor
is the effort needed to enter the data into the system. This is especially important
if highly trained change agents such as Master Black Belts, Black Belts, or
Green Belts are needed to perform the data entry (Table 6.5).
Usually viewing access is restricted to the project data according to role
played in the project, position in the organization, etc. Change access is usually
restricted to the project sponsor, leader, or Black Belt. However, to the extent
possible, it should be easy to ‘‘slice-and-dice’’ this information in a variety of
ways. Periodic reports might be created summarizing results according to
department, sponsor, Black Belt, etc. The system should also allow ad-hoc
views to be easily created, such as the simple list shown in Table 6.6.
Tracking Six Sigma Project Results 209
Worksheet 3. Project assessment summary.
Project Description
or ID Number
Priority Comments
Table 6.5. Possible information to be captured.
&Charter information (title, sponsor, membership, deadline etc.)
&Description of project in ordinary language
&Project status
&Savings type (hard, soft, cost avoidance, CTQ, etc.)
&Process or unit owner
&Key accounting information (charge numbers, etc.)
& Project originator
&Top-level strategy addressed by project
&Comments, issues
&Lessons learned
&Keywords (for future searches)
&Related documents and links
&Audit trail of changes
&Project task and schedule information
Table 6.6. A typical view of Six Sigma projects.
Title Status Black Belt Sponsor Due
Savings Costs
76 Cup
J Jones Jane Doe 3/1/04 Hard $508,000 $5,900
33 Tank
Define B Olson Sam
9/30/03 Hard $250,000 $25,000
35 SSPA Completed NHepburn Sal Davis 10/31/03 Cost
Control MLittleton Henry
9/30/03 Other NA $1,500
Financial results validation
Six Sigma financial benefits claimed for every project must be confirmed by
experts in accounting or finance. Initial savings estimates may be calculated by
Black Belts or sponsors, but final results require at least the concurrence of the
finance department. This should be built in from the start. The finance person
assigned to work with the team should be listed in the project charter. Without
this involvement the claimed savings are simply not credible. Aside from the
built-in bias involved in calculating the benefit created from one’s own project,
there is the issue of qualifications. The best qualified people to calculate financial
benefits are generally those who do such calculations for a living.
This is not to imply that the finance expert’s numbers should go unchallenged.
If the results appear to be unreasonable, either high or low, then they
should be clearly explained in terms the sponsor understands. The Six Sigma
Leader also has an interest in assuring that the numbers are valid. Invalid results
pose a threat to the viability of the Six Sigma effort itself. For example, on one
project the Black Belt claimed savings of several hundred thousand dollars for
‘‘unpaid overtime.’’ A finance person concurred. However, the Six Sigma
Leader would not accept the savings, arguing quite reasonably that the company
hadn’t saved anything if it had never paid the overtime. This isn’t to say that
the project didn’t have a benefit. Perhapsmorale improved or turnover declined
due to the shorter working hours. Care must be taken to show the benefits
The accounting or finance department should formally define the different
categories of savings. Savings are typically placed in categories such as:
Hard savings are actual reductions in dollars now being spent, such as
reduced budgets, fewer employees, reduction of prices paid on purchasing
contracts, etc. Hard savings can be used to lower prices, change bid models,
increase pro?ts, or for other purposes where a high degree of con?-
dence in the bene?t is required.
Soft savings are projected reductions that should result from the project.
For example, savings from less inventory, reduced testing, lower cycle
times, improved yields, lower rework rates, reduced scrap.
It is important that savings be integrated into the business systems of the
organization. If the institutional framework doesn’t change, the savings could
eventually be lost. For example, if a Six Sigma project improves a process
yield, be sure the MRP system’s calculations reflect the new yields.
Tracking Six Sigma Project Results 211
Financial analysis
Financial analysis of bene?t and cost
In performing benefit-cost analysis it is helpful to understand some of the
basic principles of financial analysis, in particular, break-even analysis and the
time value of money (TVM).
Let’s assume that there are two kinds of costs:
1. Variable costs are those costs which are expected to change at the same
rate as the ?rm’s level of sales. As more units are sold, total variable
costs will rise. Examples include sales commissions, shipping costs,
hourly wages and raw materials.
2. Fixed costs are those costs that are constant, regardless of the quantity
produced, over some meaningful range of production. Total ?xed cost
per unit will decline as the number of units increases. Examples of ?xed
costs include rent, salaries, depreciation of equipment, etc.
These concepts are illustrated in Figure 6.3.
Figure 6.3. Fixed and variable costs.
Break-even points
We can define the break-even point, or operating break-even point as the level
of unit sales required to make earnings before interest and taxes (EBIT) equal
to zero, i.e., the level of sales where profits cover both fixed and variable costs.
Let Q be the quantity sold, P the price per unit, V the variable cost per unit,
and F the total fixed costs. Then the quantity PV represents the variable profit
per unit and
Q?P  V?  F ? EBIT ?6:2?
If we set EBIT equal to zero in Equation 6.2 and solve for the break-even
quantity Q* we get:
Q ?
P  V ?6:3?
Example of break-even analysis
A publishing firm is selling books for $30 per unit. The variable costs are $10
per unit and fixed costs total $100,000. The break-even point is:
Q ?
P  V ?
$30  $10 ? 5,000 units
Of course, management usually wishes to earn a profit rather than to merely
break even. In this case, simply set EBIT to the desired profit rather than zero
in Equation 6.2 and we get the production quantity necessary to meet management’s
P  V ?6:4?
For example, if the publisher mentioned above wishes to earn a $5,000 profit
then the break-even level of sales becomes
P  V ?
$100,000 ? $5,000
$30  $10 ? 5,250 units
In project benefit-cost analysis these break-even quantities are compared to
the sales forecasts to determine the probability that the expected return will
actually be earned.
Tracking Six Sigma Project Results 213
The time value of money
Because money can be invested to grow to a larger amount, we say that
money has a ‘‘time value.’’ The concept of time value of money underlies much
of the theory of financial decision making. We will discuss two TVM concepts:
future value and present value.
Future value. Assume that you have $1,000 today and that you can invest this
sum and earn interest at the rate of 10% per year. Then, one year from today,
your $1,000 will have grown by $100 and it will be worth $1,100. The $1,100 figure
is the future value of your $1,000. The $1,000 is the present value. Let’s call
the future value FV, the present value PV and the interest rate i, where i is
expressed as a proportion rather than as a percentage. Then we can write this
example algebraically as follows:
FV ? PV ? PV  i ? PV?1 ? i?
Now, let’s say that you could invest at the 10% per year rate for two years.
Then your investment would grow as follows:
Observe that in year #2 you earned interest on your original $1,000 and on
the $100 interest you earned in year #1. The result is that you earned more
interest in year #2 than in year #1. This is known as compounding. The year
time interval is known as the compounding period. Thus, the FV after two years
is $1,210. Algebraically, here’s what happened:
FV ? ?$1,000?1:10?
?1:10? ? $1,000?1:10?2
Where the value between the [ ] characters represents the value at the end of
the first year. This approach can be used for any number of N compounding
periods. The equation is:
FV ? PV?1 ? i?N ?6:5?
Of course, Equation 6.5 can be solved for PV as well, which gives us the present
value of some future amount of money at a given rate of interest.
PV ?
?1 ? i?N ?6:6?
Non-annual compounding periods
Note that N can be stated in any time interval, it need not be in years. For
example, if the compounding period was quarterly then Nwould be the number
of quarters. Of course, the interest rate would also need to be stated in quarters.
For example, if the $1,000 were invested for two years at 10% per year, compounded
quarterly, then
FV ? PV?1 ? i?N ? $1,000 1 ?
4  24
? $1,000?1 ? 0:025?8 ? $1,218:40
Continuous compounding
Note that the FV is greater when a greater number of compounding periods
are used. The limit is an infinite number of compounding periods, known as
continuous compounding. For continuous compounding the PV and FV equations
FV ? PV  eit ?6:7?
PV ?
eit ?6:8?
Where t is the length of time (in years) the sum is compounded, e is a constant
2.71828, and all other terms are as previously defined. For our example, we
have a two-year period which gives
FV ? PV  eit ? $1,000  2:71828180:12 ? $1,221:40
Net present value
When evaluating project costs and benefits, it often happens that both costs
and benefits come in cash flow streams, rather than in lump sums.
Furthermore, the cash flow streams are uneven, i.e., the amounts vary from
Tracking Six Sigma Project Results 215
one period to the next. The approach described above can be used for uneven
cash flow streams as well. Simply compute the PV (or FV) of each cash flowseparately
and add the various results together. The result of applying this procedure
is called the net present value, or NPV. The procedure, while conceptually easy
to grasp, becomes tedious quite quickly. Fortunately, most spreadsheets have a
built in capability to perform this analysis.
Assume that a proposed project has the projected costs and benefits shown in
the table below.
1 $10,000 $0
2 $2,000 $500
3 $0 $5,000
4 $0 $10,000
5 $0 $15,000
Also assume that management wants a 12% return on their investment. What
is the NPV of this project?
There are two ways to approach this question, both of which produce the
same result (Figure 6.4). One method would be to compute the net difference
between the cost and benefit for each year of the project, then find the NPV of
this cash flow stream. The other method is to find the NPV of the cost cash
flow stream and benefit cash flow stream, then subtract.
The NPV of the cost column is $10,523; the NPV of the benefits is $18,824.
The project NPV can be found by subtracting the cost NPV from the benefit
NPV, or by finding the NPV of the yearly benefit minus the yearly cost.
Either way, the NPV analysis indicates that this project’s net present value
is $8,301.
Often in financial analysis of projects, it is necessary to determine the yield of
an investment in a project given its price and cash flows. For example, this may
be the way by which projects are prioritized. When faced with uneven cash
flows, the solution to this type of problem is usually done by computer. For
example, with Microsoft Excel, we need to make use of the internal rate of
return (IRR) function. The IRR is defined as the rate of return which equates
the present value of future cash flows with the cost of the investment. To find
the IRR the computer uses an iterative process. In other words, the computer
starts by taking an initial ‘‘guess’’ for the IRR, determines how close the computed
PV is to the cost of the investment, then adjusts its estimate of the IRR
either upward or downward. The process is continued until the desired degree
of precision has been achieved.
A quality improvement team in a hospital has been investigating the problem
of lost surgical instruments. They have determined that in the rush to get the
operating room cleaned up between surgeries many instruments are accidentally
thrown away with the surgical waste. A test has shown that a $1,500
metal detector can save the following amounts:
Tracking Six Sigma Project Results 217
Figure 6.4. Using Excel to ?nd the net present value of a project.
Year Savings
1 $750
2 $1,000
3 $1,250
4 $1,500
5 $1,750
After five years of use the metal detector will have a scrap value of $250. To
find the IRR for this cash flow stream we set up the Excel spreadsheet and
solve the problem as illustrated in Figure 6.5.
The Excel formula, shown in the window at the top of the figure, was built
using the Insert Formula ‘‘wizard,’’ with the cash flows in cells B2:B7 and an
initial guess of 0.1 (10%). Note that in year #5 the $250 salvage value is added
to the expected $1,750 in savings on surgical instruments. The cost is shown as
Figure 6.5. Using Excel to ?nd the internal rate of return for a project.
a negative cash flow in year 0. Excel found the IRR to be 63%. The IRR can be
one of the criteria for prioritizing projects, as an alternative to, or in addition
to, using the PPI.
The history of quality costs dates back to the first edition of Juran’s QC
Handbook in 1951. Today, quality cost accounting systems are part of every
modern organization’s quality improvement strategy. Indeed, quality cost
accounting and reporting are part of many quality standards. Quality cost systems
help management plan for Six Sigma by identifying opportunities for
greatest return on investment. However, leadership should keep in mind that
quality costs address only half of the quality equation. The quality equation
states that quality consists of doing the right things and not doing the wrong
things. ‘‘Doing the right things’’ means including product and service features
that satisfy or delight the customer. ‘‘Not doing the wrong things’’ means avoiding
defects and other behaviors that cause customer dissatisfaction. Quality
costs address only the latter aspect of quality. It is conceivable that a firm
could drive quality costs to zero and still go out of business.
A problem exists with the very name ‘‘cost of quality.’’ By using this terminology,
we automatically create the impression that quality is a cost. However,
our modern understanding makes it clear that quality is not a cost. Quality
represents a driver that produces higher profits through lower costs and the ability
to command a premium price in the marketplace. This author concurs with
such quality experts as H.J. Harrington and Frank M. Gryna that a better term
would be ‘‘cost of poor quality.’’ However, we will bow to tradition and use
the familiar term ‘‘cost of quality’’ throughout this discussion.
The fundamental principle of the cost of quality is that any cost that would
not have been expended if quality were perfect is a cost of quality. This includes
such obvious costs as scrap and rework, but it also includes many costs that are
far less obvious, such as the cost of reordering to replace defective material.
Service businesses also incur quality costs; for example, a hotel incurs a quality
cost when room service delivers a missing item to a guest. Specifically, quality
costs are a measure of the costs associated with the achievement or non-achievement
of product or service qualityLincluding all product or service requirements
established by the company and its contracts with customers and
society. Requirements include marketing specifications, end-product and process
specifications, purchase orders, engineering drawings, company procedures,
operating instructions, professional or industry standards, government
regulations, and any other document or customer needs that can affect the definition
of product or service. More specifically, quality costs are the total of the
Tracking Six Sigma Project Results 219
cost incurred by a) investing in the prevention of non-conformances to requirements;
b) appraising a product or service for conformance to requirements;
and c) failure to meet requirements (Figure 6.6).
tomer/user needs. Failure costs are divided into internal and external failure cost
Figure 6.6. Quality costsLgeneral description.
From Principles of Quality Costs, 3rd Edition, p. 5, Jack Campanella, Editor.
Copyright#1999 by ASQ Quality Press.
For most organizations, quality costs are hidden costs. Unless specific quality
cost identification efforts have been undertaken, few accounting systems
include provision for identifying quality costs. Because of this, unmeasured
quality costs tend to increase. Poor quality impacts companies in two ways:
higher cost and lower customer satisfaction. The lower satisfaction creates
price pressure and lost sales, which results in lower revenues. The combination
of higher cost and lower revenues eventually brings on a crisis that may threaten
the very existence of the company. Rigorous cost of quality measurement is
one technique for preventing such a crisis from occurring. Figure 6.7 illustrates
the hidden cost concept.
Goal of quality cost system
The goal of any quality cost system is to reduce quality costs to the lowest
practical level. This level is determined by the total of the costs of failure and
the cost of appraisal and prevention. Juran and Gryna (1988) present these
costs graphically as shown in Figure 6.8. In the figure it can be seen that the
cost of failure declines as conformance quality levels improve toward perfection,
while the cost of appraisal plus prevention increases. There is some ‘‘optimum’’
target quality level where the sum of prevention, appraisal, and failure
costs is at a minimum. Efforts to improve quality to better than the optimum
level will result in increasing the total quality costs.
Tracking Six Sigma Project Results 221
Figure 6.7. Hidden cost of quality and the multiplier e?ect.
From Principles of Quality Costs, 2nd Edition, p. 11, Jack Campanella, Editor.
Copyright#1990 by ASQQuality Press.
Juran acknowledged that in many cases the classical model of optimum quality
costs is flawed. It is common to find that quality levels can be economically
improved to literal perfection. For example, millions of stampings may be produced
virtually error-free from a well-designed and built stamping die. The classical
model created a mindset that resisted the idea that perfection was a
possibility. No obstacle is as difficult to surmount as a mindset. The new
model of optimum quality cost incorporates the possibility of zero defects and
is shown in Figure 6.9.
Quality costs are lowered by identifying the root causes of quality problems
and taking action to eliminate these causes. The tools and techniques described
in Part II are useful in this endeavor. KAIZEN, reengineering, and other continuous
improvement approaches are commonly used.
Strategy for reducing quality costs
As a general rule, quality costs increase as the detection point moves further
up the production and distribution chain. The lowest cost is generally obtained
when non-conformances are prevented in the first place. If non-conformances
occur, it is generally least expensive to detect them as soon as possible after
their occurrence. Beyond that point there is loss incurred from additional
work that may be lost. The most expensive quality costs are from non-confor-
Figure 6.8. Classical model of optimum quality costs.
From Juran’s Quality Control Handbook, 4th edition, J.M. Juran and F.M. Gryna, Editors.
Copyright#1988, McGraw-Hill.
mances detected by customers. In addition to the replacement or repair loss, a
company loses customer goodwill and their reputation is damaged when the
customer relates his experience to others. In extreme cases, litigation may result,
adding even more cost and loss of goodwill.
Another advantage of early detection is that it provides more meaningful
feedback to help identify root causes. The time lag between production and
field failure makes it very difficult to trace the occurrence back to the process
state that produced it. While field failure tracking is useful in prospectively evaluating
a ‘‘fix,’’ it is usually of little value in retrospectively evaluating a problem.
Accounting support
We have said it before, but it bears repeating, that the support of the accounting
department is vital whenever financial and accounting matters are involved.
In fact, the accounting department bears primary responsibility for accounting
matters, including cost of quality systems. The Six Sigma department’s role in
Tracking Six Sigma Project Results 223
Figure 6.9. New model of optimum quality costs.
From Juran’s Quality Control Handbook, 4th edition, J.M. Juran and F.M. Gryna, Editors.
Copyright#1988, McGraw-Hill.
development and maintenance of the cost of quality system is to provide guidance
and support to the accounting department.
The cost of quality system must be integrated into the larger cost accounting
system. It is, in fact, merely a subsystem. Terminology, format, etc., should be
consistent between the cost of quality system and the larger system. This will
speed the learning process and reduce confusion. Ideally, the cost of quality
will be so fully integrated into the cost accounting system that it will not be
viewed as a separate accounting system at all, it will be a routine part of cost
reporting and reduction. The ideal cost of quality accounting system will simply
aggregate quality costs to enhance their visibility to management and facilitate
efforts to reduce them. For most companies, this task falls under the jurisdiction
of the controller’s office.
Quality cost measurement need not be accurate to the penny to be effective.
The purpose of measuring such costs is to provide broad guidelines for management
decision-making and action. The very nature of cost of quality makes
such accuracy impossible. In some instances it will only be possible to obtain
periodic rough estimates of such costs as lost customer goodwill, cost of damage
to the company’s reputation, etc. These estimates can be obtained using special
audits, statistical sampling, and other market studies. These activities can be
jointly conducted by teams of marketing, accounting, and Six Sigma personnel.
Since these costs are often huge, these estimates must be obtained. However,
they need not be obtained every month. Annual studies are usually sufficient
to indicate trends in these measures.
Management of quality costs
In our discussion of the cost of quality subsystem, we emphasized the importance
of not creating a unique accounting system. The same holds true when discussing
management of quality costs. Quality cost management should be part
of the charter of the senior level cross-functional cost management team. It is
one part of the broader business effort to control costs. However, in all likelihood,
the business will find that quality cost reduction has greater potential to
contribute to the bottom line than the reduction of other costs. This is so
because, unlike other costs, quality costs are waste costs (Pyzdek, 1976). As
such, quality costs contribute no value to the product or service purchased by
the customer. Indeed, quality costs are often indicators of negative customer
value. The customer who brings his car in for a covered warranty expense suffers
uncompensated inconvenience, the cost of which is not captured by most
quality cost systems (although, as discussed above, we recommend that such
costs be estimated from time to time). All other costs incurred by the firm purchase
at least some value.
Effective cost of quality programs consist of taking the following steps
(Campanella, 1990, p. 34):
. Establish a quality cost measurement system
. Develop a suitable long-range trend analysis
. Establish annual improvement goals for total quality costs
. Develop short-range trend analyses with individual targets which, when
combined, meet the annual improvement goal
. Monitor progress towards the goals and take action when progress falls
short of targets
The tools and techniques described in Chapter 15 are useful for managing Six
Sigma quality cost reduction projects.
Quality cost management helps firms establish priorities for corrective
action. Without such guidance, it is likely that firms will misallocate their
resources, thereby getting less than optimal return on investment. If such
experiences are repeated frequently, the organization may even question
or abandon their quality cost reduction efforts. The most often-used tool
in setting priorities is Pareto analysis (see Chapter 8). Typically at the outset
of the quality cost reduction effort, Pareto analysis is used to evaluate
failure costs to identify those ‘‘vital few’’ areas in most need of attention.
Documented failure costs, especially external failure costs, almost certainly
understate the true cost and they are highly visible to the customer.
Pareto analysis is combined with other quality tools, such as control charts
and cause and effect diagrams, to identify the root causes of quality problems.
Of course, the analyst must constantly keep in mind the fact that
most costs are hidden. Pareto analysis cannot be effectively performed
until the hidden costs have been identified. Analyzing only those data
easiest to obtain is an example of the GIGO (garbage-in, garbage-out)
approach to analysis.
After the most significant failure costs have been identified and brought
under control, appraisal costs are analyzed. Are we spending too much on
appraisal in view of the lower levels of failure costs? Here quality cost analysis
must be supplemented with risk analysis to assure that failure and appraisal
cost levels are in balance. Appraisal cost analysis is also used to justify expenditure
in prevention costs.
Prevention costs of quality are investments in the discovery, incorporation,
and maintenance of defect prevention disciplines for all operations affecting
the quality of product or service (Campanella, 1990). As such, prevention
needs to be applied correctly and not evenly across the board. Much improvement
has been demonstrated through reallocation of prevention effort from
areas having little effect to areas where it really pays off; once again, the Pareto
principle in action.
Tracking Six Sigma Project Results 225
Cost of quality examples
I. Prevention costsLCosts incurred to prevent the occurrence of nonconformances
in the future, such as*
A. Marketing/customer/user
1. Marketing research
2. Customer/user perception surveys/clinics
3. Contract/document review
B. Product/service/design development
1. Design quality progress reviews
2. Design support activities
3. Product design qualification test
4. Service design qualification
5. Field tests
C. Purchasing
1. Supplier reviews
2. Supplier rating
3. Purchase order tech data reviews
4. Supplier quality planning
D. Operations (manufacturing or service)
1. Operations process validation
2. Operations quality planning
a. Design and development of quality measurement and control
3. Operations support quality planning
4. Operator quality education
5. Operator SPC/process control
E. Quality administration
1. Administrative salaries
2. Administrative expenses
3. Quality program planning
4. Quality performance reporting
5. Quality education
6. Quality improvement
7. Quality audits
8. Other prevention costs
*All detailed quality cost descriptions are from Principles of Quality Costs, John T. Hagan, editor. Milwaukee, WI: ASQ
Quality Press, appendix B.
II. Appraisal costsLCosts incurred in measuring and controlling current
production to assure conformance to requirements, such as
A. Purchasing appraisal costs
1. Receiving or incoming inspections and tests
2. Measurement equipment
3. Qualification of supplier product
4. Source inspection and control programs
B. Operations (manufacturing or service) appraisal costs
1. Planned operations inspections, tests, audits
a. Checking labor
b. Product or service quality audits
c. Inspection and test materials
2. Set-up inspections and tests
3. Special tests (manufacturing)
4. Process control measurements
5. Laboratory support
6. Measurement equipment
a. Depreciation allowances
b. Measurement equipment expenses
c. Maintenance and calibration labor
7. Outside endorsements and certifications
C. External appraisal costs
1. Field performance evaluation
2. Special product evaluations
3. Evaluation of field stock and spare parts
D. Review of tests and inspection data
E. Miscellaneous quality evaluations
III. Internal failure costsLCosts generated before a product is shipped as a
result of non-conformance to requirements, such as
A. Product/service design failure costs (internal)
1. Design corrective action
2. Rework due to design changes
3. Scrap due to design changes
B. Purchasing failure costs
1. Purchased material reject disposition costs
2. Purchased material replacement costs
3. Supplier corrective action
4. Rework of supplier rejects
5. Uncontrolled material losses
Tracking Six Sigma Project Results 227
C. Operations (product or service) failure costs
1. Material review and corrective action costs
a. Disposition costs
b. Troubleshooting or failure analysis costs (operations)
c. Investigation support costs
d. Operations corrective action
2. Operations rework and repair costs
a. Rework
b. Repair
3. Reinspection/retest costs
4. Extra operations
5. Scrap costs (operations)
6. Downgraded end product or service
7. Internal failure labor losses
D. Other internal failure costs
IV. External failure costsLCosts generated after a product is shipped as a
result of non-conformance to requirements, such as
A. Complaint investigation/customer or user service
B. Returned goods
C. Retrofit costs
D. Recall costs
E. Warranty claims
F. Liability costs
G. Penalties
H. Customer/user goodwill
I. Lost sales
J. Other external failure costs
Quality cost bases
The guidelines for selecting a base for analyzing quality costs are:
. The base should be related to quality costs in a meaningful way
. The base should be well-known to the managers who will receive the quality
cost reports
. The base should be a measure of business volume in the area where quality
cost measurements are to be applied
. Several bases are often necessary to get a complete picture of the relative
magnitude of quality costs
Some commonly used bases are (Campanella, 1990, p. 26):
. A labor base (such as total labor, direct labor, or applied labor)
. A cost base (such as shop cost, operating cost, or total material and labor)
. A sales base (such as net sales billed, or sales value of ?nished goods)
. A unit base (such as the number of units produced, or the volume of output)
While actual dollars spent are usually the best indicator for determining
where quality improvement projects will have the greatest impact on profits
and where corrective action should be taken, unless the production rate is relatively
constant, it will not provide a clear indication of quality cost improvement
trends. Since the goal of the cost of quality program is improvement over
time, it is necessary to adjust the data for other time-related changes such as production
rate, inflation, etc. Total quality cost compared to an applicable base
results in an index which may be plotted and analyzed using control charts,
run charts, or one of the other tools described in Chapters 11^14.
For long-range analyses and planning, net sales is the base most often used
for presentations to top management (Campanella, 1990, p. 24). If sales are relatively
constant over time, the quality cost analysis can be performed for relatively
short spans of time. In other industries this figure must be computed
over a longer time interval to smooth out large swings in the sales base. For
example, in industries such as shipbuilding or satellite manufacturing, some
periods may have no deliveries, while others have large dollar amounts. It is
important that the quality costs incurred be related to the sales for the same period.
Consider the sales as the ‘‘opportunity’’ for the quality costs to happen.
Some examples of cost of quality bases are (Campanella, 1990):
. Internal failure costs as a percent of total production costs
. External failure costs as an average percent of net sales
. Procurement appraisal costs as a percent of total purchased material cost
. Operations appraisal costs as a percent of total production costs
. Total quality costs as a percent of production costs
An example of a cost of quality report that employs some of these bases is
shown in Figure 6.10.
Quality cost trend analysis
As stated above, the purpose of collecting quality cost data is to provide a
sound basis for taking the necessary action to eliminate the causes of these
costs, and thereby eliminate the costs themselves. If the action taken is effective,
the data will indicate a positive trend. Trend analysis is most often performed
by presenting the data in run chart form and analyzing the runs (see Chapter
11). It is common to combine all of the cost of quality data on a single graph, as
shown in Figure 6.11.
Tracking Six Sigma Project Results 229
Figure 6.10. Quality costs summary report.
From Principles of Quality Costs, 2nd Edition, p. 48. Jack Campanella, Editor.
Copyright#1990 by ASQ Quality Press.
If the runs are subjected to the run tests described below, it can be shown that
the total failure and total COQ (cost of quality) trends are statistically significant.
However, for this example data, the use of formal statistical rules is
superfluousLthe improvement is obvious.
While such aggregate analysis is useful for senior management, it is of little
value to those engaged in more focused Six Sigma cost of quality projects. In
these cases the trend data should be as specific as possible to the area being studied.
Also, the measurement may be something more directly related to the
work being done by the Six Sigma team rather than dollars, and the time interval
should be shorter. For example, if it has been determined that a major internal
failure cost item is defective solder joints, then the team should plot a control
chart of the solder defect rate and analyze the process in real-time. Obviously,
reducing solder defects should reduce the cost associated with solder defects.
Implementing the quality cost program
Quality cost program introduction is a major project and should utilize the
tools and techniques described in Chapter 15. Prior to implementation, a
needs analysis should be performed to determine if, in fact, a cost of quality program
can benefit the company. The needs assessment should also include a
Tracking Six Sigma Project Results 231
Figure 6.11. Cost of quality history.
benefit/cost analysis and a plan for the implementation. The plan should
. the management presentation, designed to identify the overall opportunity
and show an example of how the program will achieve its bene?ts
. a description of the pilot program
. material designed to educate and involve all functions in the program
. outline of the internal cost of quality accounting procedures
. description of the data collection and analysis of cost of quality data at the
highest level of aggregation
. description of the cost of quality reporting system and how the data will be
used to improve quality
As with any major Six Sigma project, a sponsor should be found and management
support secured. In the case of cost of quality, the sponsor should be the
controller or one of her subordinates.
Use of quality costs
The principal use of quality cost data is to justify and support quality performance
improvement. Quality cost data help identify problem areas and
direct resources to these areas. To be effective, the cost of quality system has to
be integrated with other quality information systems to assure that root causes
will be addressed. Statistical analysis can be used to correlate quality cost trends
with other quality data to help direct attention to problem causes.
One mission of the quality management function is to educate top management
about the long-range effects of total quality performance on the profits
and quality reputation of the company. Management must understand that strategic
planning for quality is as important as strategic planning for any other
functional area. When the strategic plan addresses cost issues, quality cost consideration
should be prominent. Quality costs should be considered first
because, since they are waste costs, their reduction is always taken from the
‘‘fat’’ of the organization. The role of the quality manager in this process should
be to (Campanella, 1990, p. 56)
. analyze major trends in customer satisfaction, defects or error rates, and
quality costs, both generally and by speci?c program or project. These
trends should also be used to provide inputs for setting objectives;
. assist the other functions to ensure that costs related to quality are
included in their analyses for setting objectives;
. develop an overall quality strategic plan which incorporates all functional
quality objectives and strategic action plans, including plans and budgets
for the quality function.
Bene?ts of quality cost reduction
Quality cost reductions can have a significant impact on a company’s growth
rate and bottom line. Research done by the Chicago Graduate School of
Business showed that companies using TQM for an average of 6.5 years
increased revenues at an annual rate of 8.3% annually, versus 4.2% annually for
all U.S. manufacturers. Suminski (1994) reports that the average manufacturer’s
price of non-conformance is 25% of operating costs, for service businesses the
figure is 35%. These costs represent a direct charge against a company’s profitability.
A New England heavy equipment manufacturer reports that their
price of non-conformance was 31% of total sales when they undertook a quality
cost reduction project. In just one year they were able to lower these costs to
9%. Among their accomplishments:
. Scrap and rework reduced 30%.
. Manufacturing cost variance reduced 20%.
. Late collections reduced 46%.
. Average turnaround on receivables reduced from 62 days to 35 days.
Lessons learned capture and replication
It is often possible to apply the lessons learned from a project to other processes,
either internally or externally.Most companies have more than one person
or organizational unit performing similar or identical tasks. Many also
have suppliers and outsourcers who do work similar to that being done internally.
By replicating the changes done during a project the benefits of Six Sigma
can be multiplied many fold, often at very minimal cost. Think of it as a form
of benchmarking. Instead of looking for the best-in-class process for you to
learn from, the Six Sigma team created a best-in-class process and you want to
teach the new approach to others.
Unlike benchmarking, where the seeker of knowledge is predisposed to
change what they are doing, the process owners who might benefit from the
knowledge gained during a Six Sigma project may not even be aware that they
can benefit from a change. This needs to be accounted for when planning the
program for sharing lessons learned. The process is a combination of motivation,
education and selling the target audience on the new approach. Chances
are that those who worked the project are not the best ones to sell others on
the new approach. They can serve as technical advisers to those who will carry
the message to other areas. The Six Sigma function (Process Excellence) usually
takes the lead in developing a system for replication and sharing of lessons
Tracking Six Sigma Project Results 233
In addition to the lessons learned about business processes, a great deal will
be learned about how to conduct successful projects. In a few years even a moderately
sized Six Sigma effort will complete hundreds or thousands of projects.
These project lessons learned should be captured and used to help other project
teams. The information is usually best expressed in simple narratives by the project
Black Belt. The narratives can be indexed by search engines and used by
other Black Belts in the organization. The lessons learned database is an extremely
valuable asset to the Six Sigma organization.
^ ^ ^
Six Sigma Tools and
Copyright 2003 by The McGraw-Hill Companies, Inc. Click Here for Terms of Use.
This page intentionally left blank.
^ ^ ^
Introduction to DMAIC
and Other Improvement
Part II covers the toolkit commonly used in Six Sigma.
For the most part, these are the same tools used by the quality profession and
applied statisticians for decades. Six Sigma puts some new twists on these traditional
1. They are taught in the context of a well-de?ned improvement model
known as DMAIC (see below). Computers are used intensively.
2. They are applied at once on real projects designed to deliver tangible
results for an identi?ed stakeholder.
3. Items 1 and 2 are integrated via an intensive training regimen that is provided
to full-time change agents who work on projects while they are
being trained.
The tools of Six Sigma are most often applied within a simple performance
improvement model known as De?ne-Measure-Analyze-Improve-Control,
or DMAIC. DMAIC is summarized in Figure 7.1. DMAIC is used when a
project’s goal can be accomplished by improving an existing product, process,
or service.
Copyright 2003 by The McGraw-Hill Companies, Inc. Click Here for Terms of Use.
Figure 7.1. Overview of DMAIC.
D De?ne the goals of the improvement activity. The most important
goals are obtained from customers. At the top level the goals will be
the strategic objectives of the organization, such as greater customer
loyalty, a higher ROI or increased market share, or greater employee
satisfaction. At the operations level, a goal might be to increase the
throughput of a production department. At the project level goals
might be to reduce the defect level and increase throughput for a
particular process. Obtain goals from direct communication with
customers, shareholders, and employees.
M Measure the existing system. Establish valid and reliable metrics to
help monitor progress towards the goal(s) de?ned at the previous
A Analyze the system to identify ways to eliminate the gap between the
current performance of the system or process and the desired goal.
Begin by determining the current baseline. Use exploratory and
descriptive data analysis to help you understand the data. Use
statistical tools to guide the analysis.
I Improve the system. Be creative in ?nding new ways to do things
better, cheaper, or faster. Use project management and other
planning and management tools to implement the new approach. Use
statistical methods to validate the improvement.
C Control the new system. Institutionalize the improved system by
modifying compensation and incentive systems, policies, procedures,
MRP, budgets, operating instructions and other management
systems. You may wish to utilize standardization such as ISO 9000 to
assure that documentation is correct. Use statistical tools to monitor
stability of the new systems.
DMAIC is such an integral part of Six Sigma that it is used to organize the
material for Part II of this book. It provides a useful framework for conducting
Six Sigma projects, see Figure 7.2. DMAIC is sometimes even used to create a
‘‘gated process’’ for project control. That is, criteria for completing a particular
phase are de?ned and projects reviewed to determine if all of the criteria have
been met. If so, then the gate (e.g., De?ne) is ‘‘closed.’’
Table 7.1 shows a partial listing of tools often found to be useful in a given
stage of a project. There is considerable overlap in practice.
Design for Six Sigma project framework
Another approach, used when the goal is the development of a new or radically
redesigned product, process or service, is De?ne-Measure-Analyze-
Design-Verify, or DMADV (Figure 7.3). DMADV is part of the design for Six
Sigma (DFSS) toolkit.
DMAIC, DMADV, and learning models 239
Figure 7.2. Using DMAIC on a Six Sigma project.
Table 7.1. Six Sigma tools commonly used in each phase of a project.
Project Phase Candidate Six Sigma Tools
De?ne & Project charter
& VOC tools (surveys, focus groups, letters, comment
& Process map
& Benchmarking
Measure & Measurement systems analysis
& Exploratory data analysis
& Descriptive statistics
& Data mining
& Run charts
& Pareto analysis
Analyze & Cause-and-e?ect diagrams
& Tree diagrams
& Brainstorming
& Process behavior charts (SPC)
& Process maps
& Design of experiments
& Enumerative statistics (hypothesis tests)
& Inferential statistics (Xs and Ys)
& Simulation
Improve & Force ?eld diagrams
& 7M tools
& Project planning and management tools
& Prototype and pilot studies
Control & SPC
& ISO 900 & Change budgets, bid models, cost estimating models
& Reporting system
De?ne the goals of the design activity. What is being designed?
Why? Use QFD or the Analytic Hierarchical Process to assure that
the goals are consistent with customer demands and enterprise
Measure. Determine Critical to Stakeholder metrics. Translate
customer requirements into project goals.
Analyze the options available for meeting the goals. Determine the
performance of similar best-in-class designs.
Design the new product, service or process. Use predictive models,
simulation, prototypes, pilot runs, etc. to validate the design
concept’s e?ectiveness in meeting goals.
V Verify the design’s e?ectiveness in the real world.
Figure 7.3. Overview of DMADV.
Figure 7.4 illustrates the relationship between DMAIC and DMADV.
Learning models
Knowledge is hierarchical, meaning that some ideas have more impact
than others because they are more fundamental. Six Sigma tends to take a
very ‘‘practical view’’ of the world, but this perspective is dangerous if its
context isn’t well understood. True, the focus is on doing. But how do we
know that what we are doing is correct? If we are wrong, then our actions
may make matters worse instead of better. The question of how we know
that we know is a philosophical one, not a technical one. Technical tools,
such as statistical methods, can be used to help us answer this question,
but unless we have a deep understanding of the philosophy that underlies
the use of the tools we won’t really know how to interpret the results
Learning is the acquisition of new knowledge about the way the world
works. Both DMAIC and DMADV are learning frameworks. Learning
must occur if the project deliverable is to provide the intended bene?t.
W|thout learning to guide process change activity it’s just hit-and-miss, and
Murphy’s Law* assures that our e?orts will miss the mark more often than
DMAIC, DMADV, and learning models 241
not. There is a long and proud history of learning models that re?ect the
thinking of some of the greatest minds in twentieth century business, such as
Drs. Shewhart, Deming and Juran. There are new learning models that incorporate
recent discoveries in the ?elds of chaos theory and complexity theory.
The new models, Select-Experiment-Learn (SEL) and Select-Experiment-
Adapt (SEA), apply to systems in dynamic, far from equilibrium environments
where the traditional models break down.
Figure 7.4. DMAIC and DMADV.
*Anything that can go wrong, will.
The PDCA cycle, which Deming refers to as the PDSA cycle (Deming, 1993,
p. 134), is a ?ow chart for learning and process improvement. The basic idea
began with Shewhart’s attempt to understand the nature of knowledge.
Shewhart believed that knowledge begins and ends in experimental data but
that it does not end in the data in which it begins. He felt there were three important
components of knowledge (Shewhart, 1939, 1986): a) the data of experience
in which the process of knowing begins, b) the prediction in terms of data that
one would expect to get if one were to perform certain experiments in the future,
and c) the degree of belief in the prediction based on the original data or some
summary thereof as evidence. Shewhart arranged these three components schematically
as shown in Figure 7.5.
Since knowledge begins with the original data and ends in new data, these
future data constitute the operationally veri?able meaning of the original data.
However, since inferences or predictions based upon experimental data can
never be certain, the knowledge based upon the original data can inhere in
these data only to the extent of some degree of rational belief. In other words,
according to Shewhart, knowledge can only be probable. Also, the data are not
‘‘facts’’ in and of themselves, they are merely measurements that allow us to
draw inferences about something. In other words, we can not have facts without
some theory.
Shewhart applied these principles in many practical ways. For example, he
identi?ed the three steps of quality control in manufacturing as speci?cation,
production, and judgment of quality (inspection). He noted that, in practice,
speci?cations could not be set without ?rst having some information from
DMAIC, DMADV, and learning models 243
Figure 7.5. The three components of knowledge.
inspection to help establish process capability, and that this information could
not be obtained until some units had been produced. In short, Shewhart modi-
?ed the sequence of speci?cation-production-inspection as shown in Figure
7.6. He also observed that the speci?cation-production-inspection sequence corresponded
respectively to making a hypothesis, carrying out an experiment,
and testing the hypothesis. Together the three steps constitute a dynamic scienti
?c process of acquiring knowledge.
Note that Shewhart’s model of knowledge forms a circle. Shewhart
followed the teachings of philosopher C.I. Lewis, who believed that all good
logics are circular. The essence of this view is to see knowledge as dynamic.
It changes as new evidence comes in. As Shewhart put it (Shewhart, 1939,
1986, p. 104):
Knowing in this sense is somewhat a continuing process, or method, and
di?ers fundamentally in this respect from what it would be if it were possible
to attain certainty in the making of predictions.
Shewhart and Deming revised the above model for application to the
improvement of products and processes. The new model was ?rst called the
PDCA cycle, later revised by Deming to the Plan-Do-Study-Act, or PDSA
cycle (Deming, 1993, p. 134). The Shewhart-Deming PDSA cycle is shown in
Figure 7.7.
Figure 7.6. Scienti?c process of acquiring knowledge.
Plan a change or a test, aimed at improvement. This is the foundation for the
entire PDCA-PDSA cycle. The term ‘‘plan’’ need not be limited to large-scale
planning on an organization-wide scale, it may simply refer to a small process
change one is interested in exploring.
Do. Carry out the change or the test (preferably on a small scale). It is important
that the DO step carefully follow the plan, otherwise learning will not be
Study the results. What did we learn? What went wrong?
Act. Adopt the change, or abandon it, or run through the cycle again.
The PDCAapproach is essentially amanagement-oriented version of the original
Shewhart cycle, which focused on engineering and production. A number
of other variations have been developed, two of Deming’s variations are shown
in Figure 7.8.
Juran depicts quality as a ‘‘spiral,’’ as shown in Figure 7.9.
Because of their historical origins and logical appeal, circular diagrams are
ubiquitous in the quality ?eld. In quality management, the circle represents continuous
improvement of quality by continuous acquisition of knowledge.
The PDSA cycle describes planning and learning in an environment at or
near a stable equilibrium. The PDSA loop indicates that plans are con-
DMAIC, DMADV, and learning models 245
Figure 7.7. The Shewhart-Deming PDSA cycle for learning and improvement.
tinuously improved by studying the results obtained when the plans are
implemented, and then modifying the plans. However, the PDSA model fails
to account for the activities of other agents, which is a characteristic of complex
adaptive systems, such as a market economy. For this situation I propose
a new model, the Select-Experiment-Adapt (SEA) model depicted in Figure
Figure 7.8. Some variations of the PDCA-PDSA cycle.
In real life, experimentation goes on constantly. Experimenting involves
executing a performance rule activated by a message received from the environment.
We observe something, or induce something based on thinking
about past observations, and decide which course of action would be most
bene?cial. The action taken in response to the environmental messages is
called a performance rule. Adaptation occurs by adjusting the strength of the
performance rule based on the payo? we actually received from it. Repeated
iterations of the SEA cycle mimics what computer scientist John Holland
calls the bucket brigade algorithm (Holland, 1996) which strengthens rules
that belong to chains of action terminating in rewards. The process amounts
to a progressive con?rmation of hypotheses concerned with stage setting and
DMAIC, DMADV, and learning models 247
Figure 7.9. Juran’s spiral of progress in quality.
Figure 7.10. The Select-Experiment-Adapt (SEA) model for non-linear systems.
SEA versus PDSA
In the PDSA cycle, the plan documents the theory being tested. Deming
believed that a statement which conveys knowledge must predict future outcomes,
with risk of being wrong, and that it ?ts without failure observations of
the past. Rational prediction, according to Deming, requires theory and builds
knowledge through systematic revision and extension of theory based on comparison
of prediction with observation. W|thout theory there is nothing to
revise, so experience has no meaning (Deming, 1993).
The SEA model, unlike the PDSA cycle, contains positive feedback loops,
making this a dynamic, non-linear system. These systems act like both common
and special causes in the Shewhart-Deming model. V|rtually undetectable
minor di?erences (common causes) are greatly ampli?ed by positive feedback
and produce unpredictably large e?ects (special causes). Because of positive
feedback the behavior of even simple systems like the one shown in Figure 7.10
is unpredictable, even in principle. Of course, this illustration grossly oversimpli
?es reality. In the real world there are many competitors, competitors for
our customers, many customers, regulation, many employees changing things
in our ?rm, and so on. But the conclusion is the same: long-term forecasting is
impossible, and therefore long-term planning is invalid. The ‘‘P’’ (plan) in the
PDSA cycle cannot be used for other than short-term planning or systems in a
state of ‘‘control’’ in the Shewhart-Deming sense.
The ‘‘S’’ (study) element is also suspect. What exactly are we studying? The
e?ect of the action we took in the ‘‘A’’ (act) step? This won’t work because the
observed e?ects are also in?uenced, even overwhelmed, by actions taken by
other agents. Thus, we may falsely conclude that our actions had an e?ect when
in fact they did not, leading to superstitious learning. For example, we run a special
promotion and sales increase, leading us to conclude that the promotion
was a success. But in fact our promotion just happened to coincide with a customer
promotion that created a temporary increase in demand for our product.
Or we may conclude that our actions did not have an e?ect when in fact their
e?ect was masked by activities by other agents. For example, perhaps our new
marketing program would have worked except that our competitor had a
short-term sale and our customer was under pressure to hold costs down due
to a temporary cash ?ow problem.
Learning and the SEA model
In the SEA model, there is no ‘‘learning’’ per se. There is merely strategic
adaptation. Computers can be programmed to modify performance rules
based on payo?s, but the computer doesn’t learn anything. It merely ‘‘discovers’’
new performance rules through successful adaptations based on repeated
DMAIC, DMADV, and learning models 249
trial and error, i.e., through iterations of the SEA loop. Learning in the human
sense involves discovering principles that explain the reasons for the increased
or decreased payo?s obtained by applying the performance rules. This is a different
thing entirely than simply discovering that a particular performance
rule gives a somewhat higher payo?. Learning makes it possible to skip one or
more generations of adaptation.
One model that incorporates learning in a dynamic environment, the Select-
Experiment-Learn (SEL) model, is shown in Figure 7.11.
SELalso di?ers from PDSA. Shewhart realized the value of discovering scienti
?c principles, and he also understood that progress was possible even without
this knowledge. However, Shewhart believed that it was possible to apply
natural laws to achieve ‘‘control within limits,’’ i.e., statistical certainty. What
Shewhart called a state of statistical control, I will call statistical equilibrium.
A system exhibits statistical equilibrium when its future performance is predictable
within limits which can be determined using linear, negative feedback models.
Chaos theory and complexity theory show that in dynamic environments,
i.e., environments in?uenced by positive feedback, even this level of control is
impossible to achieve.
TheSELmodel is designed for a dynamicenvironmentandit does not attempt
to develop long-range strategic plans based on super-human knowledge and foresight.
Instead SELseeks principles that are useful for making predictions, recognizing
that positive feedback and the actions of other agents makes it di?cult to
identify the e?ects of these principles. Furthermore, other agents may also
acquire thisknowledgeandmodifytheirbehavior,thereby negating the principle.
For example, cooperation and reciprocity may appear to be a principle that
applies to all human cultures. However, since the principle applies to agents,
future behavior can not be predicted with even statistical certainty. If others realize
that you are applying this principle, they can take advantage of your predictability.
Of course, until new breakthrough principles are learned, gradual
continuous improvement can still be obtained by using the SEA model. The
cumulative improvement fromSEAcan be signi?cant (e.g., natural evolution).
Figure 7.11. The Select-Experiment-Learn (SEL) model for dynamic systems.
Essentially, when environments are dynamic the SEA and SELmodels
replace the equilibrium environment PDSA learning model with dynamic adaptation
(SEA) and agent-based learning (SEL). Centralized control schemes
(plans) are replaced by self-control or at least local control by meta-agents. Six
Sigma activities should employ all three strategies for improvement. Here are
some general guidelines to help you determine when to apply a given approach:
. SEA applies unless formal, controlled experiments are underway. Follow
a mini-max strategy: minimize central planning and control to the maximum
extent possible. Allow individual employees maximum freedom to
experiment and change their work environment and processes to seek better
ways to do things.
. When processes are in?uenced by positive feedback from other agents,
apply the SEA and SELmodels. Eliminate long-term strategic planning
and strive to cultivate an environment with maximum ability to adapt to
. When processes are at or near equilibrium and not in?uenced by positive
feedback loops, PDSA applies. Since PDSA is planning based, the use of
formal teams is justi?ed. Rigorously apply the tools of process control,
formal design of experiments, etc.
Illustration of PDSA, SEA and SEL
The chart below shows the percentage change in the S and P 500 index of
stocks over a period of 100 months. The data re?ect the buying and selling activities
of millions of investors. The data re?ect statistical control, i.e., equilibrium
behavior, for the entire period and PDSA functioned quite well for investors
during the period. Using control charts the predicted return for the next
month is between 11.2% and +12.6%.
But this process turned out to be distinctly non-linear. In the month #101,
investors (agents) reacted to a price drop by selling, which caused the price to
drop further, which caused still more selling. In other words, this dynamic process
encountered a positive feedback loop. The result of this SEA behavior: a
drop of nearly 22% (indicated by an ‘‘X’’ on the chart); a far greater drop than
predicted by linear statistical models.
Some investors were not in?uenced by the positive feedback. Using SEL
logic, they examined macro and micro factors and found no reason for the
plunge. Rather than selling, they either held on to their shares or bought more.
For a while, it appeared that this strategy would back?re: the market dropped
another 9% the next month. But it eventually recovered, regained all of the lost
ground, and moved back into positive territory.
DMAIC, DMADV, and learning models 251
^ ^ ^
Problem Solving Tools
Just as companies have organization charts, they can have process maps that
give a picture of how work flows through the company. A process map creates
a vocabulary to help people discuss process improvement. A process map is a
graphic representation of a process, showing the sequence of tasks using a modified
version of standard flowcharting symbols. The map of a work process is a
picture of how people do their work. Work process maps are similar to road
maps in that there are many alternative routes that will accomplish the objective.
In any given circumstance, one route may be better than others. By creating
a process map, the various alternatives are displayed and effective planning is
facilitated. The steps involved are as follows (Galloway, 1994):
1. Select a process to be mapped.
2. De?ne the process.
3. Map the primary process.
4. Map alternative paths.
5. Map inspection points.
6. Use themap to improve the process.
Processes correspond to natural business activities. However, in modern
organizations these natural processes are fragmented among many different
departments. A process map provides an integrated picture of the natural process.
Because of the focus on organizational hierarchies, processes tend to be
unmanaged. People are responsible for departments and budgets, but no one is
responsible for the processes.
Copyright 2003 by The McGraw-Hill Companies, Inc. Click Here for Terms of Use.
Because organizations are arranged as departments rather than processes, it
is often difficult for people to see the processes that make up the business. To
get a better handle on the processes that make up a business, Hammer and
Champy (1993, p. 118) suggest that they be given names that express their beginning
and ending states. These names should imply all the work that gets done
between their start and finish. Manufacturing, which sounds like a department
name, is better called the procurement-to-shipment process. Some other recurring
processes and their state-change names:
. Product development: concept to prototype
. Sales: prospect to order
. Order ful?llment: order to payment
. Service: inquiry to resolution
Cycle time reduction through cross-functional
process mapping
Hurley and Loew (1996) describe how Motorola uses process mapping to
help them reduce cycle times. Cross-functional process mapping involves creating
teams whose members are selected from every department involved in
the new product development cycleLfrom marketing to manufacturing to
research and development. The next phase involves mapping each step within
the product development process from start to finish. Team members are
divided into four categories:
. Project championLprovide resources and remove barriers
. Team leaderLorganize and conduct meetings, insure that information
exchange occurs
. Action item ownerLcomplete assigned tasks
. Team memberLcomplete assigned tasks
The teams develop two maps: an ‘‘as-is’’ map and a ‘‘should-be’’ map. The
As-is may detail the way the new product-development process is currently
run and identifies all the problematic issues that exist in the current way that
new product development is accomplished. Using the cross-functional format,
each step of the process is mapped out, along with the time each step takes.
The result of the exercise is twofold: a map that shows the current process, and
an appreciation among team members of the contributions of their fellow
team members. The As-is map can be used to improve the current process
(KAIZEN). If possible, any steps that do not add value in the customer’s eyes,
or that are redundant, should be deleted.
The Should-be map forms the basis of reengineering the product development
process. The Should-be map details each step in the new, more efficient
Process mapping 253
process. A list of action items is created during this mapping session. Action
items define and detail what needs to be changed in order to move from the
As-is state to the Should-be state. The project management tools and techniques
described in Chapter 15 are then used to plan and implement the necessary
steps. A Should-be process map is shown in Figure 8.1.
A process flow chart is simply a tool that graphically shows the inputs,
actions, and outputs of a given system. These terms are defined as follows:
InputsLthe factors of production: land, materials, labor, equipment, and
ActionsLthe way in which the inputs are combined and manipulated in
order to add value. Actions include procedures, handling, storage,
transportation, and processing.
OutputsLthe products or services created by acting on the inputs. Outputs
are delivered to the customer or other user. Outputs also include
unplanned and undesirable results, such as scrap, rework, pollution,
etc. Flow charts should contain these outputs as well.
Figure 8.1. New product/service development ‘‘should-be’’ process map.
From Hurley, H. and Loew, C. (1996), ‘‘A quality change for new product development,’’
The Quality Observer, January, pp. 10^13.
Flow charting is such a useful activity that the symbols have been standardized
by various ANSI standards. There are special symbols for special processes,
such as electronics or information systems. However, in most cases one
can get by with the symbols shown in Figure 8.2.
The flow chart in Figure 8.3 shows a high-level view of a process capability
analysis. The flow chart can be made either more complex or less complex. As
a rule of thumb, to paraphrase Albert Einstein, ‘‘Flow charts should be as simple
as possible, but not simpler.’’ The purpose of the flow chart is to help people
understand the process and this is not accomplished with flow charts that are
either too simple or too complex.
Check sheets are devices which consist of lists of items and some indicator of
how often each item on the list occurs. In their simplest form, checklists are
tools that make the data collection process easier by providing pre-written
descriptions of events likely to occur. A well-designed check sheet will answer
the questions posed by the investigator. Some examples of questions are: ‘‘Has
everything been done?’’ ‘‘Have all inspections been performed?’’ ‘‘How often
does a particular problem occur?’’ ‘‘Are problems more common with part X
than with part Y?’’ They also serve as reminders that direct the attention of the
data collector to items of interest and importance. Such simple check sheets
are called confirmation checksheets. Check sheets have been improved by adding
a number of enhancements, a few of which are described below
Although they are simple, check sheets are extremely useful processimprovement
and problem-solving tools. Their power is greatly enhanced
when they are used in conjunction with other simple tools, such as histograms
and Pareto analysis. Ishikawa (1985) estimated that 80% to 90% of all workplace
problems could be solved using only the simple quality improvement tools.
Check sheets 255
Figure 8.2. Selected ?ow chart symbols.
Process check sheets
These check sheets are used to create frequency distribution tally sheets that
are, in turn, used to construct histograms (see below). A process check sheet is
constructed by listing several ranges of measurement values and recording a
mark for the actual observations. An example is shown in Figure 8.4. Notice
that if reasonable care is taken in recording tick marks, the check sheet gives a
graphical picture similar to a histogram.
Figure 8.3. Flow chart of process capability analysis.
Defect check sheets
Here the different types of defects are listed and the observed frequencies
recorded.Anexample of a defect check sheet is shown in Figure 8.5. If reasonable
care is taken in recording tick marks, the check sheet resembles a bar chart.
Stratified defect check sheets
These check sheets stratify a particular defect type according to logical criteria.
This is helpful when the defect check sheet fails to provide adequate information
regarding the root cause or causes of a problem. An example is shown
in Figure 8.6.
Check sheets 257
Figure 8.4. Process check sheet.
Figure 8.5. Defect check sheet.
Defect location check sheets
These ‘‘check sheets’’ are actually drawings, photographs, layout diagrams or
maps which show where a particular problem occurs. The spatial location is
valuable in identifying root causes and planning corrective action. In Figure
8.7, the location of complaints from customers about lamination problems on
a running shoe are shown with an ‘‘X.’’ The diagram makes it easy to identify a
problem area that would be difficult to depict otherwise. In this case, a picture
is truly worth a thousand words of explanation.
Figure 8.6. Strati?ed defect check sheet.
Figure 8.7. Defect location check sheet lamination complaints.
Cause and effect diagram check sheets
Cause and effect diagrams can also serve as check sheets. Once the diagram
has been prepared, it is posted in the work area and the appropriate arrow is
marked whenever that particular cause or situation occurs. Teams can also use
this approach for historic data, when such data are available.
DefinitionLPareto analysis is the process of ranking opportunities to
determine which of many potential opportunities should be pursued
first. It is also known as ‘‘separating the vital few from the trivial many.’’
UsageLPareto analysis should be used at various stages in a quality
improvement program to determine which step to take next. Pareto
analysis is used to answer such questions as ‘‘What department should
have the next SPC team?’’ or ‘‘On what type of defect should we concentrate
our efforts?’’
How to perform a Pareto analysis
1. Determine the classi?cations (Pareto categories) for the graph. If the
desired information does not exist, obtain it by designing check sheets
and log sheets.
2. Select a time interval for analysis. The interval should be long enough to
be representative of typical performance.
3. Determine the total occurrences (i.e., cost, defect counts, etc.) for each
category. Also determine the grand total. If there are several categories
which account for only a small part of the total, group these into a category
called ‘‘other.’’
4. Compute the percentage for each category by dividing the category
total by the grand total and multiplying by 100.
5. Rank-order the categories from the largest total occurrences to the
6. Compute the ‘‘cumulative percentage’’ by adding the percentage for
each category to that of any preceding categories.
7. Construct a chart with the left vertical axis scaled from 0 to at least the
grand total. Put an appropriate label on the axis. Scale the right vertical
axis from 0 to 100%, with 100% on the right side being the same height
as the grand total on the left side.
8. Label the horizontal axis with the category names. The left-most category
should be the largest, second largest next, and so on.
Pareto analysis 259
9. Draw in bars representing the amount of each category. The height of
the bar is determined by the left vertical axis.
10. Draw a line that shows the cumulative percentage column of the Pareto
analysis table. The cumulative percentage line is determined by the
right vertical axis.
Example of Pareto analysis
The data in Table 8.1 have been recorded for peaches arriving at Super Duper
Market during August.
The Pareto table for the data in Table 8.1 is shown in Table 8.2.
Table 8.1. Raw data for Pareto analysis.
Bruised 100
Undersized 87
Rotten 235
Underripe 9
Wrong variety 7
Wormy 3
Table 8.2. Data organized for Pareto analysis.
1 Rotten 235 53.29 53.29
2 Bruised 100 22.68 75.97
3 Undersized 87 19.73 95.70
4 Other 19 4.31 100.01
Note that, as often happens, the final percentage is slightly different than
100%. This is due to round-off error and is nothing to worry about. The finished
diagram is shown in Figure 8.8.
Process improvement involves taking action on the causes of variation. With
most practical applications, the number of possible causes for any given problem
can be huge. Dr. Kaoru Ishikawa developed a simple method of graphically
displaying the causes of any given quality problem. His method is called by several
names, the Ishikawa diagram, the fishbone diagram, and the cause and
effect diagram.
Cause and effect diagrams are tools that are used to organize and graphically
display all of the knowledge a group has relating to a particular problem.
Usually, the steps are:
1. Develop a ?ow chart of the area to be improved.
2. De?ne the problem to be solved.
3. Brainstorm to ?nd all possible causes of the problem.
Cause and effect diagrams 261
Figure 8.8. The completed Pareto diagram.
4. Organize the brainstorming results in rational categories.
5. Construct a cause and e?ect diagram that accurately displays the relationships
of all the data in each category.
Once these steps are complete, constructing the cause and effect diagram is
very simple. The steps are:
1. Draw a box on the far right-hand side of a large sheet of paper and draw a
horizontal arrow that points to the box. Inside of the box, write the
description of the problem you are trying to solve.
2. Write the names of the categories above and below the horizontal line.
Think of these as branches from the main trunk of the tree.
3. Draw in the detailed cause data for each category. Think of these as limbs
and twigs on the branches.
A good cause and effect diagram will have many ‘‘twigs,’’ as shown in Figure
8.9. If your cause and effect diagram doesn’t have a lot of smaller branches and
twigs, it shows that the understanding of the problem is superficial. Chances
are you need the help of someone outside of your group to aid in the understanding,
perhaps someone more closely associated with the problem.
Cause and effect diagrams come in several basic types. The dispersion analysis
type is created by repeatedly asking ‘‘why does this dispersion occur?’’ For
example, we might want to know why all of our fresh peaches don’t have the
same color.
The production process class cause and effect diagram uses production processes
as the main categories, or branches, of the diagram. The processes are
shown joined by the horizontal line. Figure 8.10 is an example of this type of diagram.
The cause enumeration cause and effect diagram simply displays all possible
causes of a given problem grouped according to rational categories. This type
of cause and effect diagram lends itself readily to the brainstorming approach
we are using.
Cause and effect diagrams have a number of uses. Creating the diagram is an
education in itself. Organizing the knowledge of the group serves as a guide for
discussion and frequently inspires more ideas. The cause and effect diagram,
once created, acts as a record of your research. Simply record your tests and
results as you proceed. If the true cause is found to be something that wasn’t
on the original diagram, write it in. Finally, the cause and effect diagram is a display
of your current level of understanding. It shows the existing level of technology
as understood by the team. It is a good idea to post the cause and effect
diagram in a prominent location for all to see.
A variation of the basic cause and effect diagram, developed by Dr. Ryuji
Fukuda of Japan, is cause and effect diagrams with the addition of cards, or
CEDAC. The main difference is that the group gathers ideas outside of the
Cause and effect diagrams 263
meeting room on small cards, as well as in group meetings. The cards also serve
as a vehicle for gathering input from people who are not in the group; they can
be distributed to anyone involved with the process. Often the cards provide
more information than the brief entries on a standard cause and effect diagram.
The cause and effect diagram is built by actually placing the cards on the
Since Dr. Shewhart launched modern quality control practice in 1931, the
pace of change in recent years has been accelerating. The 7M tools are an example
of the rapidly changing face of quality technology. While the traditional
QCtools (Pareto analysis, control charts, etc.) are used in the analysis of quantitative
data, the 7M tools apply to qualitative data as well. The ‘‘M’’ stands for
Management, and the tools are focused on managing and planning quality
improvement activities. In recognition of the planning emphasis, these tools
are often referred to as the ‘‘7 MP’’ tools. This section will provide definitions
of the 7M tools. The reader is referred to Mizuno (1988) for additional information
on each of these techniques.
Affinity diagrams
The word affinity means a ‘‘natural attraction’’ or kinship. The affinity diagram
is a means of organizing ideas into meaningful categories by recognizing
their underlying similarity. It is a means of data reduction in that it organizes a
large number of qualitative inputs into a smaller number of major dimensions,
constructs, or categories. The basic idea is that, while there are many variables,
the variables are measuring a smaller number of important factors. For example,
Figure 8.10. Production process class cause and e?ect diagram.
if patients are interviewed about their hospital experience they may say ‘‘the
doctor was friendly,’’ ‘‘the doctor knew what she was doing,’’ and ‘‘the doctor
kept me informed.’’ Each of these statements relates to a single thing, the doctor.
Many times affinity diagrams are constructed using existing data, such as
memos, drawings, surveys, letters, and so on. Ideas are sometimes generated in
brainstorming sessions by teams. The technique works as follows:
1. Write the ideas on small pieces of paper (Post-itsTM or 3  5 cards work
very well).
2. The team works in silence to arrange the ideas into separate categories.
Silence is believed to help because the task involves pattern recognition
and some research shows that for some people, particularly males, language
processing involves the left side of the brain. Research also shows
that left-brain thinking tends to be more linear, which is thought to inhibit
creativity and pattern recognition. Thus, by working silently, the
right brain is more involved in the task. To put an idea into a category a
person simply picks up the Post-itTM and moves it.
3. The ?nal groupings are then reviewed and discussed by the team.
Usually, the grouping of ideas helps the team to develop a coherent plan.
Affinity diagrams are useful for analysis of quality problems, defect data, customer
complaints, survey results, etc. They can be used in conjunction with
other techniques such as cause and effect diagrams or interrelationship digraphs
(see below). Figure 8.11 is an example of an affinity diagram.
Tree diagrams
Tree diagrams are used to break down or stratify ideas in progressively
greater detail. The objective is to partition a big idea or problem into its smaller
components. By doing this you will make the idea easier to understand, or the
problem easier to solve. The basic idea behind this is that, at some level, a problem’s
solution becomes relatively easy to find. Figure 8.12 shows an example
of a tree diagram. Quality improvement would progress from the right-most
portion of the tree diagram to the left-most. Another common usage of tree diagrams
is to show the goal or objective on the left side and the means of accomplishing
the goal, to the right.
Process decision program charts
The process decision program chart (PDPC) is a technique designed to help
prepare contingency plans. It is modeled after reliability engineering methods
of failure mode, effects, and criticality analysis (FMECA) and fault tree analysis
(see Chapter 16). The emphasis of PDPC is the impact of the ‘‘failures’’ (pro-
7M tools 265
Figure 8.11. Software development process a?nity diagram.
From ‘‘Modern approaches to software quality improvement,’’ ?gure 3,
Australian Organization for Quality: Qualcon 90. Copyright#1990 by
Thomas Pyzdek.
blems) on project schedules. Also, PDPC seeks to describe specific actions to be
taken to prevent the problems from occurring in the first place, and to mitigate
the impact of the problems if they do occur. An enhancement to classical
PDPC is to assign subjective probabilities to the various problems and to use
these to help assign priorities. Figure 8.13 shows a PDPC.
7M tools 267
Figure 8.12. An example of a tree diagram.
Figure 8.13. Process decision program chart.
Matrix diagrams
A matrix diagram is constructed to analyze the correlations between two
groups of ideas. Actually, quality function deployment (QFD) is an enhanced
matrix diagram (see Chapter 3 for a discussion of QFD). The major advantage
of constructing matrix diagrams is that it forces you to systematically analyze
correlations. Matrix diagrams can be used in conjunction with decision trees.
To do this, simply use the most detailed level of two decision trees as the contents
of rows and columns of a matrix diagram. An example of a matrix diagram
is shown in Figure 8.14.
Interrelationship digraphs
Like affinity diagrams, interrelationship digraphs are designed as a means of
organizing disparate ideas, usually (but not always) ideas generated in brainstorming
sessions. However, while affinity diagrams seek to simply arrange
related ideas into groups, interrelationship digraphs attempt to define the
ways in which ideas influence one another. It is best to use both affinity diagrams
and interrelationship digraphs.
Figure 8.14. An example of a matrix diagram.
The interrelationship digraph begins by writing down the ideas on small
pieces of paper, such as Post-itsTM. The pieces of paper are then placed on a
large sheet of paper, such as a flip-chart sheet or a piece of large-sized blue-print
paper. Arrows are drawn between related ideas. An idea that has arrows leaving
it but none entering is a ‘‘root idea.’’ By evaluating the relationships between
ideas you will get a better picture of the way things happen. The root ideas are
often keys to improving the system. Figure 8.15 illustrates a simple interrelationship
Prioritization matrices*
To prioritize is to arrange or deal with in order of importance. A prioritization
matrix is a combination of a tree diagram and a matrix chart and it is used
to help decision makers determine the order of importance of the activities
7M tools 269
Figure 8.15. How does ‘‘people management’’ impact change?
*This chart replaces thematrix data analysis chart, formerly one of the 7Mtools. The matrix data analysis chart was based on
factor analysis or principal components analysis. This dependence on heavy-duty statistical methods made it unacceptable
as a tool for use by non-statisticians on a routine basis.
or goals being considered. Prioritization matrices are designed to rationally
narrow the focus of the team to those key issues and options which are
most important to the organization. Brassard (1989, pp. 102^103) presents
three methods for developing prioritization matrices: the full analytical
criteria method, the combination interrelationship digraph (ID)/matrix
method, and the consensus criteria method. We will discuss the three different
The full analytical criteria method is based upon work done by Saaty (1988).
Saaty’s approach is called the analytic hierarchy process (AHP). While analytically
rigorous, AHP is cumbersome in both data collection procedures and the
analysis. This author recommends that this approach be reserved for truly
‘‘heavy-duty’’ decisions of major strategic importance to the organization. In
those cases, you may wish to obtain consulting assistance to assure that the
approach is properly applied. In addition, you may want to acquire software to
assist in the analysis.* Brassard (1989) and Saaty (1988) provide detailed examples
of the application of the full analytical criteria approach.
The interrelationship digraph (ID) is a method used to uncover patterns in
cause and effect relationships (see above). This approach to creating a prioritization
matrix begins with a tree diagram (see above). Items at the right-most
level of the tree diagram (the most detailed level) are placed in a matrix (i.e.,
both the rows and columns of thematrix are the items from the right-most position
of the tree diagram) and their impact on one another evaluated. The ID
matrix is developed by starting with a single item, then adding items one by
one. As each item is added, the team answers the question ‘‘is this item caused
by X?’’ where X is another item. The process is repeated item by item until the
relationship between each item and every other item has been determined. If
the answer is ‘‘yes,’’ then an arrow is drawn between the ‘‘cause’’ item and the
‘‘effect’’ item. The strength of the relationship is determined by consensus. The
final result is an estimate of the relative strength of each item and its effect on
other items.
*At the time of this writing, software was available fromQualityAmerica, Inc., in Tucson, Arizona, and Expert Choice, Inc.,
in Pittsburgh, Pennsylvania.
In Figure 8.16, an ‘‘in’’ arrow points left and indicates that the column item
leads to the row item. On the ID, this would be indicated by an arrow from the
column item to the row item. An ‘‘out’’ arrow points upward and indicates the
opposite of an ‘‘in’’ arrow. To maintain symmetry, if an in arrow appears in a
row/column cell, an out arrow must appear in the corresponding column/row
cell, and vice versa.
Once the final matrix has been created, priorities are set by evaluating the
strength column, the total arrows column, and the relationship between the
number of in and out arrows. An item with a high strength and a large number
of out arrows would be a strong candidate because it is important (high
strength) and it influences a large number of other options (many arrows, predominately
out arrows). Items with high strength and a large number of in
arrows are candidates for outcome measures of success.
The consensus criteria method is a simplified approach to selecting from several
options according to some criteria. It begins with a matrix where the different
options under consideration are placed in rows and the criteria to be used
are shown in columns. The criteria are given weights by the team using the con-
7M tools 271
Figure 8.16. Combination ID/matrix method.
Use the best mix of marketing medium.
sensus decision rule. For example, if criterion #1 were given a weight of 3 and
the group agreed that criterion #2 was twice as important, then criterion #2
would receive a weight of 6. Another way to do the weighting is to give the
team $1 in nickels and have them ‘‘spend’’ the dollar on the various criteria.
The resulting value allocated to each criterion is its weight. The group then
rank-orders the options based on each criterion. Ranks are labeled such that
the option that best meets the criterion gets the highest rank; e.g., if there are
five options being considered for a given criterion, the option that best meets
the criterion is given a rank of 5.
The options are then prioritized by adding up the option’s rank for each criterion
multiplied by the criterion weight.
Example of consensus criteria method
A team had to choose which of four projects to pursue first. To help them
decide, they identified four criteria for selection and their weights as follows:
high impact on bottom line (weight ? 0.25), easy to implement (0.15), low cost
to implement (0.20) and high impact on customer satisfaction (0.40). The four
projects were then ranked according to each of the criteria; the results are
shown in the table below.
In the above example, the team would begin with project #2 because it has
the highest score. If the team had difficulty reaching consensus on the weights
or ranks, they could use totals or a method such as the nominal group technique
described below.
Weight !
Criteria and weights
0.25 0.15 0.2 0.4
line Easy
satisfaction Total
Project 1 1 2 2 1 1.35
Project 2 3 4 4 3 3.35
Project 3 2 1 3 4 2.85
Project 4 4 3 1 2 2.45
Activity network diagram
Activity network diagrams, sometimes called arrow diagrams, have their
roots in well-established methods used in operations research. The arrow diagram
is directly analogous to the critical path method (CPM) and the program
evaluation and review technique (PERT) discussed in Chapter 15. These two
project management tools have been used for many years to determine which
activities must be performed, when they must be performed, and in what
order. Unlike CPM and PERT, which require training in project management
or systems engineering, arrow diagrams are greatly simplified so that they can
be used with a minimum of training. An illustration of an arrow (PERT) diagram,
is reproduced in Figure 8.17.
Other continuous improvement tools
Over the years, the tools of quality improvement have proliferated. By some
estimates there are now over 400 tools in the ‘‘TQM Toolbox.’’ This author
believes that it is possible to make dramatic improvements with the tools
7M tools 273
Figure 8.17. PERT network for constructing a house.
already described, combined with the powerful statistical techniques described
in other parts of this book. However, in addition to the tools already discussed,
there are two more simple tools that the author believes deserve mention: the
nominal group technique, and force-field analysis. These tools are commonly
used to help teams move forward by obtaining input from all interested parties
and identifying the obstacles they face.
The nominal group technique (NGT) is a method for generating a ‘‘short
list’’ of items to be acted upon. The NGT uses a highly structured approach
designed to reduce the usual give-and-take among group members. Usage of
the NGT is indicated when 1) the group is new or has several new members, 2)
when the topic under consideration is controversial, or 3) when the team is
unable to resolve a disagreement. Scholtes (1988) describes the steps involved
in the NGT. A summary of the approach is shown below.
Part ILAformalized brainstorm
1. De?ne the task in the form of a question.
2. Describe the purpose of this discussion and the rules and procedures of
the NGT.
3. Introduce and clarify the question.
4. Generate ideas. Do this by having the team write down their ideas in
5. List the ideas obtained.
6. Clarify and discuss the ideas.
Part IILMaking the selection
1. Choose the top 50 ideas. Note: members can remove their ideas from
consideration if they wish, but no member can remove another’s idea.
2. Pass out index cards to each member, using the following table as a guide:
less than 20
4 cards
6 cards
8 cards
3. Each member writes down their choices from the list, one choice per
4. Each member rank-orders their choices and writes the rank on the cards.
5. Record the group’s choices and ranks.
6. Group reviews and discusses the results. Consider: How often was an
item selected? What is the total of the ranks for each item?
If the team can agree on the importance of the item(s) that got the highest
score(s) (sum of ranks), then the team moves on to preparing an action plan to
deal with the item or items selected.
Force-field analysis (FFA) is a method borrowed from the mechanical engineering
discipline known as free-body diagrams. Free-body diagrams are
drawn to help the engineer identify all the forces surrounding and acting on a
body. The objective is to ascertain the forces leading to an equilibrium state for
the body.
In FFA the ‘‘equilibrium’’ is the status quo. FFA helps the team understand
the forces that keep things the way they are. Some of the forces are
‘‘drivers’’ that move the system towards a desired goal. Other forces are
‘‘restrainers’’ that prevent the desired movement and may even cause movement
away from the goal. Once the drivers and restrainers are known, the
team can design an action plan which will 1) reduce the forces restraining
progress and 2) increase the forces which lead to movement in the desired
FFA is useful in the early stages of planning. Once restrainers are explicitly
identified, a strategic plan can be prepared to develop the drivers necessary to
overcome them. FFA is also useful when progress has stalled. By performing
FFA, people are brought together and guided toward consensus, an activity
that, by itself, often overcomes a number of obstacles. Pyzdek (1994) lists the following
steps for conducting FFA.
1. Determine the goal.
2. Create a team of individuals with the authority, expertise, and interest
needed to accomplish the goal.
3. Have the team use brainstorming or theNGTto identify restrainers and
4. Create a force-?eld diagram or table which lists the restrainers and
5. Prepare a plan for removing restrainers and increasing drivers.
An example of a force-field diagram is shown in Figure 8.18.
7M tools 275
It may be helpful to assign ‘‘strength weights’’ to the drivers and restrainers
(e.g., weak, moderate, strong).
Figure 8.18. Example of a force-?eld diagram.
From Pocket Guide to Quality Tools, p. 10. Copyright#1995 by Thomas Pyzdek.
^ ^ ^
Basic Principles of
An argument can be made for asserting that quality begins with measurement.
Only when quality is quantified can meaningful discussion about
improvement begin. Conceptually, measurement is quite simple: measurement
is the assignment of numbers to observed phenomena according to
certain rules. Measurement is a sine qua non of any science, including management
A measurement is simply a numerical assignment to something, usually a
non-numerical element. Measurements convey certain information about the
relationship between the element and other elements. Measurement involves a
theoretical domain, an area of substantive concern represented as an empirical
relational system, and a domain represented by a particular selected numerical
relational system. There is a mapping function that carries us from the empirical
system into the numerical system. The numerical system is manipulated and
the results of the manipulation are studied to help the manager better understand
the empirical system.
In reality, measurement is problematic: the manager can never know the
‘‘true’’ value of the element being measured. The numbers provide information
on a certain scale and they represent measurements of some unobservable vari-
Copyright 2003 by The McGraw-Hill Companies, Inc. Click Here for Terms of Use.
able of interest. Some measurements are richer than others, i.e., some measurements
provide more information than other measurements. The information
content of a number is dependent on the scale of measurement used. This scale
determines the types of statistical analyses that can be properly employed in
studying the numbers. Until one has determined the scale of measurement,
one cannot know if a given method of analysis is valid.
The four measurement scales are: nominal, ordinal, interval, and ratio.
Harrington (1992) summarizes the properties of each scale in Table 9.1.
Numbers on a nominal scale aren’t measurements at all, they are merely category
labels in numerical form. Nominal measurements might indicate membership
in a group (1?male, 2?female) or simply represent a designation
(John Doe is #43 on the team). Nominal scales represent the simplest and
Table 9.1. Types of measurement scales and permissible statistics.
From Quality Engineering Handbook, p. 516. Copyright#1992. Used by permission of
the publisher, ASQ Quality Press, Milwaukee, Wisconsin.
Nominal Only the presence/absence of
an attribute; can only count
go/no go;
chi-square tests
Ordinal Can say that one item has
more or less of an attribute
than another item; can order a
set of items
Interval Di?erence between any two
successive points is equal;
often treated as a ratio scale
even if assumption of equal
intervals is incorrect; can add,
subtract, order objects
calendar time;
t-tests; F-tests;
Ratio True zero point indicates
absence of an attribute; can
add, subtract, multiply and
elapsed time;
distance; weight
t-test; F-test;
weakest form of measurement. Nominal variables are perhaps best viewed as a
form of classification rather than as a measurement scale. Ideally, categories on
the nominal scale are constructed in such a way that all objects in the universe
are members of one and only one class. Data collected on a nominal scale are
called attribute data. The only mathematical operations permitted on nominal
scales are ? (which shows that an object possesses the attribute of concern)
or 6?.
An ordinal variable is one that has a natural ordering of its possible values,
but for which the distances between the values are undefined. An example is
product preference rankings such as good, better, best. Ordinal data can be analyzed
with the mathematical operators, ?(equality), 6? (inequality), >(greater
than) and < (less than). There are a wide variety of statistical techniques which
can be applied to ordinal data including the Pearson correlation. Other ordinal
models include odds-ratio measures, log-linear models and logit models, both
of which are used to analyze cross-classifications of ordinal data presented in
contingency tables. In quality management, ordinal data are commonly converted
into nominal data and analyzed using binomial or Poisson models. For
example, if parts were classified using a poor-good-excellent ordering, the quality
analyst might plot a p chart of the proportion of items in the poor category.
Interval scales consist of measurements where the ratios of differences are
invariant. For example, 908C ?1948F, 1808C ?3568F, 2708C ? 5188F, 3608C
?6808F. Now 1948F/908C 6? 3568F/1808C but
3568F  1948F
6808F  5188F ?
1808C  908C
3608C  2708C
Conversion between two interval scales is accomplished by the transformation
y ? ax ? b; a > 0. For example, 8F ? 32 ? ?95
 8C), where a ? 9=5,
b ? 32. As with ratio scales, when permissible transformations are made statistical,
results are unaffected by the interval scale used. Also, 08 (on either scale)
is arbitrary. In this example, zero does not indicate an absence of heat.
Ratio scale measurements are so called because measurements of an object in
two different metrics are related to one another by an invariant ratio. For example,
if an object’s mass was measured in pounds (x) and kilograms (y), then
x=y ? 2:2 for all values of x and y. This implies that a change from one ratio
measurement scale to another is performed by a transformation of the form
y ? ax, a > 0; e.g., pounds = 2.2  kilograms. When permissible transformations
are used, statistical results based on the data are identical regardless of
the ratio scale used. Zero has an inherent meaning: in this example it signifies
an absence of mass.
Scales of measurement 279
Fundamentally, any item measure should meet two tests:
1. The item measures what it is intended to measure (i.e., it is valid).
2. Aremeasurement would order individual responses in the same way (i.e.,
it is reliable).
The remainder of this section describes techniques and procedures
designed to assure that measurement systems produce numbers with these
properties. A good measurement system possesses certain properties. First, it
should produce a number that is ‘‘close’’ to the actual property being
measured, i.e., it should be accurate. Second, if the measurement system is
applied repeatedly to the same object, the measurements produced should be
close to one another, i.e., it should be repeatable. Third, the measurement
system should be able to produce accurate and consistent results over the
entire range of concern, i.e., it should be linear. Fourth, the measurement
system should produce the same results when used by any properly trained
individual, i.e., the results should be reproducible. Finally, when applied to
the same items the measurement system should produce the same results in
the future as it did in the past, i.e., it should be stable. The remainder of this
section is devoted to discussing ways to ascertain these properties for particular
measurement systems. In general, the methods and definitions presented
here are consistent with those described by the Automotive Industry Action
Group (AIAG).
BiasLThe difference between the average measured value and a reference
value is referred to as bias. The reference value is an agreed-upon standard,
such as a standard traceable to a national standards body (see
below). When applied to attribute inspection, bias refers to the ability
of the attribute inspection system to produce agreement on inspection
standards. Bias is controlled by calibration, which is the process of
comparing measurements to standards. The concept of bias is illustrated
in Figure 9.1.
RepeatabilityLAIAG defines repeatability as the variation in measurements
obtained with one measurement instrument when used several
times by one appraiser, while measuring the identical characteristic on
the same part. Variation obtained when the measurement system is
applied repeatedly under the same conditions is usually caused by conditions
inherent in the measurement system.
ASQ defines precision as ‘‘The closeness of agreement between randomly
selected individual measurements or test results. NOTE: The
standard deviation of the error of measurement is sometimes called
‘imprecision’.’’ This is similar to what we are calling repeatability.
Repeatability is illustrated in Figure 9.2.
ReproducibilityLReproducibility is the variation in the average of the measurements
made by different appraisers using the same measuring
instrument when measuring the identical characteristic on the same
part. Reproducibility is illustrated in Figure 9.3.
StabilityLStability is the total variation in the measurements obtained with
a measurement system on the same master or parts when measuring a
single characteristic over an extended time period. A system is said to
be stable if the results are the same at different points in time. Stability
is illustrated in Figure 9.4.
LinearityLthe difference in the bias values through the expected operating
range of the gage. Linearity is illustrated in Figure 9.5.
Reliability and validity of data 281
Figure 9.1. Bias illustrated.
Figure 9.2. Repeatability illustrated.
Figure 9.3. Reproducibility illustrated. Figure 9.4. Stability illustrated.
Enumerative versus analytic statistical methods
How would you respond to the following question?
A sample of 100 bottles taken from a ?lling process has an average of 11.95
ounces and a standard deviation of 0.1 ounce. The speci?cations are 11.9^
12.1 ounces. Based on these results, should you
a. Do nothing?
b. Adjust the average to precisely 12 ounces?
Overview of statistical methods 283
Figure 9.5. Linearity illustrated.
c. Compute a confidence interval about the mean and adjust the
process if the nominal fill level is not within the confidence
d. None of the above?
The correct answer is d, none of the above. The other choices all make the
mistake of applying enumerative statistical concepts to an analytic statistical
situation. In short, the questions themselves are wrong! For example, based on
the data, there is no way to determine if doing nothing is appropriate. ‘‘Doing
something’’ implies that there is a known cause and effect mechanism which
can be employed to reach a known objective. There is nothing to suggest that
this situation exists. Thus, we can’t simply adjust the process average to the
nominal value of 12 ounces, even though the process appears to be 5 standard
errors below this value. This might have happened because the first 50 were 10
standard errors below the nominal and the last 50 were exactly equal to the nominal
(or any of a nearly infinite number of other possible scenarios). The confidence
interval calculation fails for the same reason. Figure 9.6 illustrates some
processes that could produce the statistics provided above.
Some appropriate analytic statistics questions might be:
. Is the process central tendency stable over time?
. Is the process dispersion stable over time?
. Does the process distribution change over time?
If any of the above are answered ‘‘no,’’ then what is the cause of the instability?
To help answer this question, ask ‘‘what is the nature of the variation as
revealed by the patterns?’’ when plotted in time-sequence and stratified in various
If none of the above are answered ‘‘no,’’ then, and only then, we can ask such
questions as
. Is the process meeting the requirements?
. Can the process meet the requirements?
. Can the process be improved by recentering it?
. How can we reduce variation in the process?
Deming (1975) defines enumerative and analytic studies as follows:
Enumerative studyLa study in which action will be taken on the universe.
Analytic studyLa study in which action will be taken on a process to
improve performance in the future.
The term ‘‘universe’’ is defined in the usual way: the entire group of interest,
e.g., people, material, units of product, which possess certain properties of interest.
An example of an enumerative study would be sampling an isolated lot to
determine the quality of the lot.
In an analytic study the focus is on a process and how to improve it. The focus
is the future. Thus, unlike enumerative studies which make inferences about
the universe actually studied, analytic studies are interested in a universe
which has yet to be produced. Table 9.2 compares analytic studies with enumerative
studies (Provost, 1988).
Deming (1986) points out that ‘‘Analysis of variance, t-tests, confidence
intervals, and other statistical techniques taught in the books, however interesting,
are inappropriate because they provide no basis for prediction and
because they bury the information contained in the order of production.’’
Overview of statistical methods 285
Figure 9.6. Possible processes with similar means and sigmas.
These traditional statistical methods have their place, but they are widely
abused in the real world. When this is the case, statistics do more to cloud the
issue than to enlighten.
Analytic study methods provide information for inductive thinking, rather
than the largely deductive approach of enumerative statistics. Analytic methods
are primarily graphical devices such as run charts, control charts, histograms,
interrelationship digraphs, etc. Analytic statistics provide
operational guidelines, rather than precise calculations of probability. Thus,
such statements as ‘‘There is a 0.13% probability of a Type I error when acting
on a point outside a three-sigma control limit’’ are false (the author admits
to having made this error in the past). The future cannot be predicted with a
known level of confidence. Instead, based on knowledge obtained from every
source, including analytic studies, one can state that one has a certain degree
of belief (e.g., high, low) that such and such will result from such and such
action on a process.
Another difference between the two types of studies is that enumerative
statistics proceed from predetermined hypotheses while analytic studies try
Table 9.2. Important aspects of analytic studies.
Aim Parameter estimation Prediction
Focus Universe Process
Method of
Counts, statistics Models of the process
(e.g., ?ow charts, cause
and e?ects, mathematical
Major source of
Sampling variation Extrapolation into the
Yes No
for the study
Static Dynamic
to help the analyst generate new hypotheses. In the past, this extremely worthwhile
approach has been criticized by some statisticians as ‘‘fishing’’ or ‘‘rationalizing.’’
However, this author believes that using data to develop plausible
explanations retrospectively is a perfectly legitimate way of creating new theories
to be tested. To refuse to explore possibilities suggested by data is to
take a very limited view of the scope of statistics in quality improvement and
Enumerative statistical methods
This section discusses the basic concept of statistical inference. The reader
should also consult the Glossary in the Appendix for additional information.
Inferential statistics belong to the enumerative class of statistical methods.
The term inference is defined as 1) the act or process of deriving logical conclusions
from premises known or assumed to be true, or 2) the act of reasoning
from factual knowledge or evidence. Inferential statistics provide information
that is used in the process of inference. As can be seen from the definitions, inference
involves two domains: the premises and the evidence or factual knowledge.
Additionally, there are two conceptual frameworks for addressing premises
questions in inference: the design-based approach and the model-based
As discussed by Koch and Gillings (1983), a statistical analysis whose only
assumptions are random selection of units or random allocation of units to
experimental conditions results in design-based inferences; or, equivalently,
randomization-based inferences. The objective is to structure sampling such
that the sampled population has the same characteristics as the target population.
If this is accomplished then inferences from the sample are said to have
internal validity. A limitation on design-based inferences for experimental studies
is that formal conclusions are restricted to the finite population of subjects
that actually received treatment, that is, they lack external validity.
However, if sites and subjects are selected at random from larger eligible
sets, then models with random effects provide one possible way of addressing
both internal and external validity considerations. One important consideration
for external validity is that the sample coverage includes all relevant
subpopulations; another is that treatment differences be homogeneous across
subpopulations. A common application of design-based inference is the
Alternatively, if assumptions external to the study design are required to
extend inferences to the target population, then statistical analyses based on
postulated probability distributional forms (e.g., binomial, normal, etc.) or
other stochastic processes yield model-based inferences. A focus of distinction
Overview of statistical methods 287
between design-based and model-based studies is the population to which the
results are generalized rather than the nature of the statistical methods applied.
When using a model-based approach, external validity requires substantive justification
for the model’s assumptions, as well as statistical evaluation of the
Statistical inference is used to provide probabilistic statements regarding a
scientific inference. Science attempts to provide answers to basic questions,
such as can this machine meet our requirements? Is the quality of this lot
within the terms of our contract? Does the new method of processing produce
better results than the old? These questions are answered by conducting an
experiment, which produces data. If the data vary, then statistical inference
is necessary to interpret the answers to the questions posed. A statistical
model is developed to describe the probabilistic structure relating the
observed data to the quantity of interest (the parameters), i.e., a scientific
hypothesis is formulated. Rules are applied to the data and the scientific
hypothesis is either rejected or not. In formal tests of a hypothesis, there are
usually two mutually exclusive and exhaustive hypotheses formulated: a null
hypothesis and an alternate hypothesis. Formal hypothesis testing is discussed
later in this chapter.
Data are said to be discrete when they take on only a finite number of points
that can be represented by the non-negative integers. An example of discrete
data is the number of defects in a sample. Data are said to be continuous when
they exist on an interval, or on several intervals. An example of continuous
data is the measurement of pH. Quality methods exist based on probability
functions for both discrete and continuous data.
Enumeration involves counting techniques for very large numbers of possible
outcomes. This occurs for even surprisingly small sample sizes. In Six
Sigma, these methods are commonly used in a wide variety of statistical procedures.
The basis for all of the enumerative methods described here is the multiplication
principle. The multiplication principle states that the number of possible
outcomes of a series of experiments is equal to the product of the number
of outcomes of each experiment. For example, consider flipping a coin twice.
On the first flip there are two possible outcomes (heads/tails) and on the second
flip there are also two possible outcomes. Thus, the series of two flips can result
in 2  2 ? 4 outcomes. Figure 9.7 illustrates this example.
An ordered arrangement of elements is called a permutation. Suppose that
you have four objects and four empty boxes, one for each object. Consider
how many different ways the objects can be placed into the boxes. The first
object can be placed in any of the four boxes. Once this is done there are three
boxes to choose from for the second object, then two boxes for the third object
and finally one box left for the last object. Using the multiplication principle
you find that the total number of arrangements of the four objects into the
four boxes is 4  3  2  1 ? 24. In general, if there are n positions to be filled
with n objects there are
n?n  1? . . . ?2??1? ? n! ?9:1?
possible arrangements. The symbol n! is read n factorial. By de?nition, 0! ? 1.
In applying probability theory to discrete variables in quality control we frequently
encounter the need for efficient methods of counting. One counting
technique that is especially useful is combinations. The combination formula
is shown in Equation 9.2.
r ?
r!?n  r?! ?9:2?
Overview of statistical methods 289
Figure 9.7. Multiplication principle applied to coin ?ips.
Combinations tell how many unique ways you can arrange n objects taking
them in groups of r objects at a time, where r is a positive integer less than or
equal to n. For example, to determine the number of combinations we can
make with the letters X,Y, and Z in groups of 2 letters at a time, we note that
n ? 3 letters, r ? 2 letters at a time and use the above formula to find
2 ?
2!?3  2?! ?
3  2  1
?2  1??1? ?
2 ? 3
The 3 combinations are XY, XZ, and YZ. Notice that this method does not
count reversing the letters as separate combinations, i.e., XY and YX are considered
to be the same.
Assumptions and robustness of tests
It is important at the outset to comment on what we are not discussing here
when we use the term ‘‘robustness.’’ First, we are not talking about the sensitivity
of a particular statistic to outliers. This concept is more properly
referred to as resistance and it is discussed in the exploratory data analysis
section of this book. We are also not speaking of a product design that can
perform well under a wide variety of operating conditions. This design-based
definition of robustness is discussed in the Taguchi robustness concepts
All statistical procedures rest upon certain assumptions. For example,
ANOVA assumes that the data are normally distributed with equal variances.
When we use the term robustness here, we mean the ability of the statistical procedure
to produce the correct final result when the assumptions are violated. A
statistical procedure is said to be robust when it can be used even when the
basic assumptions are violated to a small degree.
How large a departure from the assumptions is acceptable? Or, equivalently,
how small is a ‘‘small’’ degree of error? For a given violation of the assumptions,
how large an error in the result is acceptable? Regrettably, there is no rigorous
mathematical definition of the term ‘‘robust.’’
In practice, robustness is commonly addressed in two ways. One approach is
to test the underlying assumptions prior to using a given statistical procedure.
In the case of ANOVA, for example, the practitioner would test the assumptions
of normality and constant variance on the data set before accepting the
results of the ANOVA.
Another approach is to use robust statistical procedures. Some ways of dealing
with the issue are:
. Use procedures with less restrictive underlying assumptions (e.g., nonparametric
. Drop ‘‘gross outliers’’ from the data set before proceeding with the analysis
(using an acceptable statistical method to identify the outliers).
. Use more resistant statistics (e.g., the median instead of the arithmetic
Distributions are a set of numbers collected from a well-defined universe of
possible measurements arising from a property or relationship under study.
Distributions show the way in which the probabilities are associated with
the numbers being studied. Assuming a state of statistical control, by consulting
the appropriate distribution one can determine the answer to such
questions as:
. What is the probability that x will occur?
. What is the probability that a value less than x will occur?
. What is the probability that a value greater than x will occur?
. What is the probability that a value will occur that is between x and y?
By examining plots of the distribution shape, one can determine how rapidly
or slowly probabilities change over a given range of values. In short, distributions
provide a great deal of information.
Afrequency distribution is an empirical presentation of a set of observations.
If the frequency distribution is ungrouped, it simply shows the observations
and the frequency of each number. If the frequency distribution is grouped,
then the data are assembled into cells, each cell representing a subset of the
total range of the data. The frequency in each cell completes the grouped frequency
distribution. Frequency distributions are often graphically displayed
in histograms or stem-and-leaf plots.
While histograms and stem-and-leaf plots show the frequency of specific
values or groups of values, analysts often wish to examine the cumulative frequency
of the data. The cumulative frequency refers to the total up to and
including a particular value. In the case of grouped data, the cumulative frequency
is computed as the total number of observations up to and including a
cell boundary. Cumulative frequency distributions are often displayed on an
ogive, as depicted in Figure 9.8.
Overview of statistical methods 291
In most Six Sigma projects involving enumerative statistics, we deal with
samples, not populations. In the previous section, sample data were used to construct
an ogive and, elsewhere in this book, sample data are used to construct
histograms, stem-and-leaf plots, boxplots, and to compute various statistics.
We now consider the estimation of certain characteristics or parameters of the
distribution from the data.
The empirical distribution assigns the probability 1/n to each Xi in the sample,
thus the mean of this distribution is
X ?Xn
n ?9:3?
The symbol XX is called ‘‘X bar.’’ Since the empirical distribution is determined
by a sample, XX is simply called the sample mean.
The variance of the empirical distribution is given by
S2 ?
n  1Xn
i?1 ?Xi  XX?2 ?9:4?
Figure 9.8. Ogive of rod diameter data.
This equation for S2 is commonly referred to as the sample variance. The
unbiased sample standard deviation is given by
S ? ffiffiffiffi S2 p
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi Pn
i?1?Xi  XX?2
n  1
Another sampling statistic of special interest in Six Sigma is the standard
deviation of the sample average, also referred to as the standard error of the
mean or simply the standard error. This statistic is given by
S XX ?
ffiffiffi n p ?9:6?
As can be seen, the standard error of the mean is inversely proportional to the
square root of the sample size. That is, the larger the sample size, the smaller
the standard deviation of the sample average. This relationship is shown in
Figure 9.9. It can be seen that averages of n=4 have a distribution half as variable
as the population from which the samples are drawn.
Overview of statistical methods 293
Figure 9.9. E?ect of sample size on the standard error.
Probability distributions for Six Sigma
This section discusses the following probability distributions often used in
Six Sigma:
. Binomial distribution
. Poisson distribution
. Hypergeometric distribution
. Normal distribution
. Exponential distribution
. Chi-square distribution
. Student’s t distribution
. F distribution
Assume that a process is producing some proportion of non-conforming
units, which we will call p. If we are basing p on a sample we find p by dividing
the number of non-conforming units in the sample by the number of items
sampled. The equation that will tell us the probability of getting x defectives in
a sample of n units is shown by Equation 9.7.
P?x? ? Cn
xpx?1  p?nx ?9:7?
This equation is known as the binomial probability distribution. In addition
to being useful as the exact distribution of non-conforming units for processes
in continuous production, it is also an excellent approximation to the cumbersome
hypergeometric probability distribution when the sample size is less
than 10% of the lot size.
Example of applying the binomial probability distribution
Aprocess is producing glass bottles on a continuous basis. Past history shows
that1%of the bottles have one or more flaws. Ifwedraw a sample of 10 units from
the process, what is the probability that there will be 0 non-conforming bottles?
Using the above information, n = 10, p = .01, and x = 0. Substituting these
values into Equation 9.7 gives us
P?0? ? C10
0 0:010?1  0:01?100 ? 1  1  0:9910 ? 0:904 ? 90:4%
Anotherwayof interpreting theaboveexampleis thata samplingplan ‘‘inspect
10 units, accept the process if no non-conformances are found’’ has a 90.4% probability
of accepting a process that is averaging 1%non-conforming units.
Example of binomial probability calculations using Microsoft
Microsoft Excel has a built-in capability to analyze binomial probabilities. To
solve the above problem using Excel, enter the sample size, p value, and x value
as shown in Figure 9.10. Note the formula result near the bottom of the screen.
Poisson distribution
Another situation encountered often in quality control is that we are not just
concerned with units that don’t conform to requirements, instead we are concerned
with the number of non-conformances themselves. For example, let’s
say we are trying to control the quality of a computer. A complete audit of the
finished computer would almost certainly reveal some non-conformances,
even though these non-conformances might be of minor importance (for example,
a decal on the back panel might not be perfectly straight). If we tried to use
the hypergeometric or binomial probability distributions to evaluate sampling
plans for this situation, we would find they didn’t work because our lot or pro-
Overview of statistical methods 295
Figure 9.10. Example of ?nding binomial probability using Microsoft Excel.
cess would be composed of 100% non-conforming units. Obviously, we are
interested not in the units per se, but in the non-conformances themselves. In
other cases, it isn’t even possible to count sample units per se. For example, the
number of accidents must be counted as occurrences. The correct probability
distribution for evaluating counts of non-conformances is the Poisson distribution.
The pdf is given in Equation 9.8.
P?x? ?
x! ?9:8?
In Equation 9.8, m is the average number of non-conformances per unit, x is
the number of non-conformances in the sample, and e is the constant approximately
equal to 2.7182818. P(x) gives the probability of exactly x occurrences in
the sample.
Example of applying the Poisson distribution
A production line is producing guided missiles. When each missile is completed,
an audit is conducted by an Air Force representative and every nonconformance
to requirements is noted. Even though any major nonconformance
is cause for rejection, the prime contractor wants to control
minor non-conformances as well. Such minor problems as blurred stencils,
small burrs, etc., are recorded during the audit. Past history shows that on the
average each missile has 3 minor non-conformances. What is the probability
that the next missile will have 0 non-conformances?
We have m ? 3, x ? 0. Substituting these values into Equation 9.8 gives us
P?0? ?
0! ?
1  0:05
1 ? 0:05 ? 5%
In other words, 100%5% ? 95%of the missiles will have at least one nonconformance.
The Poisson distribution, in addition to being the exact distribution for the
number of non-conformances, is also a good approximation to the binomial distribution
in certain cases. To use the Poisson approximation, you simply let
 ? np in Equation 9.8. Juran (1988) recommends considering the Poisson
approximation if the sample size is at least 16, the population size is at least 10
times the sample size, and the probability of occurrence p on each trial is less
than 0.1. The major advantage of this approach is that it allows you to use the
tables of the Poisson distribution, such as Table 7 in the Appendix. Also, the
approach is useful for designing sampling plans.
Example of Poisson probability calculations using Microsoft
Microsoft Excel has a built-in capability to analyze Poisson probabilities. To
solve the above problem using Excel, enter the average and x values as shown
in Figure 9.11. Note the formula result near the bottom of the screen.
Assume we have received a lot of 12 parts from a distributor. We need the
parts badly and are willing to accept the lot if it has fewer than 3 non-conforming
parts. We decide to inspect only 4 parts since we can’t spare the time to check
every part. Checking the sample, we find 1 part that doesn’t conform to the
requirements. Should we reject the remainder of the lot?
This situation involves sampling without replacement. We draw a unit from
the lot, inspect it, and draw another unit from the lot. Furthermore, the lot is
Overview of statistical methods 297
Figure 9.11. Example of ?nding Poisson probability using Microsoft Excel.
quite small, the sample is 25% of the entire lot. The formula needed to compute
probabilities for this procedure is known as the hypergeometric probability distribution,
and it is shown in Equation 9.9.
P?x? ?
nx Cm
n ?9:9?
In the above equation, N is the lot size, m is the number of defectives in the
lot, n is the sample size, x is the number of defectives in the sample, and P(x) is
the probability of getting exactly x defectives in the sample. Note that the
numerator term CNm
nx gives the number of combinations of non-defectives
while Cm
x is the number of combinations of defectives. Thus the numerator
gives the total number of arrangements of samples from lots of size N with m
defectives where the sample n contains exactly x defectives. The term CN
n in
the denominator is the total number of combinations of samples of size n from
lots of size N, regardless of the number of defectives. Thus, the probability is a
ratio of the likelihood of getting the result under the assumed conditions.
For our example, we must solve the above equation for x = 0 as well as x = 1,
since we would also accept the lot if we had no defectives. The solution is
shown as follows.
P?0? ?
40 C3
4 ?
126  1
495 ? 0:255
P?1? ?
41 C3
4 ?
84  3
495 ?
495 ? 0:509
P?1 or less? ? P?0? ? P?1?
Adding the two probabilities tells us the probability that our sampling plan
will accept lots of 12 with 3 non-conforming units. The plan of inspecting 4
parts and accepting the lot if we have 0 or 1 non-conforming has a probability
of .255 + .509 = .764, or 76.4%, of accepting this ‘‘bad’’ quality lot. This is the
‘‘consumer’s risk’’ for this sampling plan. Such a high sampling risk would be
unacceptable to most people.
Example of hypergeometric probability calculations using
Microsoft Excel
Microsoft Excel has a built-in capability to analyze hypergeometric probabilities.
To solve the above problem using Excel, enter the population and sample
values as shown in Figure 9.12. Note the formula result near the bottom of the
screen (0.509) gives the probability for x?1. To find the cumulative probability
you need to sum the probabilities for x ?0 and x ?1 etc.
The most common continuous distribution encountered in Six Sigma work
is, by far, the normal distribution. Sometimes the process itself produces an
approximately normal distribution, other times a normal distribution can be
obtained by performing a mathematical transformation on the data or by
using averages. The probability density function for the normal distribution is
given by Equation 9.10.
f ?x? ?
 ffiffiffiffiffi 2 p e?x?2=22
If f ?x? is plotted versus x, the well-known ‘‘bell curve’’ results. The normal
distribution is also known as the Gaussian distribution. An example is shown
in Figure 9.13.
Overview of statistical methods 299
Figure 9.12. Example of ?nding hypergeometric probability using Microsoft Excel.
In Equation 9.10, m is the population average or mean and s is the population
standard deviation. These parameters have been discussed earlier in this
Example of calculating m, s2 and s
Find m, s2 and s for the following data:
i xi
1 17
2 23
3 5
Table 11.4 gives the equation for the population mean as:
Figure 9.13. The normal/Gaussian curve.
To find the mean for our data we compute
3 ?17 ? 23 ? 5? ? 15
The variance and standard deviation are both measures of dispersion or
spread. The equations for the population variance s2 and standard deviation s
are given in Table 11.4.
2 ? PNi
?1 ?xi?2
 ? ffiffiffi  p 2
Referring to the data above with a mean m of 15, we compute s2 and s as
i xi xi   ?xi  ?2
1 17 2 4
2 23 8 64
3 5 10 100
SUM 168
2 ? 168=3 ? 56
 ? ffiffiffi  p 2 ? ffiffi5 p 6  7:483
Usually we have only a sample and not the entire population. A population is
the entire set of observations from which the sample, a subset, is drawn.
Calculations for the sample mean, variance, and standard deviation were
shown earlier in this chapter.
The areas under the normal curve can be found by integrating Equation 9.10
using numerical methods, but, more commonly, tables are used. Table 2 in the
Appendix gives areas under the normal curve. The table is indexed by using the
Z transformation, which is
Z ?
Overview of statistical methods 301
for population data, or
Z ?
xi  XX
s ?9:12?
for sample data.
By using the Z transformation, we can convert any normal distribution into a
normal distribution with a mean of 0 and a standard deviation of 1. Thus, we
can use a single normal table to find probabilities.
The normal distribution is very useful in predicting long-term process yields.
Assume we have checked the breaking strength of a gold wire bonding process
used in microcircuit production and we have found that the process average
strength is 9# and the standard deviation is 4#. The process distribution is
normal. If the engineering specification is 3# minimum, what percentage of
the process will be below the low specification?
Since our data are a sample, we must compute Z using Equation 9.12.
Z ?
3  9
4 ? 6
4 ? 1:5
Figure 9.14 illustrates this situation.
Figure 9.14. Illustration of using Z tables for normal areas.
Entering Table 2 in the Appendix for Z ? 1:5, we find that 6.68% of the
area is below this Z value. Thus 6.68% of our breaking strengths will be below
our low specification limit of 3#. In quality control applications, we usually
try to have the average at least 3 standard deviations away from the specification.
To accomplish this, we would have to improve the process by either raising the
average breaking strength or reducing the process standard deviation, or both.
Example of normal probability calculations using Microsoft
Microsoft Excel has a built-in capability to analyze normal probabilities. To
solve the above problem using Excel, enter the average, sigma and x values as
shown in Figure 9.15. The formula result near the bottom of the screen gives
the desired probability.
Overview of statistical methods 303
Figure 9.15. Example of ?nding normal probability using Microsoft Excel.
Another distribution encountered often in quality control work is the exponential
distribution. The exponential distribution is especially useful in analyzing
reliability (see Chapter 16). The equation for the probability density
function of the exponential distribution is
f ?x? ?

ex=; x  0 ?9:13?
Unlike the normal distribution, the shape of the exponential distribution is
highly skewed and there is a much greater area below the mean than above it.
In fact, over 63% of the exponential distribution falls below the mean. Figure
9.16 shows an exponential pdf.
Unlike the normal distribution, the exponential distribution has a closed
form cumulative density function (cdf), i.e., there is an equation which gives
the cumulative probabilities directly. Since the probabilities can be determined
directly from the equation, no tables are necessary. See equation 9.14.
P?X  x? ? 1  ex= ?9:14?
Figure 9.16. Exponential pdf curve.
Example of using the exponential cdf
A city water company averages 500 system leaks per year. What is the probability
that the weekend crew, which works from 6 p.m. Friday to 6 a.m.
Monday, will get no calls?
We have m ? 500 leaks per year, which we must convert to leaks per hour.
There are 365 days of 24 hours each in a year, or 8760 hours. Thus, mean time
between failures (MTBF) is 8760/500 =17.52 hours. There are 60 hours between
6 p.m. Friday and 6 a.m.Monday. Thus x ? 60. Using Equation 9.14 gives
P?X  60? ? 1  e60=17:52 ? 0:967 ? 96:7%
Thus, the crew will get to loaf away 3.3% of the weekends.
Example of exponential probability calculations using
Microsoft Excel
Microsoft Excel has a built-in capability to analyze exponential probabilities.
To solve the above problem using Excel, enter the average and x values as
shown in Figure 9.17. Note that Excel uses ‘‘Lambda’’ rather than the average
Overview of statistical methods 305
Figure 9.17. Example of ?nding exponential probability using Microsoft Excel.
in its calculations; lambda is the reciprocal of the average. The formula result
near the bottom of the screen gives the desired probability.
These three distributions are used in Six Sigma to test hypotheses, construct
confidence intervals, and compute control limits.
Many characteristics encountered in Six Sigma have normal or approximately
normal distributions. It can be shown that in these instances the distribution
of sample variances has the form (except for a constant) of a chi-square
distribution, symbolized 	2. Tables have been constructed giving abscissa
values for selected ordinates of the cumulative 	2 distribution. One such table
is Table 4 in the Appendix.
The 	2 distribution varies with the quantity v, which for our purposes is
equal to the sample size minus 1. For each value of v there is a different 	2 distribution.
Equation 9.15 gives the pdf for the 	2.
f ?	2? ?
e	2=2?	2??v2?=2
2v=2 v  2
! ?9:15?
Figure 9.18 shows the pdf for v ? 4.
The use of 	2 is illustrated in this example to find the probability that the variance
of a sample of n items from a specified normal universe will equal or
exceed a given value s2; we compute 	2 ? ?n  1?s2=2. Now, let’s suppose
that we sample n ? 10 items from a process with 2 ? 25 and wish to determine
the probability that the sample variance will exceed 50. Then
?n  1?s2
2 ?
25 ? 18
We enter Appendix Table 4 (	2) at the line for v ? 10  1 ? 9 and note that
18 falls between the columns for the percentage points of 0.025 and 0.05. Thus,
the probability of getting a sample variance in excess of 50 is about 3%.
It is also possible to determine the sample variance that would be exceeded
only a stated percentage of the time. For example, we might want to be alerted
when the sample variance exceeded a value that should occur only once in 100
times. Then we set up the 	2 equation, find the critical value from Table 4 in the
Appendix, and solve for the sample variance. Using the same values as above,
the value of s2 that would be exceeded only once in 100 times is found as follows:
2 ?
25 ? 21:7 ) s2 ?
21:7  25
9 ? 60:278
In other words, the variance of samples of size 10, taken from this process,
should be less than 60.278, 99% of the time.
Example of chi-squared probability calculations using
Microsoft Excel
Microsoft Excel has a built-in capability to calculate chi-squared probabilities.
To solve the above problem using Excel, enter the n and x values as
shown in Figure 9.19. Note that Excel uses degrees of freedom rather than the
sample size in its calculations; degrees of freedom is the sample size minus
one, as shown in the Deg___freedom box in Figure 9.19. The formula result near
the bottom of the screen gives the desired probability.
Overview of statistical methods 307
Figure 9.18. 	2 pdf for v ? 4.
Example of inverse chi-squared probability calculations using
Microsoft Excel
Microsoft Excel has a built-in capability to calculate chi-squared probabilities,
making it unnecessary to look up the probabilities in tables. To find the critical
chi-squared value for the above problem using Excel, use the CHIINV function
and enter the desired probability and degrees of freedom as shown in Figure 9.20.
The formula result near the bottom of the screen gives the desired critical value.
Figure 9.19. Example of ?nding chi-squared probability using Microsoft Excel.
Figure 9.20. Example of ?nding inverse chi-squared probability using Microsoft Excel.
The t statistic is commonly used to test hypotheses regarding means, regression
coefficients and a wide variety of other statistics used in quality engineering.
‘‘Student’’ was the pseudonym of W.S. Gosset, whose need to
quantify the results of small scale experiments motivated him to develop and
tabulate the probability integral of the ratio which is now known as the t
statistic and is shown in Equation 9.16.
t ?
s= ffiffiffi n p ?9:16?
In Equation 9.16, the denominator is the standard deviation of the sample
mean. Percentage points of the corresponding distribution function of t may
be found in Table 3 in the Appendix. There is a t distribution for each sample
size of n > 1. As the sample size increases, the t distribution approaches the
shape of the normal distribution, as shown in Figure 9.21.
One of the simplest (and most common) applications of the Student’s t test
involves using a sample from a normal population with mean m and variance
s2. This is demonstrated in the Hypothesis testing section later in this chapter.
Overview of statistical methods 309
Figure 9.21. Student’s t distributions.
Suppose we have two random samples drawn from a normal population. Let
be the variance of the first sample and s22
be the variance of the second sample.
The two samples need not have the same sample size. The statistic F given by
F ?
has a sampling distribution called the F distribution. There are two sample variances
involved and two sets of degrees of freedom, n11 in the numerator and
n21 in the denominator. The Appendix includes tables for 1% and 5% percentage
points for the F distribution. The percentages refer to the areas to the
right of the values given in the tables. Figure 9.22 illustrates two F distributions.
Figure 9.22. F distributions.
Statistical inference
All statements made in this section are valid only for stable processes, i.e.,
processes in statistical control. The statistical methods described in this section
are enumerative. Although most applications of Six Sigma are analytic, there
are times when enumerative statistics prove useful. In reading this material,
the analyst should keep in mind the fact that analytic methods should also be
used to identify the underlying process dynamics and to control and improve
the processes involved. The subject of statistical inference is large and it is covered
in many different books on introductory statistics. In this book we review
that part of the subject matter of particular interest in Six Sigma.
So far, we have introduced a number of important statistics including the
sample mean, the sample standard deviation, and the sample variance. These
sample statistics are called point estimators because they are single values used
to represent population parameters. It is also possible to construct an interval
about the statistics that has a predetermined probability of including the true
population parameter. This interval is called a confidence interval. Interval estimation
is an alternative to point estimation that gives us a better idea of the magnitude
of the sampling error. Confidence intervals can be either one-sided or
two-sided. A one-sided or confidence interval places an upper or lower bound
on the value of a parameter with a specified level of confidence. A two-sided
confidence interval places both upper and lower bounds.
In almost all practical applications of enumerative statistics, including Six
Sigma applications, we make inferences about populations based on data from
samples. In this chapter, we have talked about sample averages and standard
deviations; we have even used these numbers to make statements about future
performance, such as long term yields or potential failures. A problem arises
that is of considerable practical importance: any estimate that is based on a sample
has some amount of sampling error. This is true even though the sample estimates
are the ‘‘best estimates’’ in the sense that they are (usually) unbiased
estimators of the population parameters.
Estimates of the mean
For random samples with replacement, the sampling distribution of XX has a
mean m and a standard deviation equal to = ffiffiffi n p . For large samples the sampling
distribution of XX is approximately normal and normal tables can be used to
find the probability that a sample mean will be within a given distance of m.
Overview of statistical methods 311
For example, in 95% of the samples we will observe a mean within 1:96= ffiffiffi n p
of m. In other words, in 95% of the samples the interval from XX  1:96= ffiffiffi n p to
X ? 1:96= ffiffiffi n p will include m. This interval is called a ‘‘95% confidence interval
for estimating m.’’ It is usually shown using inequality symbols:
X  1:96= ffiffiffi n p <  < XX ? 1:96= ffiffiffi n p
The factor 1.96 is the Z value obtained from the normal Table 2 in the
Appendix. It corresponds to the Z value beyond which 2.5% of the population
lie. Since the normal distribution is symmetric, 2.5% of the distribution lies
above Z and 2.5% below Z. The notation commonly used to denote Z values
for confidence interval construction or hypothesis testing is Za=z where
100?1  a) is the desired confidence level in percent. For example, if we want
95% confidence, a ? 0:05, 100?1  a? ? 95%, and Z0.025?1.96. In hypothesis
testing the value of a is known as the significance level.
Example: estimating m when s is known
Suppose thatsis known to be 2.8. Assume that we collect a sample of n ? 16
and compute XX ? 15:7. Using the above equation we find the 95% confidence
interval for m as follows:
X  1:96= ffiffiffi n p <  < XX ? 1:96= ffiffiffi n p
15:7  1:96?2:8= ffiffiffiffiffi 16 p ? <  < 15:7 ? 1:96?2:8= ffiffiffiffiffi 16 p ?
14:33 <  < 17:07
There is a 95% level of confidence associated with this interval. The numbers
14.33 and 17.07 are sometimes referred to as the confidence limits.
Note that this is a two-sided confidence interval. There is a 2.5% probability
that 17.07 is lower than m and a 2.5% probability that 14.33 is greater than m. If
we were only interested in, say, the probability that m were greater than 14.33,
then the one-sided confidence interval would be m > 14:33 and the one-sided
confidence level would be 97.5%.
Example of using Microsoft Excel to calculate the con?dence
interval for the mean when sigma is known
Microsoft Excel has a built-in capability to calculate confidence intervals for
the mean. The dialog box in Figure 9.23 shows the input. The formula result
near the bottom of the screen gives the interval width as 1.371972758. To find
the lower confidence limit subtract the width from the mean. To find the upper
confidence limit add the width to the mean.
Example: estimating m when s is unknown
When s is not known and we wish to replace s with s in calculating confidence
intervals for m, we must replace Za=2 with ta=2 and obtain the percentiles
from tables for Student’s t distribution instead of the normal tables. Let’s revisit
the example above and assume that instead of knowing s, it was estimated
from the sample, that is, based on the sample of n ? 16, we computed s ? 2:8
and XX ? 15:7. Then the 95% confidence interval becomes:
X ? 2:131s= ffiffiffi n p <  < XX ? 2:131s= ffiffiffi n p
15:7  2:131?2:8= ffiffiffiffiffi 16 p ? <  < 15:7 ? 2:131?2:8= ffiffiffiffiffi 16 p ?
14:21 <  < 17:19
It can be seen that this interval is wider than the one obtained for known s.
The ta=2 value found for 15 df is 2.131 (see Table 3 in the Appendix), which is
greater than Za=2 ? 1:96 above.
Overview of statistical methods 313
Figure 9.23. Example of ?nding the con?dence interval when sigma is known using
Microsoft Excel.
Example of using Microsoft Excel to calculate the con?dence
interval for the mean when sigma is unknown
Microsoft Excel has no built-in capability to calculate confidence intervals
for the mean when sigma is not known. However, it does have the ability to calculate
t-values when given probabilities and degrees of freedom. This information
can be entered into an equation and used to find the desired confidence
limits. Figure 9.24 illustrates the approach. The formula bar shows the formula
for the 95% upper confidence limit for the mean in cell B7.
Hypothesis testing/Type I and Type II errors
Statistical inference generally involves four steps:
1. Formulating a hypothesis about the population or ‘‘state of nature,’’
2. Collecting a sample of observations from the population,
3. Calculating statistics based on the sample,
4. Either accepting or rejecting the hypothesis based on a predetermined
acceptance criterion.
Figure 9.24. Example of ?nding the con?dence interval when sigma is unknown using
Microsoft Excel.
There are two types of error associated with statistical inference:
Type I error (a error)LThe probability that a hypothesis that is actually
true will be rejected. The value of a is known as the significance level
of the test.
Type II error (b error)LThe probability that a hypothesis that is actually
false will be accepted.
Type II errors are often plotted in what is known as an operating characteristics
Confidence intervals are usually constructed as part of a statistical test of
hypotheses. The hypothesis test is designed to help us make an inference
about the true population value at a desired level of confidence. We will look
at a few examples of how hypothesis testing can be used in Six Sigma applications.
Example: hypothesis test of sample mean
Experiment: The nominal specification for filling a bottle with a test chemical
is 30 cc. The plan is to draw a sample of n?25 units from a stable process
and, using the sample mean and standard deviation, construct a two-sided confidence
interval (an interval that extends on either side of the sample average)
that has a 95% probability of including the true population mean. If the interval
includes 30, conclude that the lot mean is 30, otherwise conclude that the
lot mean is not 30.
Result: A sample of 25 bottles was measured and the following statistics computed
X ? 28 cc
s ? 6 cc
The appropriate test statistic is t, given by the formula
t ?
s= ffiffiffi n p ?
28  30
6= ffiffiffiffiffi 25 p ? 1:67
Table 3 in the Appendix gives values for the t statistic at various degrees of
freedom. There are n  1 degrees of freedom (df). For our example we need
the t.975 column and the row for 24 df. This gives a t value of 2.064. Since the
absolute value of this t value is greater than our test statistic, we fail to reject
the hypothesis that the lot mean is 30 cc. Using statistical notation this is
shown as:
Overview of statistical methods 315
H0:  ? 30 cc (the null hypothesis)
H1:  is not equal to 30 cc (the alternate hypothesis)
a ? :05 (Type I error or level of signi?cance)
Critical region: 2:064  t0  ?2:064
Test statistic: t ? 1:67.
Since t lies inside the critical region, fail to reject H0, and accept the hypothesis
that the lot mean is 30 cc for the data at hand.
Example: hypothesis test of two sample variances
The variance of machine X’s output, based on a sample of n=25 taken from a
stable process, is 100. Machine Y’s variance, based on a sample of 10, is 50. The
manufacturing representative from the supplier of machine X contends that
the result is a mere ‘‘statistical fluke.’’ Assuming that a ‘‘statistical fluke’’ is
something that has less than 1 chance in 100, test the hypothesis that both
variances are actually equal.
The test statistic used to test for equality of two sample variances is the F
statistic, which, for this example, is given by the equation
F ?
50 ? 2, numerator df ? 24; denominator df ? 9
Using Table 5 in the Appendix for F.99 we find that for 24 df in the numerator
and 9 df in the denominator F = 4.73. Based on this we conclude that the manufacturer
of machine X could be right, the result could be a statistical fluke. This
example demonstrates the volatile nature of the sampling error of sample
variances and standard deviations.
Example: hypothesis test of a standard deviation compared to
a standard value
A machine is supposed to produce parts in the range of 0.500 inches plus or
minus 0.006 inches. Based on this, your statistician computes that the absolute
worst standard deviation tolerable is 0.002 inches. In looking over your capability
charts you find that the best machine in the shop has a standard deviation
of 0.0022, based on a sample of 25 units. In discussing the situation with the statistician
and management, it is agreed that the machine will be used if a onesided
95% confidence interval on sigma includes 0.002.
The correct statistic for comparing a sample standard deviation with a standard
value is the chi-square statistic. For our data we have s?0.0022, n?25,
and s0 ? 0:002. The 	2 statistic has n  1 ? 24 degrees of freedom. Thus,
	2 ? ?n  1?s2
2 ?
24  ?0:0022?2
?0:002?2 ? 29:04
Appendix Table 4 gives, in the 0.05 column (since we are constructing a onesided
confidence interval) and the df ? 24 row, the critical value 	2 ? 36:42.
Since our computed value of 	2 is less than 36.42, we use the machine. The
reader should recognize that all of these exercises involved a number of assumptions,
e.g., that we ‘‘know’’ that the best machine has a standard deviation of
0.0022. In reality, this knowledge must be confirmed by a stable control chart.
A number of criticisms have been raised regarding the methods used for estimation
and hypothesis testing:
. They are not intuitive.
. They are based on strong assumptions (e.g., normality) that are often not
met in practice.
. They are di?cult to learn and to apply.
. They are error-prone.
In recent years a new method of performing these analyses has been developed.
It is known as resampling or bootstrapping. The new methods are conceptually
quite simple: using the data from a sample, calculate the statistic of
interest repeatedly and examine the distribution of the statistic. For example,
say you obtained a sample of n ? 25 measurements from a lot and you wished
to determine a confidence interval on the statistic Cpk.* Using resampling, you
would tell the computer to select a sample of n ? 25 from the sample results,
compute Cpk, and repeat the process many times, say 10,000 times. You would
then determine whatever percentage point value you wished by simply looking
at the results. The samples would be taken ‘‘with replacement,’’ i.e., a particular
value from the original sample might appear several times (or not at all) in a
Resampling has many advantages, especially in the era of easily available,
low-cost computer power. Spreadsheets can be programmed to resample and
calculate the statistics of interest. Compared with traditional statistical methods,
resampling is easier for most people to understand. It works without strong
Overview of statistical methods 317
*See Chapter 13.
assumptions, and it is simple. Resampling doesn’t impose as much baggage
between the engineering problem and the statistical result as conventional
methods. It can also be used for more advanced problems, such as modeling,
design of experiments, etc.
For a discussion of the theory behind resampling, see Efron (1982). For a presentation
of numerous examples using a resampling computer program see
Simon (1992).
Terms and concepts
A central concept in statistical process control (SPC) is that every measurable
phenomenon is a statistical distribution. In other words, an observed set
of data constitutes a sample of the effects of unknown common causes. It follows
that, after we have done everything to eliminate special causes of variations,
there will still remain a certain amount of variability exhibiting the state
of control. Figure 9.25 illustrates the relationships between common causes,
special causes, and distributions.
There are three basic properties of a distribution: location, spread, and
shape. The location refers to the typical value of the distribution, such as the
mean. The spread of the distribution is the amount by which smaller values
differ from larger ones. The standard deviation and variance are measures of
distribution spread. The shape of a distribution is its patternLpeakedness,
symmetry, etc. A given phenomenon may have any one of a number of distribution
shapes, e.g., the distribution may be bell-shaped, rectangular-shaped,
Figure 9.25. Distributions.
From Continuing Process Control and Process Capability Improvement, p. 4a. Copyright
#1983 by Ford Motor Company. Used by permission of the publisher.
The central limit theorem can be stated as follows:
Irrespective of the shape of the distribution of the population or universe,
the distribution of average values of samples drawn from that universe
will tend toward a normal distribution as the sample size grows without
It can also be shown that the average of sample averages will equal the average
of the universe and that the standard deviation of the averages equals the standard
deviation of the universe divided by the square root of the sample size.
Shewhart performed experiments that showed that small sample sizes were
needed to get approximately normal distributions from even wildly non-normal
universes. Figure 9.26 was created by Shewhart using samples of four measurements.
Principles of statistical process control 319
Figure 9.26. Illustration of the central limit theorem.
From Economic Control of Quality of Manufactured Product, ?gure 59. Copyright#1931,
1980 by ASQC Quality Press. Used by permission of the publisher.
The practical implications of the central limit theorem are immense.
Consider that without the central limit theorem effects, we would have to
develop a separate statistical model for every non-normal distribution encountered
in practice. This would be the only way to determine if the system were
exhibiting chance variation. Because of the central limit theorem we can use
averages of small samples to evaluate any process using the normal distribution.
The central limit theorem is the basis for the most powerful of statistical process
control tools, Shewhart control charts.
Objectives and benefits
Without SPC, the bases for decisions regarding quality improvement are
based on intuition, after-the-fact product inspection, or seat-of-the-pants ‘‘data
analysis.’’ SPC provides a scientific basis for decisions regarding process
A process control system is essentially a feedback system that links process
outcomes with process inputs. There are four main elements involved, the process
itself, information about the process, action taken on the process, and
action taken on the output from the process. The way these elements fit together
is shown in Figure 9.27.
Figure 9.27. A process control system.
By the process, we mean the whole combination of people, equipment,
input materials, methods, and environment that work together to produce
output. The performance information is obtained, in part, from evaluation
of the process output. The output of a process includes more than product,
it also includes information about the operating state of the process such as
temperature, cycle times, etc. Action taken on a process is future-oriented in
the sense that it will affect output yet to come. Action on the output is pastoriented
because it involves detecting out-of-specification output that has
already been produced.
There has been a tendency in the past to concentrate attention on the detection-
oriented strategy of product inspection. With this approach, we wait until
an output has been produced, then the output is inspected and either accepted
or rejected. SPC takes you in a completely different direction: improvement in
the future. A key concept is the smaller the variation around the target, the better.
Thus, under this school of thought, it is not enough to merely meet the
requirements; continuous improvement is called for even if the requirements
are already being met. The concept of never-ending, continuous improvement
is at the heart of SPC and Six Sigma.
Common and special causes of variation
Shewhart (1931, 1980) defined control as follows:
A phenomenon will be said to be controlled when, through the use of past
experience, we can predict, at least within limits, how the phenomenon
may be expected to vary in the future. Here it is understood that prediction
within limits means that we can state, at least approximately, the
probability that the observed phenomenon will fall within the given limits.
The critical point in this definition is that control is not defined as the complete
absence of variation. Control is simply a state where all variation is predictable
variation. A controlled process isn’t necessarily a sign of good
management, nor is an out-of-control process necessarily producing non-conforming
In all forms of prediction there is an element of risk. For our purposes, we
will call any unknown random cause of variation a chance cause or a common
cause, the terms are synonymous and will be used as such. If the influence of
any particular chance cause is very small, and if the number of chance causes
of variation are very large and relatively constant, we have a situation where
the variation is predictable within limits. You can see from the definition
above, that a system such as this qualifies as a controlled system. Where Dr.
Shewhart used the term chance cause, Dr. W. Edwards Deming coined the
Principles of statistical process control 321
Figure 9.28. Should these variations be left to chance?
FromEconomic Control of Quality of Manufactured Product, p. 13. Copyright#1931, 1980 by
ASQC Quality Press. Used by permission of the publisher.
Figure 9.29. Types of variation.
term common cause to describe the same phenomenon. Both terms are encountered
in practice.
Needless to say, not all phenomena arise from constant systems of common
causes. At times, the variation is caused by a source of variation that is not part
of the constant system. These sources of variation were called assignable causes
by Shewhart, special causes of variation by Deming. Experience indicates that
Principles of statistical process control 323
Figure 9.30. Charts from Figure 9.28 with control limits shown.
From Economic Control of Quality of Manufactured Product, p. 13. Copyright#1931, 1980
by ASQC Quality Press. Used by permission of the publisher.
special causes of variation can usually be found without undue difficulty, leading
to a process that is less variable.
Statistical tools are needed to help us effectively separate the effects of special
causes of variation from chance cause variation. This leads us to another definition:
Statistical process controlLthe use of valid analytical statistical methods to
identify the existence of special causes of variation in a process.
The basic rule of statistical process control is:
Variation from common-cause systems should be left to chance, but
special causes of variation should be identi?ed and eliminated.
This is Shewhart’s original rule. However, the rule should not be misinterpreted
as meaning that variation from common causes should be ignored.
Rather, common-cause variation is explored ‘‘off-line.’’ That is, we look for
long-term process improvements to address common-cause variation.
Figure 9.28 illustrates the need for statistical methods to determine the
category of variation.
The answer to the question ‘‘should these variations be left to chance?’’ can
only be obtained through the use of statistical methods. Figure 9.29 illustrates
the basic concept.
In short, variation between the two ‘‘control limits’’ designated by the dashed
lines will be deemed as variation from the common-cause system. Any variability
beyond these fixed limits will be assumed to have come from special causes
of variation. We will call any system exhibiting only common-cause variation,
‘‘statistically controlled.’’ It must be noted that the control limits are not simply
pulled out of the air, they are calculated from actual process data using valid statistical
methods. Figure 9.28 is shown below as Figure 9.30, only with the control
limits drawn on it; notice that process (a) is exhibiting variations from special
causes, while process (b) is not. This implies that the type of action needed to
reduce the variability in each case is of a different nature. Without statistical guidance
there could be endless debate over whether special or common causes
were to blame for variability.
^ ^ ^
Measurement Systems
Discrimination, stability, bias, repeatability,
reproducibility, and linearity
Modern measurement system analysis goes well beyond calibration. A gage
can be perfectly accurate when checking a standard and still be entirely unacceptable
for measuring a product or controlling a process. This section illustrates
techniques for quantifying discrimination, stability, bias, repeatability,
reproducibility and variation for a measurement system. We also show how to
express measurement error relative to the product tolerance or the process variation.
For the most part, the methods shown here use control charts. Control
charts provide graphical portrayals of the measurement processes that enable
the analyst to detect special causes that numerical methods alone would not
Discrimination, sometimes called resolution, refers to the ability of the
measurement system to divide measurements into ‘‘data categories.’’ All
parts within a particular data category will measure the same. For example,
Copyright 2003 by The McGraw-Hill Companies, Inc. Click Here for Terms of Use.
if a measurement system has a resolution of 0.001 inches, then items measuring
1.0002, 1.0003, 0.9997 would all be placed in the data category 1.000, i.e.,
they would all measure 1.000 inches with this particular measurement system.
A measurement system’s discrimination should enable it to divide the region
of interest into many data categories. In Six Sigma, the region of interest is
the smaller of the tolerance (the high specification minus the low specification)
or six standard deviations. A measurement system should be able to
divide the region of interest into at least five data categories. For example, if
a process was capable (i.e., Six Sigma is less than the tolerance) and
s ? 0:0005, then a gage with a discrimination of 0.0005 would be acceptable
(six data categories), but one with a discrimination of 0.001 would not
(three data categories). When unacceptable discrimination exists, the range
chart shows discrete ‘‘jumps’’ or ‘‘steps.’’ This situation is illustrated in
Figure 10.1.
Figure 10.1. Inadequate gage discrimination on a control chart.
Note that on the control charts shown in Figure 10.1, the data plotted are the
same, except that the data on the bottom two charts were rounded to the nearest
25. The effect is most easily seen on the R chart, which appears highly stratified.
As sometimes happens (but not always), the result is to make the X-bar chart
go out of control, even though the process is in control, as shown by the control
charts with unrounded data. The remedy is to use a measurement system capable
of additional discrimination, i.e., add more significant digits. If this cannot
be done, it is possible to adjust the control limits for the round-off error by
using a more involved method of computing the control limits, see Pyzdek
(1992a, pp. 37^42) for details.
Measurement system stability is the change in bias over time when using a
measurement system to measure a given master part or standard. Statistical stability
is a broader term that refers to the overall consistency of measurements
over time, including variation from all causes, including bias, repeatability,
reproducibility, etc. A system’s statistical stability is determined through the
use of control charts. Averages and range charts are typically plotted on measurements
of a standard or a master part. The standard is measured repeatedly
over a short time, say an hour; then the measurements are repeated at predetermined
intervals, say weekly. Subject matter expertise is needed to determine
the subgroup size, sampling intervals and measurement procedures to be followed.
Control charts are then constructed and evaluated.A(statistically) stable
system will show no out-of-control signals on an X-control chart of the averages’
readings. No ‘‘stability number’’ is calculated for statistical stability; the system
either is or is not statistically stable.
Once statistical stability has been achieved, but not before, measurement system
stability can be determined. One measure is the process standard deviation
based on the R or s chart.
R chart method:
^  ?
s chart method:
^  ?
The values d2 and c4 are constants from Table 11 in the Appendix.
R&R studies for continuous data 327
Bias is the difference between an observed average measurement result and a
reference value. Estimating bias involves identifying a standard to represent
the reference value, then obtaining multiple measurements on the standard.
The standard might be amaster part whose value has been determined by ameasurement
system with much less error than the system under study, or by a standard
traceable to NIST. Since parts and processes vary over a range, bias is
measured at a point within the range. If the gage is non-linear, bias will not be
the same at each point in the range (see the definition of linearity above).
Bias can be determined by selecting a single appraiser and a single reference
part or standard. The appraiser then obtains a number of repeated measurements
on the reference part. Bias is then estimated as the difference between
the average of the repeated measurement and the known value of the reference
part or standard.
Example of computing bias
A standard with a known value of 25.4 mm is checked 10 times by one
mechanical inspector using a dial caliper with a resolution of 0.025 mm. The
readings obtained are:
25.425 25.425 25.400 25.400 25.375
25.400 25.425 25.400 25.425 25.375
The average is found by adding the 10 measurements together and dividing by
X ?
10 ? 25:4051 mm
The bias is the average minus the reference value, i.e.,
bias ? average  reference value
? 25:4051 mm  25:400 mm ? 0:0051 mm
The bias of the measurement system can be stated as a percentage of the tolerance
or as a percentage of the process variation. For example, if this measurement
system were to be used on a process with a tolerance of  0.25 mm
% bias ? 100  jbiasj=tolerance
? 100  0:0051=0:5 ? 1%
This is interpreted as follows: this measurement system will, on average, produce
results that are 0.0051 mm larger than the actual value. This difference
represents 1% of the allowable product variation. The situation is illustrated in
Figure 10.2.
Ameasurement system is repeatable if its variability is consistent. Consistent
variability is operationalized by constructing a range or sigma chart based on
repeated measurements of parts that cover a significant portion of the process
variation or the tolerance, whichever is greater. If the range or sigma chart is
out of control, then special causes are making the measurement system inconsistent.
If the range or sigma chart is in control then repeatability can be estimated
by finding the standard deviation based on either the average range or the average
standard deviation. The equations used to estimate sigma are shown in
Chapter 9.
Example of estimating repeatability
The data in Table 10.1 are from a measurement study involving two inspectors.
Each inspector checked the surface finish of five parts, each part was
checked twice by each inspector. The gage records the surface roughness in minches
(micro-inches). The gage has a resolution of 0.1 m-inches.
R&R studies for continuous data 329
Figure 10.2. Bias example illustrated.
We compute:
Ranges chart
? 0:51
UCL ? D4 RR ? 3:267  0:51 ? 1:67
Averages chart
X ? 118:85
LCL ? XX  A2 RR ? 118:85  1:88  0:109 ? 118:65
UCL ? XX ? A2 RR ? 118:85 ? 1:88  0:109 ? 119:05
Table 10.1. Measurement system repeatability study data.
1 111.9 112.3 112.10 0.4
2 108.1 108.1 108.10 0.0
3 124.9 124.6 124.75 0.3
4 118.6 118.7 118.65 0.1
5 130.0 130.7 130.35 0.7
1 111.4 112.9 112.15 1.5
2 107.7 108.4 108.05 0.7
3 124.6 124.2 124.40 0.4
4 120.0 119.3 119.65 0.7
5 130.4 130.1 130.25 0.3
The data and control limits are displayed in Figure 10.3. The R chart analysis
shows that all of the R values are less than the upper control limit. This indicates
that the measurement system’s variability is consistent, i.e., there are no special
causes of variation.
Note that many of the averages are outside of the control limits. This is the
way it should be! Consider that the spread of the X-bar chart’s control limits is
based on the average range, which is based on the repeatability error. If the
averages were within the control limits it wouldmean that the part-to-part variation
was less than the variation due to gage repeatability error, an undesirable
situation. Because the R chart is in control we can now estimate the standard
deviation for repeatability or gage variation:
e ?
d2 ?10:1?
where d2 is obtained from Table 13 in the Appendix. Note that we are using d2
and not d2. The d2 values are adjusted for the small number of subgroups typically
involved in gage R&R studies. Table 13 is indexed by two values: m is the
number of repeat readings taken (m ? 2 for the example), and g is the number
of parts times the number of inspectors (g ? 5  2 ? 10 for the example).
This gives, for our example
e ?
d2 ?
1:16 ? 0:44
R&R studies for continuous data 331
Figure 10.3. Repeatability control charts.
The repeatability from this study is calculated by 5:15e ? 5:15 0:44 ? 2:26. The value 5.15 is the Z ordinate which includes 99% of a standard
normal distribution.
A measurement system is reproducible when different appraisers produce
consistent results. Appraiser-to-appraiser variation represents a bias due to
appraisers. The appraiser bias, or reproducibility, can be estimated by comparing
each appraiser’s average with that of the other appraisers. The standard
deviation of reproducibility (o) is estimated by finding the range between
appraisers (Ro) and dividing by d2 . Reproducibility is then computed as 5.15o.
Reproducibility example (AIAGmethod)
Using the data shown in the previous example, each inspector’s average is
computed and we find:
Inspector #1 average ? 118:79-inches
Inspector #2 average ? 118:90-inches
Range ? Ro ? 0:11-inches
Looking in Table 13 in the Appendix for one subgroup of two appraisers we
find d2 ? 1:41 ?m ? 2, g ? 1), since there is only one range calculation g ? 1.
Using these results we find Ro=d2 ? 0:11=1:41 ? 0:078.
This estimate involves averaging the results for each inspector over all of the
readings for that inspector. However, since each inspector checked each part
repeatedly, this reproducibility estimate includes variation due to repeatability
error. The reproducibility estimate can be adjusted using the following equation:
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi 5:15
d2  2
nr s ? ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffififfiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi 5:15 
1:41  2
?5:15  0:44?2
5  2 s
? ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi 0:16  0:51 p ? 0
As sometimes happens, the estimated variance from reproducibility exceeds
the estimated variance of repeatability + reproducibility. When this occurs the
estimated reproducibility is set equal to zero, since negative variances are theoretically
impossible. Thus, we estimate that the reproducibility is zero.
The measurement system standard deviation is
m ? ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi 2
e ? 2
o p ? ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi ?0:44?2 ? 0 q ? 0:44 ?10:2?
and the measurement system variation, or gage R&R, is 5.15m. For our data
gage R&R ?5:15  0:44 ? 2:27.
Reproducibility example (alternative method)
One problem with the above method of evaluating reproducibility error is
that it does not produce a control chart to assist the analyst with the evaluation.
The method presented here does this. This method begins by rearranging the
data in Table 10.1 so that all readings for any given part become a single row.
This is shown in Table 10.2.
Observe that when the data are arranged in this way, theRvalue measures the
combined range of repeat readings plus appraisers. For example, the smallest
reading for part #3 was from inspector #2 (124.2) and the largest was from
inspector #1 (124.9). Thus, R represents two sources of measurement error:
repeatability and reproducibility.
R&R studies for continuous data 333
Table 10.2. Measurement error data for reproducibility evaluation.
Part Reading 1 Reading 2 Reading 1 Reading 2 X bar R
1 111.9 112.3 111.4 112.9 112.125 1.5
2 108.1 108.1 107.7 108.4 108.075 0.7
3 124.9 124.6 124.6 124.2 124.575 0.7
4 118.6 118.7 120 119.3 119.15 1.4
5 130 130.7 130.4 130.1 130.3 0.7
Averages! 118.845 1
The control limits are calculated as follows:
Ranges chart
R ? 1:00
UCL ? D4 RR ? 2:282  1:00 ? 2:282
Note that the subgroup size is 4.
Averages chart
X ? 118:85
LCL ? XX  A2 RR ? 118:85  0:729  1 ? 118:12
UCL ? XX ? A2 RR ? 118:85 ? 0:729  1 ? 119:58
The data and control limits are displayed in Figure 10.4. The R chart analysis
shows that all of the R values are less than the upper control limit. This indicates
that the measurement system’s variability due to the combination of repeatability
and reproducibility is consistent, i.e., there are no special causes of variation.
Using this method, we can also estimate the standard deviation of reproducibility
plus repeatability, as we can find o ? Ro=d2 ? 1=2:08 ? 0:48.
Now we know that variances are additive, so
repeatability?reproducibility ? 2
repeatability ? 2
reproducibility ?10:3?
Figure 10.4. Reproducibility control charts.
which implies that
reproducibility ? ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffififfiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi 2
repeatability?reproducibility  2
repeatability q
In a previous example, we computed repeatability ? 0:44. Substituting these
values gives
reproducibility ? ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffififfiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi 2
repeatability?reproducibility  2
repeatability q
? ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi ?0:48?2  ?0:44?2 q ? 0:19
Using this we estimate reproducibility as 5:15  0:19 ? 1:00.
The X-bar charts show the part-to-part variation. To repeat, if the measurement
system is adequate, most of the parts will fall outside of the X -bar chart control
limits. If fewer than half of the parts are beyond the control limits, then
the measurement system is not capable of detecting normal part-to-part variation
for this process.
Part-to-part variation can be estimated once the measurement process is
shown to have adequate discrimination and to be stable, accurate, linear (see
below), and consistent with respect to repeatability and reproducibility. If the
part-to-part standard deviation is to be estimated from the measurement system
study data, the following procedures are followed:
1. Plot the average for each part (across all appraisers) on an averages control
chart, as shown in the reproducibility error alternate method.
2. Con?rm that at least 50% of the averages fall outside the control limits. If
not, ?nd a better measurement system for this process.
3. Find the range of the part averages, Rp.
4. Compute p ? Rp=d2 , the part-to-part standard deviation. The value of
d2 is found in Table 13 in the Appendix using m ? the number of parts
and g ? 1, since there is only one R calculation.
5. The 99% spread due to part-to-part variation (PV) is found as 5.15p.
Once the above calculations have been made, the overall measurement system
can be evaluated.
1. The total process standard deviation is found as t ? ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi 2m
? 2
p q . Where
m ? the standard deviation due to measurement error.
2. Total variability (TV) is 5.15t.
3. The percent repeatability and reproducibility (R&R) is 100  ?m=t?%.
R&R studies for continuous data 335
4. The number of distinct data categories that can be created with this measurement
system is 1.41 (PV/R&R).
1. Plot the average for each part (across all appraisers) on an averages control
chart, as shown in the reproducibility error alternate method.
Done above, see Figure 10.3.
2. Con?rm that at least 50% of the averages fall outside the control limits. If
not, ?nd a better measurement system for this process.
4 of the 5 part averages, or 80%, are outside of the control limits. Thus,
the measurement system error is acceptable.
3. Find the range of the part averages, Rp.
Rp ? 130:3  108:075 ? 22:23.
4. Compute p ? Rp=d2 , the part-to-part standard deviation. The value of
d2 is found in Table 13 in the Appendix using m ? the number of parts
and g ? 1, since there is only one R calculation.
m ? 5, g ? 1, d2 ? 2:48, p ? 22:23=2:48 ? 8:96.
5. The 99% spread due to part-to-part variation (PV) is found as 5.15p.
5:15  8:96 ? PV ? 46:15.
Once the above calculations have been made, the overall measurement system
can be evaluated.
1. The total process standard deviation is found as t ? ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi 2m
? 2
p q
t ? ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi 2m
? 2
p q ? ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi ?0:44?2 ? ?8:96?2 q ? ffiffiffiffiffiffiffiffi 80:5 p ? 8:97
2. Total variability (TV) is 5.15t.
5:15  8:97 ? 46:20
3. The percent R&R is 100  ?m=t?%
% ? 100
8:97 ? 4:91%
4. The number of distinct data categories that can be created with this measurement
system is 1:41  ?PV=R&R?.
2:27 ? 28:67 ? 28
Since the minimum number of categories is five, the analysis indicates that
this measurement system is more than adequate for process analysis or process
Gage R&R analysis using Minitab
Minitab has a built-in capability to perform gage repeatability and reproducibility
studies. To illustrate these capabilities, the previous analysis will be
repeated using Minitab. To begin, the data must be rearranged into the format
expected by Minitab (Figure 10.5). For reference purposes, columns C1^C4
contain the data in our original format and columns C5^C8 contain the same
data in Minitab’s preferred format.
Minitab offers two different methods for performing gage R&R studies:
crossed and nested. Use gage R&R nested when each part can be measured by
only one operator, as with destructive testing. Otherwise, choose gage R&R
crossed. To do this, select Stat > Quality Tools > Gage R&R Study (Crossed)
to reach the Minitab dialog box for our analysis (Figure 10.6). In addition to
choosing whether the study is crossed or nested, Minitab also offers both the
R&R studies for continuous data 337
Figure 10.5. Data formatted for Minitab input.
ANOVA and the X-bar and R methods. You must choose the ANOVA option
to obtain a breakdown of reproducibility by operator and operator by part. If
the ANOVA method is selected, Minitab still displays the X-bar and R charts
so you won’t lose the information contained in the graphics. We will use
ANOVA in this example. Note that the results of the calculations will differ
slightly from those we obtained using the X-bar and R methods.
There is an option in gage R&R to include the process tolerance. This will
provide comparisons of gage variation with respect to the specifications in addition
to the variability with respect to process variation. This is useful information
if the gage is to be used to make product acceptance decisions. If the
process is ‘‘capable’’ in the sense that the total variability is less than the tolerance,
then any gage that meets the criteria for checking the process can also be
used for product acceptance. However, if the process is not capable, then its output
will need to be sorted and the gage used for sorting may need more discriminatory
power than the gage used for process control. For example, a gage
capable of 5 distinct data categories for the process may have 4 or fewer for the
product. For the purposes of illustration, we entered a value of 40 in the process
tolerance box in the Minitab options dialog box (Figure 10.7).
Minitab produces copious output, including six separate graphs, multiple
tables, etc. Much of the output is identical to what has been discussed earlier in
this chapter and won’t be shown here.
Figure 10.6. Minitab gage R&R (crossed) dialog box.
Table 10.3 shows the analysis of variance for the R&R study. In the ANOVA
the MS for repeatability (0.212) is used as the denominator or error term for calculating
the F-ratio of the Operator*PartNum interaction; 0.269/0.212 = 1.27.
The F-ratio for the Operator effect is found by using the Operator*PartNum
interaction MS term as the denominator, 0.061/0.269 = 0.22. The F-ratios are
used to compute the P values, which show the probability that the observed variation
for the source row might be due to chance. By convention, a P value less
than 0.05 is the critical value for deciding that a source of variation is ‘‘signifi-
R&R studies for continuous data 339
Figure 10.7. Minitab gage R&R (crossed) options dialog box.
Table 10.3. Two-way ANOVA table with interaction.
Source DF SS MS F P
PartNum 4 1301.18 325.294 1208.15 0
Operator 1 0.06 0.061 0.22 0.6602
Operator*PartNum 4 1.08 0.269 1.27 0.34317
Repeatability 10 2.12 0.212
Total 19 1304.43
cant,’’ i.e., greater than zero. For example, the P value for the PartNum row is 0,
indicating that the part-to-part variation is almost certainly not zero. The P
values for Operator (0.66) and the Operator*PartNum interaction (0.34) are
greater than 0.05 so we conclude that the differences accounted for by these
sources might be zero. If the Operator term was significant (P < 0.05) we
would conclude that there were statistically significant differences between
operators, prompting an investigation into underlying causes. If the interaction
term was significant, we would conclude that one operator has obtained different
results with some, but not all, parts.
Minitab’s next output is shown in Table 10.4. This analysis has removed the
interaction term from the model, thereby gaining 4 degrees of freedom for the
error term and making the test more sensitive. In some cases this might identify
a significant effect that was missed by the larger model, but for this example
the conclusions are unchanged.
Minitab also decomposes the total variance into components, as shown in
Table 10.5. The VarComp column shows the variance attributed to each source,
while the % of VarComp shows the percentage of the total variance accounted
for by each source. The analysis indicates that nearly all of the variation is
between parts.
The variance analysis shown in Table 10.5, while accurate, is not in original
units. (Variances are the squares of measurements.) Technically, this is the correct
way to analyze information on dispersion because variances are additive,
while dispersion measurements expressed in original units are not. However,
there is a natural interest in seeing an analysis of dispersion in the original
units so Minitab provides this. Table 10.6 shows the spread attributable to the
Table 10.4. Two-way ANOVA table without interaction.
Source DF SS MS F P
PartNum 4 1301.18 325.294 1426.73 0
Operator 1 0.06 0.061 0.27 0.6145
Repeatability 14 3.19 0.228
Total 19 1304.43
different sources. The StdDev column is the standard deviation, or the square
root of the VarComp column in Table 10.5. The Study Var column shows the
99% confidence interval using the StdDev. The % Study Var column is the
Study Var column divided by the total variation due to all sources. And the %
Tolerance is the Study Var column divided by the tolerance. It is interesting
that the % Tolerance column total is greater than 100%. This indicates that the
measured process spread exceeds the tolerance. Although this isn’t a process
capability analysis, the data do indicate a possible problem meeting tolerances.
The information in Table 10.6 is presented graphically in Figure 10.8.
Linearity can be determined by choosing parts or standards that cover all or
most of the operating range of the measurement instrument. Bias is determined
at each point in the range and a linear regression analysis is performed.
Linearity is defined as the slope times the process variance or the slope times
the tolerance, whichever is greater. A scatter diagram should also be plotted
from the data.
The following example is taken from Measurement Systems Analysis, published
by the Automotive Industry Action Group.
R&R studies for continuous data 341
Table 10.5. Components of variance analysis.
Source VarComp % of VarComp
Total gage R&R 0.228 0.28
Repeatability 0.228 0.28
Reproducibility 0 0
Operator 0 0
Part-to-Part 81.267 99.72
Total Variation 81.495 100
Table 10.6. Analysis of spreads.
Source StdDev
% StudyVar
Total gage R&R 0.47749 2.4591 5.29 6.15
Repeatability 0.47749 2.4591 5.29 6.15
Reproducibility 0 0 0 0
Operator 0 0 0 0
Part-to-Part 9.0148 46.4262 99.86 116.07
Total Variation 9.02743 46.4913 100 116.23
Figure 10.8. Graphical analysis of components of variation.
A plant foreman was interested in determining the linearity of a measurement
system. Five parts were chosen throughout the operating range of the measurement
system based upon the process variation. Each part was measured by
a layout inspection to determine its reference value. Each part was then measured
twelve times by a single appraiser. The parts were selected at random.
The part average and bias average were calculated for each part as shown in
Figure 10.9. The part bias was calculated by subtracting the part reference
value from the part average.
A linear regression analysis was performed. In the regression, x is the reference
value and y is the bias. The results are shown in Figure 10.10.
R&R studies for continuous data 343
Figure 10.9. Gage data summary.
Figure 10.10. Regression analysis of linearity summary data.
The P-values indicate that the result is statistically significant, i.e., there is
actually a bias in the gage. The slope of the line is 0.132, and the intercept is
0.74. R2 ? 0:98, indicating that the straight line explains about 98% of the variation
in the bias readings. The results can be summarized as follows:
Bias b ? ax ? 0:74  0:132 (Reference Value)
Linearity jslopej Process Variation ? 0:132  6 ? 0:79, where 6 is the
% Linearity 100%  jslopej ? 13:2%
Note that the zero bias point is found at
x ? 
slope  ? 
0:132  ? 5:61
In this case, this is the point of least bias. Greater bias exists as you move
further from this value.
This information is summarized graphically in Figure 10.11.
Figure 10.11. Graphical analysis of linearity.
Minitab has a built-in capability to perform gage linearity analysis. Figure
10.12 shows the data layout and dialog box. Figure 10.13 shows the Minitab output.
Note that Minitab doesn’t show the P-values for the analysis so it is necessary
to perform a supplementary regression analysis anyway to determine the statistical
significance of the results. For this example, it is obvious from the scatter
plot that the slope of the line isn’t zero, so a P-value isn’t required to conclude
that non-linearity exists. The results aren’t so clear for bias, which is only
0.867%. In fact, if we perform a one-sample t test of the hypothesis that the
mean bias is 0, we get the results shown in Figure 10.14, which indicate the bias
could be 0 (P = 0.797).*
R&R studies for continuous data 345
Figure 10.12. Minitab gage linearity dialog box.
*A problem with this analysis is that the datum for each part is an average of twelve measurements, not individual measurements.
If we could obtain the 60 actual measurements the P-value would probably be different because the standard error
would be based on 60 measurements rather than five. On the other hand, the individual measurements would also be more
variable, so the exact magnitude of the difference is impossible to determine without the raw data.
Attribute data consist of classifications rather than measurements.
Attribute inspection involves determining the classification of an item, e.g., is
it ‘‘good’’ or ‘‘bad’’? The principles of good measurement for attribute inspection
are the same as for measurement inspection (Table 10.7). Thus, it is possible
to evaluate attribute measurement systems in much the same way as we
Figure 10.13. Minitab gage linearity output.
Figure 10.14. One-sample t-test of bias.
evaluate variable measurement systems. Much less work has been done on
evaluating attribute measurement systems. The proposals provided in this
book are those I’ve found to be useful for my employers and clients. The
ideas are not part of any standard and you are encouraged to think about
them critically before adopting them. I also include an example of Minitab’s
attribute gage R&R analysis.
Attribute measurement error analysis 347
Table 10.7. Attribute measurement concepts.
Interpretation for
Attribute Data Suggested Metrics and Comments
Accuracy Items are correctly
Number of times correctly classified by all
Total number of evaluations by all
Requires knowledge of the ‘‘true’’ value.
Bias The proportion of
items in a given
category is correct.
Overall average proportion in a given category (for all
inspectors) minus correct proportion in a given
category. Averaged over all categories.
Requires knowledge of the ‘‘true’’ value.
Repeatability When an inspector
evaluates the same
item multiple
times in a short
time interval, she
assigns it to the
same category
every time.
For a given inspector:
Total number of times repeat classifications agree
Total number of repeat classifications
Overall: Average of repeatabilities
Reproducibility When all
inspectors evaluate
the same item,
they all assign it to
the same category.
Total number of times classifications for all concur
Total number of classifications
Continued on next page . . .
Operational definitions
An operational definition is defined as a requirement that includes a means
of measurement. ‘‘High quality solder’’ is a requirement that must be operationalized
by a clear definition of what ‘‘high quality solder’’ means. This might
include verbal descriptions, magnification power, photographs, physical comparison
specimens, and many more criteria.
1. Operational de?nition of the Ozone Transport Assessment Group’s
(OTAG) goal
Goal: To identify reductions and recommend transported ozone
and its precursors which, in combination with other measures, will
enable attainment and maintenance of the ozone standard in the
OTAG region.
Interpretation for
Attribute Data Suggested Metrics and Comments
Stability The variability
between attribute
R&R studies at
di?erent times.
‘‘Linearity’’ When an inspector
evaluates items
covering the full
set of categories,
her classi?cations
are consistent
across the
Range of inaccuracy and bias across all categories.
Requires knowledge of the ‘‘true’’ value.
Note: Because there is no natural ordering for
nominal data, the concept of linearity doesn’t really
have a precise analog for attribute data on this scale.
However, the suggested metrics will highlight
interactions between inspectors and speci?c categories.
Metric Stability Measure for Metric
Repeatability Standard deviation of
Reproducibility Standard deviation of
Accuracy Standard deviation of accuracies
Bias Average bias
Table 10.7 (cont.)
Suggested operational de?nition of the goal:
1. A general modeled reduction in ozone and ozone precursors
aloft throughout the OTAG region; and
2. A reduction of ozone and ozone precursors both aloft and at
ground level at the boundaries of non-attainment area modeling
domains in the OTAG region; and
3. A minimization of increases in peak ground level ozone concentrations
in the OTAG region. (This component of the operational
de?nition is in review.)
2. Wellesley College Child Care Policy Research Partnership operational
de?nition of unmet need
1. Standard of comparison to judge the adequacy of neighborhood services:
the median availability of services in the larger region
(Hampden County).
2. Thus, our de?nition of unmet need: The di?erence between the care
available in the neighborhood and the median level of care in the surrounding
region (stated in terms of child care slots indexed to the
age-appropriate child populationL‘‘slots-per-tots’’).
3. Operational de?nitions of acids and bases
1. An acid is any substance that increases the concentration of theH+
ion when it dissolves in water.
2. A base is any substance that increases the concentration of theOH^
ion when it dissolves in water.
4. Operational de?nition of ‘‘intelligence’’
1. Administer the Stanford-Binet IQ test to a person and score the
result. The person’s intelligence is the score on the test.
5. Operational de?nition of ‘‘darkblue carpet’’
A carpet will be deemed to be dark blue if
1. Judged by an inspector medically certi?ed as having passed the U.S.
Air Force test for color-blindness
1.1. It matches thePANTONEcolor card 7462Cwhen both carpet
and card are illuminated byGE‘‘cool white’’ ?uorescent tubes;
1.2. Card and carpet are viewed at a distance between 16 inches and
24 inches.
Some commonly used approaches to attribute inspection analysis are shown
in Table 10.8.
Attribute measurement error analysis 349
Table 10.8. Methods of evaluating attribute inspection.
True Value Method of Evaluation Comments
Expert Judgment: An
expert looks at the
classi?cations after the
operator makes normal
classi?cations and decides
which are correct and
which are incorrect.
& Metrics:
Percent correct
& Quanti?es the accuracy of the
& Simple to evaluate.
& Who says the expert is correct?
& Care must be taken to include all types of
& Di?cult to compare operators since
di?erent units are classi?ed by di?erent
& Acceptable level of performance must be
decided upon. Consider cost, impact on
customers, etc.
Round Robin Study: A set
of carefully identi?ed
objects is chosen to
represent the full range of
1. Each item is evaluated
by an expert and its
condition recorded.
2. Each item is evaluated
by every inspector at
least twice.
& Metrics:
1. Percent correct by inspector
2. Inspector repeatability
3. Inspector reproducibility
4. Stability
5. Inspector ‘‘linearity’’
& Full range of attributes included.
& All aspects of measurement error
& People know they’re being watched, may
a?ect performance.
& Not routine conditions.
& Special care must be taken to insure rigor.
& Acceptable level of performance must be
decided upon for each type of error.
Consider cost, impact on customers, etc.
Continued on next page . . .
Example of attribute inspection error analysis
Two sheets with identical lithographed patterns are to be inspected under
carefully controlled conditions by each of the three inspectors. Each sheet has
been carefully examined multiple times by journeymen lithographers and they
have determined that one of the sheets should be classified as acceptable, the
other as unacceptable. The inspectors sit on a stool at a large table where the
sheet will be mounted for inspection. The inspector can adjust the height of
the stool and the angle of the table. A lighted magnifying glass is mounted to
the table with an adjustable arm that lets the inspector move it to any part of
the sheet (see Figure 10.15).
Attribute measurement error analysis 351
True Value Method of Evaluation Comments
Inspector Concurrence
Study: A set of carefully
identi?ed objects is
chosen to represent the
full range of attributes, to
the extent possible.
1. Each item is evaluated
by every inspector at
least twice.
& Metrics:
1. Inspector repeatability
2. Inspector reproducibility
3. Stability
4. Inspector ‘‘linearity’’
& Like a round robin, except true value isn’t
& No measures of accuracy or bias are
possible. Can only measure agreement
between equally quali?ed people.
& Full range of attributes included.
& People know they’re being watched, may
a?ect performance.
& Not routine conditions.
& Special care must be taken to insure rigor.
& Acceptable level of performance must be
decided upon for each type of error.
Consider cost, impact on customers, etc.
Table 10.8. (cont.)
Each inspector checks each sheet once in the morning and again in the afternoon.
After each inspection, the inspector classifies the sheet as either acceptable
or unacceptable. The entire study is repeated the following week. The
results are shown in Table 10.9.
Figure 10.15. Lithography inspection station table, stool and magnifying glass.
Table 10.9. Results of lithography attribute inspection study.
1 Part Standard InspA InspB InspC Date T|me Reproducible Accurate
2 1 1 1 1 1 Today Morning 1 1
3 1 1 0 1 1 Today Afternoon 0 0
4 2 0 0 0 0 Today Morning 1 0
5 2 0 0 0 1 Today Afternoon 0 0
6 1 1 1 1 1 LastWeek Morning 1 1
7 1 1 1 1 0 LastWeek Afternoon 0 0
8 2 0 0 0 1 LastWeek Morning 0 0
9 2 0 0 0 0 LastWeek Afternoon 1 0
In Table 10.9 the Part column identifies which sheet is being inspected, and
the Standard column is the classification for the sheet based on the journeymen’s
evaluations. A 1 indicates that the sheet is acceptable, a 0 that it is unacceptable.
The columns labeled InspA, InspB, and InspC show the
classifications assigned by the three inspectors respectively. The Reproducible
column is a 1 if all three inspectors agree on the classification, whether their classification
agrees with the standard or not. The Accurate column is a 1 if all
three inspectors classify the sheet correctly as shown in the Standard column.
Individual inspector accuracy is determined by comparing each inspector’s
classification with the Standard. For example, in cell C2 of Table 10.9
Inspector A classified the unit as acceptable, and the standard column in the
same row indicates that the classification is correct. However, in cell C3 the
unit is classified as unacceptable when it actually is acceptable. Continuing
this evaluation shows that Inspector A made the correct assessment 7 out of 8
times, for an accuracy of 0.875 or 87.5%. The results for all inspectors are given
in Table 10.10.
Repeatability and pairwise reproducibility
Repeatability is defined in Table 10.7 as the same inspector getting the same
result when evaluating the same item more than once within a short time interval.
Looking at InspA we see that when she evaluated Part 1 in the morning of
‘‘Today’’ she classified it as acceptable (1), but in the afternoon she said it was
unacceptable (0). The other three morning/afternoon classifications matched
each other. Thus, her repeatability is 3/4 or 75%.
Pairwise reproducibility is the comparison of each inspector with every other
inspector when checking the same part at the same time on the same day. For
example, on Part 1/Morning/Today, InspA’s classification matched that of
InspB. However, for Part 1/Afternoon/Today InspA’s classification was differ-
Attribute measurement error analysis 353
Table 10.10. Inspector accuracies.
Inspector A B C
Accuracy 87.5% 100.0% 62.5%
ent than that of InspB. There are eight such comparisons for each pair of inspectors.
Looking at InspA versus InspB we see that they agreed 7 of the 8 times,
for a pairwise repeatability of 7/8 = 0.875.
In Table 10.11 the diagonal values are the repeatability scores and the offdiagonal
elements are the pairwise reproducibility scores. The results are
shown for ‘‘Today’’, ‘‘Last Week’’ and both combined.
Information is always lost when summary statistics are used, but the data
reduction often makes the tradeoff worthwhile. The calculations for the overall
statistics are operationally defined as follows:
& Repeatability is the average of the repeatability scores for the two days
combined; i.e., (0:75 ? 1:00 ? 0:25?=3 ? 0:67.
& Reproducibility is the average of the reproducibility scores for the two
days combined (see Table 10.9); i.e.,
1 ? 0 ? 1 ? 0
4 ?
1 ? 0 ? 0 ? 1
4  2 ? 0:50
& Accuracy is the average of the accuracy scores for the two days combined
(see Table 10.9); i.e.,
1 ? 0 ? 0 ? 0
4 ?
1 ? 0 ? 0 ? 0
4  2 ? 0:25:
Table 10.11. Repeatability and pairwise reproducibility for both days combined.
Overall Today Last Week
A 0.75 0.88 0.50
B 1.00 0.50
C 0.25
A 0.50 0.75 0.50
B 1.00 0.75
C 0.50
A 1.00 1.00 0.50
B 1.00 0.50
C 0.00
& Bias is the estimated proportion in a category minus the true proportion in
the category. In this example the true percent defective is 50% (1 part in
2). Of the twenty-four evaluations, twelve evaluations classi?ed the item
as defective. Thus, the bias is 0:5  0:5 ? 0.
Stability is calculated for each of the above metrics separately, as shown in
Table 10.12.
1. The system overall appears to be unbiased and accurate. However, the
evaluation of individual inspectors indicates that there is room for
2. The results of the individual accuracy analysis indicate that Inspector C
has a problem with accuracy, see Table 10.10.
3. The results of the R&R (pairwise) indicate that Inspector C has a problem
with both repeatability and reproducibility, see Table 10.11.
4. The repeatability numbers are not very stable (Table 10.12). Comparing
the diagonal elements for Today with those of Last Week in Table
10.11, we see that Inspectors A and C tended to get di?erent results for
the di?erent weeks. Otherwise the system appears to be relatively stable.
5. Reproducibility of Inspectors A and B is not perfect. Some bene?t might
be obtained from looking at reasons for the di?erence.
Attribute measurement error analysis 355
Table 10.12. Stability analysis.
Stability of . . . Operational De?nition of Stability
Repeatability Standard deviation of the six repeatabilities (0.5, 1, 0.5, 1, 1, 1) 0.41
Reproducibility Standard deviation of the average repeatabilities. For data in
Table 10.9, =STDEV(AVERAGE(H2:H5),AVERAGE(H6:H9)) 0.00
Accuracy Standard deviation of the average accuracies. For data in Table
10.9, =STDEV(AVERAGE(I2:I5),AVERAGE(I6:I9)) 0.00
Bias Average of bias over the two weeks 0.0
6. Since Inspector B’s results are more accurate and repeatable, studying
her might lead to the discovery of best practices.
Minitab attribute gage R&R example
Minitab includes a built-in capability to analyze attribute measurement systems,
known as ‘‘attribute gage R&R.’’ We will repeat the above analysis using
Minitab can’t work with the data as shown in Table 10.9, it must be rearranged.
Once the data are in a format acceptable to Minitab, we enter the
Attribute Gage R&R Study dialog box by choosing Stat > Quality Tools >
Attribute Gage R&R Study (see Figure 10.16). Note the checkbox ‘‘Categories
of the attribute data are ordered.’’ Check this box if the data are ordinal and
have more than two levels. Ordinal data means, for example, a 1 is in some
sense ‘‘bigger’’ or ‘‘better’’ than a 0. For example, if we ask raters in a taste test
a question like the following: ‘‘Rate the flavor as 0 (awful), 1 (OK), or 2 (delicious).’’
Our data are ordinal (acceptable is better than unacceptable), but
there are only two levels, so we will not check this box.
Figure 10.16. Attribute gage R&R dialog box and data layout.
Within appraiser analysis
Minitab evaluates the repeatability of appraisers by examining how often the
appraiser ‘‘agrees with him/herself across trials.’’ It does this by looking at all
of the classifications for each part and counting the number of parts where all
classifications agreed. For our example each appraiser looked at two parts four
times each. Minitab’s output, shown in Figure 10.17, indicates that InspA rated
50% of the parts consistently, InspB 100%, and InspC 0%. The 95% confidence
interval on the percentage agreement is also shown. The results are displayed
graphically in Figure 10.18.
Attribute measurement error analysis 357
Figure 10.17. Minitab within appraiser assessment agreement.
Figure 10.18. Plot of within appraiser assessment agreement.
Accuracy Analysis
Minitab evaluates accuracy by looking at how often all of an appraiser’s classifications
for a given part agree with the standard. Figure 10.19 shows the
results for our example. As before, Minitab combines the results for both days.
The plot of these results is shown in Figure 10.20.
Figure 10.19. Minitab appraiser vs standard agreement.
Figure 10.20. Plot of appraiser vs standard assessment agreement.
Minitab also looks at whether or not there is a distinct pattern in the disagreements
with the standard. It does this by counting the number of times the
appraiser classified an item as a 1 when the standard said it was a 0 (the#1/0
Percent column), how often the appraiser classified an item as a 0 when it was
a 1 (the # 0/1 Percent column), and how often the appraiser’s classifications
were mixed, i.e., is not repeatable (the # Mixed Percent column). The results
are shown in Figure 10.21. The results indicate that there is no consistent bias,
defined as consistently putting a unit into the same wrong category. The problem,
as was shown in the previous analysis, is that appraisers A and C are not
Next, Minitab looks at all of the appraiser assessments for each part and
counts how often every appraiser agrees on the classification of the part. The
results, shown in Figure 10.22, indicate that this never happened during our
experiment. The 95% confidence interval is also shown.
Attribute measurement error analysis 359
Figure 10.21. Minitab appraiser assessment disagreement analysis.
Figure 10.22. Minitab between appraisers assessment agreement.
Finally, Minitab looks at all of the appraiser assessments for each part and
counts how often every appraiser agrees on the classification of the part
and their classification agrees with the standard. This can’t be any better
than the between appraiser assessment agreement shown in Figure 10.22.
Unsurprisingly, the results, shown in Figure 10.23, indicate that this never
happened during our experiment. The 95% confidence interval is also shown.
Figure 10.23. Minitab assessment vs standard agreement across all appraisers.
^ ^ ^
Knowledge Discovery
Getting the correct answer begins with asking the right question. The tools
and techniques described in this section help the Six Sigma team learn which
questions to ask. These simple tools are properly classified as data presentation
tools. Many are graphically based, creating easy to understand pictures from
the numbers and categories in the data. Others summarize the data, reducing
incomprehensible data in massive tables to a few succinct numbers that convey
essential information.
In addition to these traditional tools of the trade, the reader should determine
if they have access to on-line analytic processing (OLAP) tools. OLAP is
discussed briefly in Chapter 2. Contact your organization’s Information
Systems department for additional information regarding OLAP.
In this section, we will address the subject of time series analysis on a relatively
simple level. First, we will look at statistical methods that can be used
when we believe the data are from a stable process. This involves analysis of patterns
in runs of data in a time-ordered sequence.Wediscuss the problem of autocorrelation
in time series data and provide a method of dealing with this
problem in Chapter 12, EWMA charts.
Run charts
Run charts are plots of data arranged in time sequence. Analysis of run charts
is performed to determine if the patterns can be attributed to common causes
Copyright 2003 by The McGraw-Hill Companies, Inc. Click Here for Terms of Use.
of variation, or if special causes of variation were present. Run charts should be
used for preliminary analysis of any data measured on a continuous scale that
can be organized in time sequence. Run chart candidates include such things as
fuel consumption, production throughput, weight, size, etc. Run charts answer
the question ‘‘was this process in statistical control for the time period
observed?’’ If the answer is ‘‘no,’’ then the process was influenced by one or
more special causes of variation. If the answer is ‘‘yes,’’ then the long-term performance
of the process can be estimated using process capability analysis
methods. The run chart tests shown are all non-parametric, i.e., there are no
assumptions made regarding the underlying distribution.
1. Plot a line chart of the data in time sequence.
2. Find the median of the data. This can be easily done by using the line
chart you constructed in the above step. Simply place a straightedge or
a piece of paper across the top of the chart, parallel to the bottom axis.
Lower the straightedge until half of the data points appear above the
straightedge, or on it. Draw a horizontal line across the chart at that
point and label the line ‘‘Median’’ or ~XX. This procedure is shown in
Figure 11.1.
Figure 11.1. Using a straightedge to ?nd the median.
As you might expect, run charts are evaluated by examining the ‘‘runs’’ on the
chart. A ‘‘run’’ is a time-ordered sequence of points. There are several different
statistical tests that can be applied to the runs.
A run to the median is a series of consecutive points on the same side of the
median. Unless the process is being influenced by special causes, it is unlikely
that a long series of consecutive points will all fall on the same side of the
median. Thus, checking run length is one way of checking for special causes
of variation. The length of a run is found by simply counting the number of
consecutive points on the same side of the median. However, it may be that
some values are exactly equal to the median. If only one value is exactly on
the median line, ignore it. There will always be at least one value exactly on
the median if you have an odd number of data points. If more than one
value is on the line, assign them to one side or the other in a way that results
in 50% being on one side and 50% on the other. On the run chart, mark
those that will be counted as above the median with an a and those that will
be counted below the median with a b. The run length concept is illustrated
in Figure 11.2.
Knowledge discovery tools 363
Figure 11.2. Determination of run length.
After finding the longest run, compare the length of the longest run to the
values in Table 11.1. If the longest run is longer than the maximum allowed,
then the process was probably influenced by a special cause of variation
(a ? 0:05). With the example, there are 20 values plotted and the longest run
was 8. Table 11.1 indicates that a run of 7 would not be too unusual for 20 plotted
points but a run of 8 would be. Since our longest run is 8, we conclude that a
special cause of variation is indicated and conduct an investigation to identify
the special cause.
The number of runs we expect to find from a controlled process can also be
mathematically determined. A process that is not being influenced by special
causes will not have either too many runs or too few runs. The number of runs
is found by simple counting. Referring to Figure 11.3, we see that there are 5
Table 11.2 is used to evaluate the number of runs. If you have fewer runs
than the smallest allowed or more runs than the largest allowed then there is
a high probability (a ? 0:05) that a special cause is present. With the example,
we have 20 values plotted and 5 runs. Table 11.2 indicates that for 20 plotted
points, 6 to 15 runs are expected, so we conclude that a special cause was
Table 11.1. Maximum run length.
10 5
15 6
20 7
30 8
40 9
50 10
The run chart should not have any unusually long series of consecutive
increases or decreases. If it does, then a trend is indicated and it is probably
due to a special cause of variation (a?0.05). Compare the longest count of consecutive
increases or decreases to the longest allowed shown in Table 11.3, and
if your count exceeds the table value then it is likely that a special cause of variation
caused the process to drift.
Figure 11.4 shows the analysis of trends. Note that the trend can extend on
both sides of the median, i.e., for this particular run test the median is ignored.
When counting increases or decreases, ignore ‘‘no change’’ values. For
example, the trend length in the series 2, 3, 3, 5, 6 is four.
Run charts should not be used if too many of the numbers are the same. As a
rule of thumb, don’t use run charts if more than 30% of the values are the same.
For example, in the data set 1, 2, 3, 3, 6, 7, 7, 11, 17, 19, the number 3 appears
twice and the number 7 appears twice. Thus, 4 of the 10, or 40% of the values
are the same.
Run charts are preliminary analysis tools, so if you have continuous data in
time-order always sketch a quick run chart before doing any more complex
Knowledge discovery tools 365
Figure 11.3. Determination of number of runs.
Table 11.2. Limits on the number of runs.
Number of Plotted
Values Smallest Run Count Largest Run Count
10 3 8
12 3 10
14 4 11
16 5 12
18 6 13
20 6 15
22 7 16
24 8 17
26 9 18
28 10 19
30 11 20
32 11 22
34 12 23
36 13 24
38 14 25
40 15 26
42 16 27
44 17 28
46 17 30
48 18 31
50 19 32
analysis. Often the patterns on a run chart will point you in the right direction
without any further work.
Run charts are one of the least sensitive SPC techniques. They are unable to
detect ‘‘freaks,’’ i.e., single points dramatically different from the rest. Thus,
run charts may fail to find a special cause even if a special cause was present. In
Knowledge discovery tools 367
Figure 11.4. Determination of trend length.
Table 11.3. Maximum consecutive increases/decreases.
5 to 8 4
9 to 20 5
21 to 100 6
101 or more 7
statistical parlance, run charts tend to have large Type II errors, i.e., they have a
high probability of accepting the hypothesis of no special cause even when the
special cause actually exists. Use run charts to aid in troubleshooting. The different
run tests indicate different types of special causes. A long run on the same
side of the median indicates a special cause that created a process shift. A long
series of consecutively increasing or decreasing values indicates a special cause
that created a trend. Too many runs often indicates a mixture of several sources
of variation in the sample. Too few runs often occur in conjunction with a process
shift or trend. If you have too few runs and they are not caused by a process
shift or trend, then too few runs may indicate a mixture that follows a definite
pattern (e.g., an operator who is periodically relieved).
Descriptive statistics
Typically, descriptive statistics are computed to describe properties of
empirical distributions, that is, distributions of data from samples. There are
three areas of interest: the distribution’s location or central tendency, its dispersion,
and its shape. The analyst may also want some idea of the magnitude
of possible error in the statistical estimates. Table 11.4 describes some of the
more common descriptive statistical measures.
Table 11.4. Common descriptive statistics.
Measures of location
Population mean The center of gravity or
centroid of the distribution.
 ? 1N
where x is an observation,Nis
the population size.
Sample mean The center of gravity or
centroid of a sample from a
X ? 1n
where x is an observation, n is
the sample size.
Knowledge discovery tools 369
Median The 50%/50% split point.
Precisely half of the data set
will be above the median,
and half below it.
Mode The value that occurs most
often. If the data are
grouped, the mode is the
group with the highest
Measures of dispersion
Range The distance between the
sample extreme values.
Population variance A measure of the variation
around the mean; units are
the square of the units used
for the original data.
2 ? PN
?xi  ?2
Population standard
A measure of the variation
around the mean, in the
same units as the original
 ? ffiffiffiffiffi 2 p
Sample variance A measure of the variation
around the mean; units are
the square of the units used
for the original data.
s2 ? Pn
?xi  XX?2
n  1
Sample standard deviation A measure of the variation
around the mean, in the
same units as the original
s ? ffiffiffiffi s2 p
Continued on next page . . .
Figures 11.5^11.8 illustrate distributions with different descriptive statistics.
Measures of shape
Ameasure of asymmetry.
Zero indicates perfect
symmetry; the normal
distribution has a skewness
of zero. Positive skewness
indicates that the ‘‘tail’’ of the
distribution is more stretched
on the side above the mean.
Negative skewness indicates
that the tail of the
distribution is more stretched
on the side below themean.
k ?
3 XXXn
n ? 2 XX3
k ?
3 XXXn
n ? 2  X3
Kurtosis Kurtosis is a measure of ?atness of the distribution. Heavier
tailed distributions have larger kurtosis measures. The normal
distribution has a kurtosis of 3.

2 ? Pn
n  4 X
n ? 6 X
n  3 XX4
Table 11.4. (cont.)
Figure 11.5. Illustration of mean, median, and mode.
A histogram is a pictorial representation of a set of data. It is created by
grouping the measurements into ‘‘cells.’’ Histograms are used to determine the
shape of a data set. Also, a histogram displays the numbers in a way that makes
it easy to see the dispersion and central tendency and to compare the distrib-
Knowledge discovery tools 371
Figure 11.6. Illustration of sigma.
Figure 11.7. Illustration of skewness.
Figure 11.8. Illustration of kurtosis.
ution to requirements. Histograms can be valuable troubleshooting aids.
Comparisons between histograms from different machines, operators, vendors,
etc., often reveal important differences.
1. Find the largest and the smallest value in the data.
2. Compute the range by subtracting the smallest value from the largest
3. Select a number of cells for the histogram. Table 11.5 provides some useful
guidelines. The ?nal histogram may not have exactly the number of
cells you choose here, as explained below.
As an alternative, the number of cells can be found as the square root of
the number in the sample. For example, if n?100, then the histogram
would have 10 cells. Round to the nearest integer.
4. Determine the width of each cell.Wewill use the letterWto stand for the
cell width.Wis computed from Equation 11.1.
W ?
Number of Cells ?11:1?
The number W is a starting point. Round W to a convenient number.
RoundingWwill a?ect the number of cells in your histogram.
5. Compute ‘‘cell boundaries.’’ A cell is a range of values and cell boundaries
de?ne the start and end of each cell. Cell boundaries should have
one more decimal place than the raw data values in the data set; for example,
if the data are integers, the cell boundaries would have one decimal
place. The low boundary of the ?rst cell must be less than the smallest
value in the data set. Other cell boundaries are found by adding W to
Table 11.5. Histogram cell determination guidelines.
100 or less 7 to 10
101^200 11 to 15
201 or more 13 to 20
the previous boundary. Continue until the upper boundary is larger than
the largest value in the data set.
6. Go through the raw data and determine into which cell each value falls.
Mark a tick in the appropriate cell.
7. Count the ticks in each cell and record the count, also called the frequency,
to the right of the tick marks.
8. Construct a graph from the table. The vertical axis of the graph will show
the frequency in each cell. The horizontal axis will show the cell boundaries.
Figure 11.9 illustrates the layout of a histogram.
9. Draw bars representing the cell frequencies. The bars should all be the
same width, the height of the bars should equal the frequency in the cell.
Assume you have the data in Table 11.6 on the size of a metal rod. The rods
were sampled every hour for 20 consecutive hours and 5 consecutive rods were
checked each time (20 subgroups of 5 values per group).
1. Find the largest and the smallest value in the data set. The smallest value
is 0.982 and the largest is 1.021. Both values are marked with an (*) in
Table 11.6.
Knowledge discovery tools 373
Figure 11.9. Layout of a histogram.
From Pyzdek’s Guide to SPCVolume One: Fundamentals, p. 61.
Copyright#1990 by Thomas Pyzdek.
2. Compute the range, R, by subtracting the smallest value from the largest
value. R?1.021 ^ 0.982?0.039.
3. Select a number of cells for the histogram. Since we have 100 values, 7 to
10 cells are recommended. We will use 10 cells.
4. Determine the width of each cell, W. Using Equation 11.1, we compute
W?0.039/10?0.0039. We will round this to 0.004 for convenience.
Table 11.6. Data for histogram.
From Pyzdek’s Guide to SPCVolume One: Fundamentals, p. 62.
Copyright#1990 by Thomas Pyzdek.
1 1.002 0.995 1.000 1.002 1.005
2 1.000 0.997 1.007 0.992 0.995
3 0.997 1.013 1.001 0.985 1.002
4 0.990 1.008 1.005 0.994 1.012
5 0.992 1.012 1.005 0.985 1.006
6 1.000 1.002 1.006 1.007 0.993
7 0.984 0.994 0.998 1.006 1.002
8 0.987 0.994 1.002 0.997 1.008
9 0.992 0.988 1.015 0.987 1.006
10 0.994 0.990 0.991 1.002 0.988
11 1.007 1.008 0.990 1.001 0.999
12 0.995 0.989 0.982* 0.995 1.002
13 0.987 1.004 0.992 1.002 0.992
14 0.991 1.001 0.996 0.997 0.984
15 1.004 0.993 1.003 0.992 1.010
16 1.004 1.010 0.984 0.997 1.008
17 0.990 1.021* 0.995 0.987 0.989
18 1.003 0.992 0.992 0.990 1.014
19 1.000 0.985 1.019 1.002 0.986
20 0.996 0.984 1.005 1.016 1.012
5. Compute the cell boundaries. The low boundary of the ?rst cell must be
below our smallest value of 0.982, and our cell boundaries should have
one decimal place more than our raw data. Thus, the lower cell boundary
for the ?rst cell will be 0.9815. Other cell boundaries are found by adding
W?0.004 to the previous cell boundary until the upper boundary is
greater than our largest value of 1.021. This gives us the cell boundaries
in Table 11.7.
6. Go through the raw data and mark a tick in the appropriate cell for each
data point.
7. Count the tick marks in each cell and record the frequency to the right of
each cell. The results of all we have done so far are shown in Table 11.8.
Table 11.8 is often referred to as a ‘‘frequency table’’ or ‘‘frequency tally
Knowledge discovery tools 375
Table 11.7. Histogram cell boundaries.
From Pyzdek’s Guide to SPCVolume One: Fundamentals, p. 63.
Copyright#1990 by Thomas Pyzdek.
1 0.9815 0.9855
2 0.9855 0.9895
3 0.9895 0.9935
4 0.9935 0.9975
5 0.9975 1.0015
6 1.0015 1.0055
7 1.0055 1.0095
8 1.0095 1.0135
9 1.0135 1.0175
10 1.0175 1.0215
Construct a graph from the table in step 7. The frequency column will be
plotted on the vertical axis, and the cell boundaries will be shown on the horizontal
(bottom) axis. The resulting histogram is shown in Figure 11.10.
Minitab’s histogram function expects to have the data in a single column. If your
data are not arranged this way, you can use Minitab’s Manip-Stack Columns
function to put multiple columns into a single column. Once the data are in
the proper format, use Minitab’s Graph-Histogram function (Figure 11.11) to
create the histogram (Figure 11.12).
Table 11.8. Frequency tally sheet.
From Pyzdek’s Guide to SPCVolume One: Fundamentals, p. 64.
Copyright# 1990 by Thomas Pyzdek.
1 0.9815 0.9855 IIII IIII 8
2 0.9855 0.9895 IIII IIIII 9
3 0.9895 0.9935 IIII IIIII IIIII III 17
4 0.9935 0.9975 IIII IIIII IIIII III 16
5 0.9975 1.0015 IIII IIIII 9
6 1.0015 1.0055 IIII IIIII IIIII IIIII 19
7 1.0055 1.0095 IIII IIIII III 11
8 1.0095 1.0135 IIII II 6
9 1.0135 1.0175 III 3
10 1.0175 1.0215 II 2
Knowledge discovery tools 377
Figure 11.10. Completed histogram.
From Pyzdek’s Guide to SPCVolume One: Fundamentals, p. 61.
Copyright#1990 by Thomas Pyzdek.
Figure 11.11. Minitab’s histogram dialog box.
It is often helpful to see a distribution curve superimposed over the histogram.
Minitab has the ability to put a wide variety of distribution curves on
histograms, although the procedure is tedious. It will be illustrated here for the
normal distribution. Minitab’s help facility also describes the procedure
shown below.
1. Sort the data using Manip > Sort. Store the sorted data in a di?erent
column than the one containing the original data.
2. Determine the mean and sigma value using Stat > Basic Statistics >
Store Descriptive Statistics. For our data:
3. Get the probability distribution using Calc >Probability Distributions
> Normal. Enter the mean and standard deviation. Store the results in
a separate column (e.g., NormProbData).
Figure 11.12. Histogram.
4. Adjust the NormProbData values in accordance with your histogram
cell interval. In the previous example we let Minitab choose the cell interval
for use. Here we will use a cell interval of 0.005. Choose Calc >
Calculator and enter the information as shown below:
5. Create the histogram with the normal curve:
Choose Graph >Histogram
For X, select StackedData
Knowledge discovery tools 379
Click Options. Under Type of Intervals, choose CutPoint. Under
De?nition of Intervals, choose Midpoint/Cutpoint positions and
type 0.98:1.02/0.005. Click OK.
Choose Annotation > Line. In Points, choose Use Variables and
select StackedData NormProbData. Click OK.
The completed histogram is shown in Figure 11.13.
. Histograms can be used to compare a process to requirements if you draw
the speci?cation lines on the histogram. If you do this, be sure to scale
the histogram accordingly.
. Histograms should not be used alone. Always construct a run chart or a
control chart before constructing a histogram. They are needed because
histograms will often conceal out of control conditions due to the fact
that they don’t show the time sequence of the data.
. Evaluate the pattern of the histogram to determine if you can detect
changes of any kind. The changes will usually be indicated by multiple
modes or ‘‘peaks’’ on the histogram. Most real-world processes produce
histograms with a single peak. However, histograms from small samples
often have multiple peaks that merely represent sampling variation. Also,
multiple peaks are sometimes caused by an unfortunate choice of the
number of cells. Also, processes heavily in?uenced by behavior patterns
are often multi-modal. For example, tra?c patterns have distinct ‘‘rush-
Figure 11.13. Histogram with normal curve superimposed.
hours,’’ and prime time is prime time precisely because more people tend
to watch television at that time.
. Compare histograms from di?erent periods of time. Changes in histogram
patterns from one time period to the next can be very useful in ?nding
ways to improve the process.
. Stratify the data by plotting separate histograms for di?erent sources of
data. For example, with the rod diameter histogram we might want to
plot separate histograms for shafts made from di?erent vendors’ materials
or made by di?erent operators or machines. This can sometimes reveal
things that even control charts don’t detect.
Exploratory data analysis
Data analysis can be divided into two broad phases: an exploratory phase and
a confirmatory phase. Data analysis can be thought of as detective work.
Before the ‘‘trial’’ one must collect evidence and examine it thoroughly. One
must have a basis for developing a theory of cause and effect. Is there a gap in
the data? Are there patterns that suggest some mechanism? Or, are there patterns
that are simply mysterious (e.g., are all of the numbers even or odd)? Do
outliers occur? Are there patterns in the variation of the data? What are the
shapes of the distributions? This activity is known as exploratory data analysis
(EDA). Tukey’s 1977 book with this title elevated this task to acceptability
among ‘‘serious’’ devotees of statistics.
Four themes appear repeatedly throughout EDA: resistance, residuals, reexpression,
and visual display. Resistance refers to the insensitivity of a method
to a small change in the data. If a small amount of the data is contaminated, the
method shouldn’t produce dramatically different results. Residuals are what
remain after removing the effect of a model or a summary. For example, one
might subtract the mean from each value, or look at deviations about a regression
line. Re-expression involves examination of different scales on which the
data are displayed. Tukey focused most of his attention on simple power transformations
such as y ? ffiffiffi x p , y ? x2, y ? 1=x . Visual display helps the analyst
examine the data graphically to grasp regularities and peculiarities in the data.
EDA is based on a simple basic premise: it is important to understand what
you can do before you learn to measure how well you seem to have done it
(Tukey, 1977). The objective is to investigate the appearance of the data, not to
confirm some prior hypothesis. While there are a large number of EDA methods
and techniques, there are two which are commonly encountered in Six
Sigma work: stem-and-leaf plots and boxplots. These techniques are commonly
included in most statistics packages. (SPSS was used to create the figures used
Knowledge discovery tools 381
in this book.) However, the graphics of EDA are simple enough to be done
easily by hand.
Stem-and-leaf plots are a variation of histograms and are especially useful for
smaller data sets (n<200). A major advantage of stem-and-leaf plots over the
histogram is that the raw data values are preserved, sometimes completely and
sometimes only partially. There is a loss of information in the histogram
because the histogram reduces the data by grouping several values into a single
Figure 11.14 is a stem-and-leaf plot of diastolic blood pressures. As in a histogram,
the length of each row corresponds to the number of cases that fall
into a particular interval. However, a stem-and-leaf plot represents each case
with a numeric value that corresponds to the actual observed value. This is
done by dividing observed values into two componentsLthe leading digit or
digits, called the stem, and the trailing digit, called the leaf. For example, the
value 75 has a stem of 7 and a leaf of 5.
Figure 11.14. Stem-and-leaf plot of diastolic blood pressures.
From SPSS forW|ndows Base System User’s Guide, p. 183. Copyright#1993. Used by
permission of the publisher, SPSS, Inc., Chicago, IL.
In this example, each stem is divided into two rows. The first row of each pair
has cases with leaves of 0 through 4, while the second row has cases with leaves
of 5 through 9. Consider the two rows that correspond to the stem of 11. In the
first row, we can see that there are four cases with diastolic blood pressure of
110 and one case with a reading of 113. In the second row, there are two cases
with a value of 115 and one case each with a value of 117, 118, and 119.
The last row of the stem-and-leaf plot is for cases with extreme values (values
far removed from the rest). In this row, the actual values are displayed in
parentheses. In the frequency column, we see that there are four extreme cases.
Their values are 125, 133, and 160. Only distinct values are listed.
When there are few stems, it is sometimes useful to subdivide each stem even
further. Consider Figure 11.15 a stem-and-leaf plot of cholesterol levels. In this
figure, stems 2 and 3 are divided into five parts, each representing two leaf
values. The first row, designated by an asterisk, is for leaves of 0 and 1; the
next, designated by t, is for leaves of 2’s and 3’s; the third, designated by f, is for
leaves of 4’s and 5’s; the fourth, designated by s, is for leaves of 6’s and 7’s; and
the fifth, designated by a period, is for leaves of 8’s and 9’s. Rows without cases
are not represented in the plot. For example, in Figure 11.15, the first two rows
for stem 1 (corresponding to 0-1 and 2-3) are omitted.
Knowledge discovery tools 383
Figure 11.15. Stem-and-leaf plot of cholesterol levels.
From SPSS forW|ndows Base System User’s Guide, p. 185. Copyright#1993. Used by
permission of the publisher, SPSS, Inc., Chicago, IL.
This stem-and-leaf plot differs from the previous one in another way. Since
cholesterol values have a wide rangeLfrom 106 to 515 in this exampleLusing
the first two digits for the stem would result in an unnecessarily detailed plot.
Therefore, we will use only the hundreds digit as the stem, rather than the first
two digits. The stem setting of 100 appears in the row labeled Stem width. The
leaf is then the tens digit. The last digit is ignored. Thus, from this particular
stem-and-leaf plot, it is not possible to determine the exact cholesterol level for
a case. Instead, each is classified by only its first two digits.
A display that further summarizes information about the distribution of the
values is the boxplot. Instead of plotting the actual values, a boxplot displays
summary statistics for the distribution. It is a plot of the 25th, 50th, and 75th percentiles,
as well as values far removed from the rest.
Figure 11.16 shows an annotated sketch of a boxplot. The lower boundary of
the box is the 25th percentile. Tukey refers to the 25th and 75th percentile
‘‘hinges.’’ Note that the 50th percentile is the median of the overall data set,
the 25th percentile is the median of those values below the median, and the
75th percentile is the median of those values above the median. The horizontal
line inside the box represents the median. 50% of the cases are included within
the box. The box length corresponds to the interquartile range, which is the difference
between the 25th and 75th percentiles.
The boxplot includes two categories of cases with outlying values. Cases with
values that are more than 3 box-lengths from the upper or lower edge of the
box are called extreme values. On the boxplot, these are designated with an
asterisk (*). Cases with values that are between 1.5 and 3 box-lengths from the
upper or lower edge of the box are called outliers and are designated with a circle.
The largest and smallest observed values that aren’t outliers are also shown.
Lines are drawn from the ends of the box to these values. (These lines are sometimes
called whiskers and the plot is then called a box-and-whiskers plot.)
Despite its simplicity, the boxplot contains an impressive amount of information.
From the median you can determine the central tendency, or location.
From the length of the box, you can determine the spread, or variability, of
your observations. If the median is not in the center of the box, you know that
the observed values are skewed. If the median is closer to the bottom of the box
than to the top, the data are positively skewed. If the median is closer to the
top of the box than to the bottom, the opposite is true: the distribution is negatively
skewed. The length of the tail is shown by the whiskers and the outlying
and extreme points.
Boxplots are particularly useful for comparing the distribution of values in
several groups. Figure 11.17 shows boxplots for the salaries for several different
job titles.
The boxplot makes it easy to see the different properties of the distributions.
The location, variability, and shapes of the distributions are obvious at a glance.
This ease of interpretation is something that statistics alone cannot provide.
The process baseline is best described as ‘‘what were things like before the
project?’’ There are several reasons for obtaining this information:
Establishing the process baseline 385
Figure 11.16. Annotated boxplot.
& To determine if the project should be pursued. Although the project charter
provides a business case for the project, it sometimes happens that additional,
detailed information fails to support it. It may be that the situation
isn’t as bad as people think, or the project may be addressing an unimportant
aspect of the problem.
& To orient the project team. The process baseline helps the team identify
CTQs and other hard metrics. The information on the historic performance
of these metrics may point the team to strategies. For example, if
the process is erratic and unstable the team would pursue a di?erent
strategy than if it was operating at a consistently poor level.
& To provide data that will be used to estimate savings. Baseline information
will be invaluable when the project is over and the team is trying to determine
the magnitude of the savings or improvement. Many a Black Belt
has discovered after the fact that the information they need is no longer
available after the completion of the project, making it impossible to
determine what bene?t was obtained. For example, a project that streamlined
a production control system was aimed at improving morale by
reducing unpaid overtime worked by exempt employees. However, no
measure of employee morale was obtained ahead of time. Nor was the
unpaid overtime documented anywhere. Consequently, the Black Belt
wasn’t able to substantiate his claims of improvement and his certi?cation
(and pay increase) was postponed.
Figure 11.17. Boxplots of salary by job category.
Describing the Process Baseline
The process baseline should be described in both qualitative and quantitative
terms. It’s not enough to report survey results or complaint counts, the voice
of the customer (VOC) should be made heard. If your customers are saying
that your service stinks, then don’t mince words, tell it like the customer tells
it.* Likewise, include glowing praise. The new process might make the average
performance better and improve consistency, but if it creates ho-hum satisfaction
at the expense of eliminating delight, that fact should be known. It may be
that the new system takes the joy out of customer service work, which will
have adverse consequences on employee morale and might lead to unexpected
consequences that need to be considered.
(VOE). Do employees say they feel great about what they do? Or do they dread
coming in each day? Is it a great place to work? Why or why not? Are employees
eager to transfer to other jobs just to getawayfromthe stress?Whatdoemployees
think would make things better? If you find a workplace that is a delight to the
employees, you might want to think twice about changing it. If the workplace is a
chamber of horrors, you may want to speed up the pace of the project.
A descriptive narrative by the team members should be considered. Every
team member should spend time in the work area. There is no substitute for
firsthand observation. It need not be a formal audit. Just go and look at the
way things are done. Talk to people doing the work and actively listen to what
they have to say. Watch what they do. Look for differences between the way different
people do similar tasks. Document your observations and share them
with the team and the sponsor.
Collect information contained in memos, email, reports, studies, etc.
Organize the information using affinity analysis and other methods of categorizing.
Arrange it in time-order and look for patterns. Were things once better than
they are now? Are things getting worse? What might account for these trends?
Quantifying the process baseline involves answering some simple questions:
& What are the key metrics for this process? (Critical to quality, cost,
schedule, customer satisfaction, etc.)
* What are the operational de?nitions of these metrics?
Establishing the process baseline 387
*Profanity and obscenity excepted, of course.
* Are these the metrics that will be used after completing the project to
measure success?
& What data are available for these metrics?
* If none, how will data be obtained?
* What is the quality of the data?
Once the metrics are identified and the data located and validated, perform
analyses to answer these questions:
& When historical data are looked at over a period of time, are there any patterns
or trends? What might be causing this? (Run charts, time series
charts, process behavior charts (control charts))
& Were things ever better than they are now? Why?
& Should the data be transformed to make them easier to analyze?
& What is the historical central tendency? (Mean, median and mode)
& What is the historical variability? (Inter-quartile range, standard deviation)
& What is the historical shape or distribution? (Histograms, stem-and-leaf
plots, box plots, dot plots)
& Are there any interesting relationships between variables? Cross tabulations
should be created to evaluate possible relationships with categorical
data. Scatterplots and correlation studies can be used to study continuous
The analyses can be performed by any team member, with the guidance of a
Black Belt. The results should be shared with the team. Brief summaries of especially
important findings should be reported to the process owner and sponsor.
If the results indicate a need to change the project charter, the sponsor should
be informed.
Virtually all Six Sigma projects address business processes that have an
impact on a top-level enterprise strategy. In previous chapters a great deal of
attention was devoted to developing a list of project candidates by meticulously
linking projects and strategies using dashboards, QFD, structured decision
making, business process mapping, and many other tools and techniques.
However, Six Sigma teams usually find that although this approach succeeds
in identifying important projects, these projects tend to have too large a scope
to be completed within the time and budget constraints. More work is needed
to clearly define that portion of the overall business process to be improved by
the project. One way to do this is to apply process flowcharting or mapping to
subprocesses until reaching the part of the process that has been assigned to
the team for improvement. A series of questions are asked, such as:
1. For which stakeholder does this process primarily exist?
2. What value does it create? What output is produced?
3. Who is the owner of this process?
4. Who provides inputs to this process?
5. What are the inputs?
6. What resources does this process use?
7. What steps create the value?
8. Are there subprocesses with natural start and end points?
These questions, which are common to nearly all processes addressed by Six
Sigma projects, have been arranged into a standard format known as SIPOC.
SIPOC stands for Suppliers-Inputs-Process-Outputs-Customers.
Process for creating a SIPOC diagram
SIPOCs begin with people who know something about the process. This may
involve people who are not full-time members of the Six Sigma team. Bring the
people together in a room and conduct a ‘‘focused brainstorming’’ session. To
begin, briefly describe the process and obtain consensus on the definition. For
& ‘‘Make it easy for the customer to reach technical support by phone’’
& ‘‘Reduce the space needed to store tooling’’
& ‘‘Reduce the downtime on the Niad CNC machine’’
& ‘‘Get roo?ng crew to the work site on time’’
& ‘‘Reduce extra trips taken by copier maintenance person’’
Post flipcharts labeled suppliers, inputs, process, outputs, customers. Once
the process has been described, create the SIPOC diagram as follows:
1. Create a simple, high-level process map of the process. Display this conspicuously
while the remaining steps are taken to provide a reminder to
the team.
Perform the steps below using brainstorming rules. Write down all ideas without
critiquing them.
2. Identify the outputs of this process. Record on the Outputs ?ip chart.
3. Identify the customers who will receive the outputs. Record on the
Customers ?ip chart.
4. Identify the inputs needed for the process to create the outputs. Record
on the Inputs ?ip chart.
5. Identify the suppliers of the inputs. Record on the Suppliers ?ip chart.
6. Clean up the lists by analyzing, rephrasing, combining, moving, etc.
7. Create a SIPOC diagram.
8. Review the SIPOC with the project sponsor and process owner. Modify
as necessary.
SIPOC example
A software company wants to improve overall customer satisfaction (Big Y).
Research has indicated that a key component of overall satisfaction is satisfaction
with technical support (Little Y). Additional drill down of customer comments
indicates that one important driver of technical support satisfaction is
the customer’s perception that it is easy to contact technical support. There
Figure 11.18. Process map for contacting technical support by telephone.
are several different types of technical support available, such as self-help built
into the product, the web, or the phone. The process owner commissioned Six
Sigma projects for each type of contact. This team’s charter is telephone support.
To begin, the team created the process map shown in Figure 11.18.
Next the team determined that there were different billing options and
created a work breakdown structure with each billing option being treated as a
subproject. For this example we will follow the subproject relating to the
billing-by-the-minute (BBTM) option. After completing the process described
above, the team produced the SIPOC shown in Figure 11.19.
Note that the process is mapped at a very low level. At this level the process
map is usually linear, with no decision boxes shown. The typical SIPOC shows
the process as it is supposed to behave. Optionally, the SIPOC can show the
unintended or undesirable outcomes, as shown in Figure 11.20.
Figure 11.19. SIPOC for easy to contact BBTM.
This ‘‘bad process’’ SIPOC is used only for team troubleshooting. It helps the
team formulate hypotheses to be tested during the analyze phase.
SIPOC analyses focus on the Xs that drive the Ys. It helps the team understand
which ‘‘dials to turn’’ to make the top-level dashboard’s Big Y move. In
the example, let’s assume that the team collects information and determines
that a significant percentage of the customers can’t find the phone number for
technical support. A root cause of the problem then, is the obscure location of
the support center phone number. Improving overall customer satisfaction is
linked to making it easier for the customer to locate the correct number, perhaps
by placing a big, conspicuous sticker with the phone number on the cover
of the manual. The Big Y and the root cause X are separated by several levels,
but the process mapping and SIPOC analysis chain provides a methodology
for making the connection.
Figure 11.20. SIPOC for undesirable outcomes.
^ ^ ^
Statistical Process Control
Types of control charts
There are two broad categories of control charts: those for use with continuous
data (e.g., measurements) and those for use with attributes data (e.g.,
counts). This section describes the various charts used for these different data.
In statistical process control (SPC), the mean, range, and standard deviation
are the statistics most often used for analyzing measurement data. Control
charts are used to monitor these statistics. An out-of-control point for any of
these statistics is an indication that a special cause of variation is present and
that an immediate investigation should be made to identify the special cause.
Averages and ranges control charts
Averages charts are statistical tools used to evaluate the central tendency of a
process over time. Ranges charts are statistical tools used to evaluate the dispersion
or spread of a process over time.
Copyright 2003 by The McGraw-Hill Companies, Inc. Click Here for Terms of Use.
Averages charts answer the question: ‘‘Has a special cause of variation caused
the central tendency of this process to change over the time period observed?’’
Ranges charts answer the question: ‘‘Has a special cause of variation caused the
process distribution to become more or less consistent?’’ Averages and ranges
charts can be applied to any continuous variable such as weight, size, etc.
The basis of the control chart is the rational subgroup. Rational subgroups (see
page 420) are composed of items which were produced under essentially the
same conditions. The average and range are computed for each subgroup separately,
then plotted on the control chart. Each subgroup’s statistics are compared
to the control limits, and patterns of variation between subgroups are analyzed.
Subgroup equations for averages and ranges charts
X ?
sum of subgroup measurements
subgroup size ?12:1?
R ? Largest in subgroup  Smallest in subgroup ?12:2?
Control limit equations for averages and ranges charts
Control limits for both the averages and the ranges charts are computed such
that it is highly unlikely that a subgroup average or range from a stable process
would fall outside of the limits. All control limits are set at plus and minus
three standard deviations from the center line of the chart. Thus, the control
limits for subgroup averages are plus and minus three standard deviations of
the mean from the grand average; the control limits for the subgroup ranges
are plus and minus three standard deviations of the range from the average
range. These control limits are quite robust with respect to non-normality in
the process distribution.
To facilitate calculations, constants are used in the control limit equations.
Table 11 in the Appendix provides control chart constants for subgroups of 25
or less. The derivation of the various control chart constants is shown in Burr
(1976, pp. 97^105).
Control limit equations for ranges charts
R ?
sum of subgroup ranges
number of subgroups ?12:3?
LCL ? D3 RR ?12:4?
UCL ? D4 RR ?12:5?
Control limit equations for averages charts using R-bar
X ?
sum of subgroup averages
number of subgroups ?12:6?
LCL ? XX  A2 RR ?12:7?
UCL ? XX ? A2 RR ?12:8?
Example of averages and ranges control charts
Table 12.1 contains 25 subgroups of five observations each.
The control limits are calculated from these data as follows:
Ranges control chart example
R ?
sum of subgroup ranges
number of subgroups ?
25 ? 14:76
LCLR ? D3 RR ? 0  14:76 ? 0
UCLR ? D4 RR ? 2:115  14:76 ? 31:22
Since it is not possible to have a subgroup range less than zero, the LCL is not
shown on the control chart for ranges.
Averages control chart example
X ?
sum of subgroup averages
number of subgroups ?
25 ? 99:5
Statistical process control (SPC) 395
Table 12.1. Data for averages and ranges control charts.
Sample 1 Sample 2 Sample 3 Sample 4 Sample 5 AVERAGE RANGE
110 93 99 98 109 101.8 17
103 95 109 95 98 100.0 14
97 110 90 97 100 98.8 20
96 102 105 90 96 97.8 15
105 110 109 93 98 103.0 17
110 91 104 91 101 99.4 19
100 96 104 93 96 97.8 11
93 90 110 109 105 101.4 20
90 105 109 90 108 100.4 19
103 93 93 99 96 96.8 10
97 97 104 103 92 98.6 12
103 100 91 103 105 100.4 14
90 101 96 104 108 99.8 18
97 106 97 105 96 100.2 10
99 94 96 98 90 95.4 9
106 93 104 93 99 99.0 13
90 95 98 109 110 100.4 20
96 96 108 97 103 100.0 12
109 96 91 98 109 100.6 18
90 95 94 107 99 97.0 17
91 101 96 96 109 98.6 18
108 97 101 103 94 100.6 14
96 97 106 96 98 98.6 10
101 107 104 109 104 105.0 8
96 91 96 91 105 95.8 14
LCLXX ? XX  A2 RR ? 99:5  0:577  14:76 ? 90:97
UCLXX ? XX ? A2 RR ? 99:5 ? 0:577  14:76 ? 108:00
The completed averages and ranges control charts are shown in Figure 12.1.
Statistical process control (SPC) 397
Figure 12.1. Completed averages and ranges control charts.
The above charts show a process in statistical control. This merely means that
we can predict the limits of variability for this process. To determine the capability
of the process with respect to requirements one must use the methods
described in Chapter 13, Process Capability Analysis.
Averages and standard deviation (sigma) control charts
Averages and standard deviation control charts are conceptually identical to
averages and ranges control charts. The difference is that the subgroup standard
deviation is used to measure dispersion rather than the subgroup range. The
subgroup standard deviation is statistically more efficient than the subgroup
range for subgroup sizes greater than 2. This efficiency advantage increases as
the subgroup size increases. However, the range is easier to compute and easier
for most people to understand. In general, this author recommends using subgroup
ranges unless the subgroup size is 10 or larger. However, if the analyses
are to be interpreted by statistically knowledgeable personnel and calculations
are not a problem, the standard deviation chart may be preferred for all subgroup
Subgroup equations for averages and sigma charts
X ?
sum of subgroup measurements
subgroup size ?12:9?
s ?
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi Pn
i?1?xi  XX?2
n  1
The standard deviation, s, is computed separately for each subgroup, using
the subgroup average rather than the grand average. This is an important
point; using the grand average would introduce special cause variation if the
process were out of control, thereby underestimating the process capability,
perhaps significantly.
Control limit equations for averages and sigma charts
Control limits for both the averages and the sigma charts are computed such
that it is highly unlikely that a subgroup average or sigma from a stable process
would fall outside of the limits. All control limits are set at plus and minus
three standard deviations from the center line of the chart. Thus, the control
limits for subgroup averages are plus and minus three standard deviations of
the mean from the grand average. The control limits for the subgroup sigmas
are plus and minus three standard deviations of sigma from the average sigma.
These control limits are quite robust with respect to non-normality in the process
To facilitate calculations, constants are used in the control limit equations.
Table 11 in the Appendix provides control chart constants for subgroups of 25
or less.
Control limit equations for sigma charts based on s-bar
s ?
sum of subgroup sigmas
number of subgroups ?12:11?
LCL ? B3s ?12:12?
UCL ? B4s ?12:13?
Control limit equations for averages charts based on s-bar
X ?
sum of subgroup averages
number of subgroups ?12:14?
LCL ? XX  A3s ?12:15?
UCL ? XX ? A3s ?12:16?
Example of averages and standard deviation control charts
To illustrate the calculations and to compare the range method to the standard
deviation results, the data used in the previous example will be reanalyzed
using the subgroup standard deviation rather than the subgroup range (Table
Statistical process control (SPC) 399
Table 12.2. Data for averages and sigma control charts.
Sample 1 Sample 2 Sample 3 Sample 4 Sample 5 AVERAGE SIGMA
110 93 99 98 109 101.8 7.396
103 95 109 95 98 100.0 6.000
97 110 90 97 100 98.8 7.259
96 102 105 90 96 97.8 5.848
105 110 109 93 98 103.0 7.314
110 91 104 91 101 99.4 8.325
100 96 104 93 96 97.8 4.266
93 90 110 109 105 101.4 9.290
90 105 109 90 108 100.4 9.607
103 93 93 99 96 96.8 4.266
97 97 104 103 92 98.6 4.930
103 100 91 103 105 100.4 5.550
90 101 96 104 108 99.8 7.014
97 106 97 105 96 100.2 4.868
99 94 96 98 90 95.4 3.578
106 93 104 93 99 99.0 6.042
90 95 98 109 110 100.4 8.792
96 96 108 97 103 100.0 5.339
109 96 91 98 109 100.6 8.081
90 95 94 107 99 97.0 6.442
91 101 96 96 109 98.6 6.804
108 97 101 103 94 100.6 5.413
96 97 106 96 98 98.6 4.219
101 107 104 109 104 105.0 3.082
96 91 96 91 105 95.8 5.718
The control limits are calculated from this data as follows:
Sigma control chart
s ?
sum of subgroup sigmas
number of subgroups ?
25 ? 6:218
LCLs ? B3s ? 0  6:218 ? 0
UCLs ? B4s ? 2:089  6:218 ? 12:989
Since it is not possible to have a subgroup sigma less than zero, the LCL is not
shown on the control chart for sigma for this example.
Averages control chart
X ?
sum of subgroup averages
number of subgroups ?
25 ? 99:5
LCLXX ? XX  A3s ? 99:5  1:427  6:218 ? 90:63
UCLXX ? XX ? A3s ? 99:5 ? 1:427  6:218 ? 108:37
The completed averages and sigma control charts are shown in Figure 12.2.
Note that the control limits for the averages chart are only slightly different
than the limits calculated using ranges.
Note that the conclusions reached are the same as when ranges were used.
Control charts for individual measurements (X charts)
Individuals control charts are statistical tools used to evaluate the central tendency
of a process over time. They are also called X charts or moving range
charts. Individuals control charts are used when it is not feasible to use averages
for process control. There are many possible reasons why averages control
charts may not be desirable: observations may be expensive to get (e.g., destructive
testing), output may be too homogeneous over short time intervals (e.g.,
pHof a solution), the production rate may be slow and the interval between successive
observations long, etc. Control charts for individuals are often used to
monitor batch process, such as chemical processes, where the within-batch variation
is so small relative to between-batch variation that the control limits on
Statistical process control (SPC) 401
a standard XX chart would be too close together. Range charts are used in conjunction
with individuals charts to help monitor dispersion.*
Figure 12.2. Completed averages and sigma control charts.
*There is considerable debate over the value ofmoving R charts. Academic researchers have failed to show statistical value in
them. However, many practitioners (including the author) believe thatmoving R charts provide valuable additional information
that can be used in troubleshooting.
Calculations for moving ranges charts
As with averages and ranges charts, the range is computed as shown above,
R?Largest in subgroup  Smallest in subgroup
Where the subgroup is a consecutive pair of process measurements. The
range control limit is computed as was described for averages and ranges charts,
using the D4 constant for subgroups of 2, which is 3.267. That is,
LCL ? 0 ?for n ? 2?
UCL ? 3:267  R-bar
Control limit equations for individuals charts
X ?
sum of measurements
number of measurements ?12:17?
LCL ? XX  E2 RR ? XX  2:66  RR ?12:18?
UCL ? XX ? E2 RR ? XX ? 2:66  RR ?12:19?
Where E2 ? 2:66 is the constant used when individual measurements are
plotted, and RR is based on subgroups of n ? 2.
Example of individuals and moving ranges control charts
Table 12.3 contains 25 measurements. To facilitate comparison, the
measurements are the first observations in each subgroup used in the previous
average/ranges and average/standard deviation control chart examples.
The control limits are calculated from this data as follows:
Moving ranges control chart control limits
R ?
sum of ranges
number of ranges ?
24 ? 8:17
LCLR ? D3 RR ? 0  8:17 ? 0
Statistical process control (SPC) 403
UCLR ? D4 RR ? 3:267  8:17 ? 26:69
Since it is not possible to have a subgroup range less than zero, the LCL is not
shown on the control chart for ranges.
Individuals control chart control limits
X ?
sum of measurements
number of measurements ?
25 ? 99:0
LCLX ? XX  E2 RR ? 99:0  2:66  8:17 ? 77:27
UCLX ? XX ? E2 RR ? 99:0 ? 2:66  8:17 ? 120:73
The completed individuals and moving ranges control charts are shown in
Figure 12.3.
Table 12.3. Data for individuals and moving ranges control charts.
Continued at right . . .
110 None
103 7
97 6
96 1
105 9
110 5
100 10
93 7
90 3
103 13
97 6
103 6
90 13
97 7
99 2
106 7
90 16
96 6
109 13
90 19
91 1
108 17
96 12
101 5
96 5
In this case, the conclusions are the same as with averages charts. However,
averages charts always provide tighter control than X charts. In some cases, the
additional sensitivity provided by averages charts may not be justified on either
an economic or an engineering basis. When this happens, the use of averages
charts will merely lead to wasting money by investigating special causes that
are of minor importance.
Statistical process control (SPC) 405
Figure 12.3. Completed individuals and moving ranges control charts.
Control charts for proportion defective ( pcharts)
p charts are statistical tools used to evaluate the proportion defective, or
proportion non-conforming, produced by a process.
p charts can be applied to any variable where the appropriate performance
measure is a unit count. p charts answer the question: ‘‘Has a special cause of
variation caused the central tendency of this process to produce an abnormally
large or small number of defective units over the time period observed?’’
p chart control limit equations
Like all control charts, p charts consist of three guidelines: center line, a lower
control limit, and an upper control limit. The center line is the average proportion
defective and the two control limits are set at plus and minus three standard
deviations. If the process is in statistical control, then virtually all proportions
should be between the control limits and they should fluctuate randomly
about the center line.
p ?
subgroup defective count
subgroup size ?12:20?
p ?
sum of subgroup defective counts
sum of subgroup sizes ?12:21?
LCL ? p  3 ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi p?1  p?
n r ?12:22?
UCL ? p ? 3 ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi p?1  p?
n r ?12:23?
In the above equations, n is the subgroup size. If the subgroup sizes varies, the
control limits will also vary, becoming closer together as n increases.
Analysis of pcharts
As with all control charts, a special cause is probably present if there are any
points beyond either the upper or the lower control limit. Analysis of p chart
patterns between the control limits is extremely complicated if the sample size
varies because the distribution of p varies with the sample size.
Example of pchart calculations
The data in Table 12.4 were obtained by opening randomly selected crates
from each shipment and counting the number of bruised peaches. There are
250 peaches per crate. Normally, samples consist of one crate per shipment.
However, when part-time help is available, samples of two crates are taken.
Statistical process control (SPC) 407
Table 12.4. Raw data for p chart.
1 1 250 47 0.188
2 1 250 42 0.168
3 1 250 55 0.220
4 1 250 51 0.204
5 1 250 46 0.184
6 1 250 61 0.244
7 1 250 39 0.156
8 1 250 44 0.176
9 1 250 41 0.164
10 1 250 51 0.204
11 2 500 88 0.176
12 2 500 101 0.202
13 2 500 101 0.202
14 1 250 40 0.160
15 1 250 48 0.192
16 1 250 47 0.188
17 1 250 50 0.200
18 1 250 48 0.192
19 1 250 57 0.228
20 1 250 45 0.180
21 1 250 43 0.172
22 2 500 105 0.210
23 2 500 98 0.196
24 2 500 100 0.200
25 2 500 96 0.192
TOTALS 8,000 1,544
Using the above data the center line and control limits are found as follows:
p ?
subgroup defective count
subgroup size
these values are shown in the last column of Table 12.4.
p ?
sum of subgroup defective counts
sum of subgroup sizes ?
8,000 ? 0:193
which is constant for all subgroups.
n ? 250 (1 crate):
LCL ? p  3 ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi p?1  p?
n r ? 0:193  3 ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi 0:193  ?1  0:193?
250 r ? 0:118
UCL ? p ? 3 ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi p?1  p?
n r ? 0:193 ? 3 ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi 0:193  ?1  0:193?
250 r ? 0:268
n ? 500 ?2 crates?:
LCL ? 0:193  3 ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi 0:193  ?1  0:193?
500 r ? 0:140
UCL ? 0:193 ? 3 ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi 0:193  ?1  0:193?
500 r ? 0:246
The control limits and the subgroup proportions are shown in Figure 12.24.
Pointers for using p charts
Determine if ‘‘moving control limits’’ are really necessary. It may be possible
to use the average sample size (total number inspected divided by the number
of subgroups) to calculate control limits. For instance, with our example the
sample size doubled from 250 peaches to 500 but the control limits hardly
changed at all. Table 12.5 illustrates the different control limits based on 250
peaches, 500 peaches, and the average sample size which is 8,000	25 = 320
Notice that the conclusions regarding process performance are the same
when using the average sample size as they are using the exact sample sizes.
This is usually the case if the variation in sample size isn’t too great. There are
many rules of thumb, but most of them are extremely conservative. The best
way to evaluate limits based on the average sample size is to check it out as
shown above. SPC is all about improved decision-making. In general, use the
most simple method that leads to correct decisions.
Control charts for count of defectives ( npcharts)
np charts are statistical tools used to evaluate the count of defectives, or count
of items non-conforming, produced by a process. np charts can be applied to
any variable where the appropriate performance measure is a unit count and
the subgroup size is held constant. Note that wherever an np chart can be used,
a p chart can be used too.
Statistical process control (SPC) 409
Figure 12.4. Completed p control chart.
Table 12.5. E?ect of using average sample size.
250 0.1181 0.2679
500 0.1400 0.2460
320 0.1268 0.2592
Control limit equations for npcharts
Like all control charts, np charts consist of three guidelines: center line, a
lower control limit, and an upper control limit. The center line is the average
count of defectives-per-subgroup and the two control limits are set at plus and
minus three standard deviations. If the process is in statistical control, then
virtually all subgroup counts will be between the control limits, and they will
fluctuate randomly about the center line.
np ? subgroup defective count ?12:24?
np ?
sum of subgroup defective counts
number of subgroups ?12:25?
LCL ? np  3 ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi np?1  p? p ?12:26?
UCL ? np ? 3 ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi np?1  p? p ?12:27?
Note that
p ?
n ?12:28?
Example of npchart calculation
The data in Table 12.6 were obtained by opening randomly selected crates
from each shipment and counting the number of bruised peaches. There are
250 peaches per crate (constant n is required for np charts).
Using the above data the center line and control limits are found as follows:
np ?
sum of subgroup defective counts
number of subgroups ?
30 ? 27:93
LCL ? np  3 ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi np?1  p? p ? 27:93  3 ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi 27:93  1 
 r ? 12:99
UCL ? np ? 3 ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi np?1  p? p ? 27:93 ? 3 ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi 27:93  1 
 r ? 42:88
The control limits and the subgroup defective counts are shown in Figure
Control charts for average occurrences-per-unit ( ucharts)
u charts are statistical tools used to evaluate the average number of occurrences-
per-unit produced by a process. u charts can be applied to any variable
where the appropriate performance measure is a count of how often a particular
event occurs. u charts answer the question: ‘‘Has a special cause of variation
caused the central tendency of this process to produce an abnormally large or
Statistical process control (SPC) 411
Table 12.6. Raw data for np chart.
1 20
2 28
3 24
4 21
5 32
6 33
7 31
8 29
9 30
10 34
11 32
12 24
13 29
14 27
15 37
Continued at right . . .
16 23
17 27
18 28
19 31
20 27
21 30
22 23
23 23
24 27
25 35
26 29
27 23
28 23
29 30
30 28
small number of occurrences over the time period observed?’’ Note that, unlike
p or np charts, u charts do not necessarily involve counting physical items.
Rather, they involve counting of events. For example, when using a p chart one
would count bruised peaches. When using a u chart one would count the bruises.
Control limit equations for ucharts
Like all control charts, u charts consist of three guidelines: center line, a lower
control limit, and an upper control limit. The center line is the average number
of occurrences-per-unit and the two control limits are set at plus and minus
three standard deviations. If the process is in statistical control then virtually
all subgroup occurrences-per-unit should be between the control limits and
they should fluctuate randomly about the center line.
u ?
subgroup count of occurrences
subgroup size in units ?12:29?
u ?
sum of subgroup occurrences
sum of subgroup sizes in units ?12:30?
LCL ? u  3 ffiffiffiu
n r ?12:31?
Figure 12.5. Completed np control chart.
UCL ? u ? 3 ffiffiffiu
n r ?12:32?
In the above equations, n is the subgroup size in units. If the subgroup size
varies, the control limits will also vary.
One way of helping determine whether or not a particular set of data is suitable
for a u chart or a p chart is to examine the equation used to compute the
center line for the control chart. If the unit of measure is the same in both the
numerator and the denominator, then a p chart is indicated, otherwise a u
chart is indicated. For example, if
Center Line ?
bruises per crate
number of crates
then the numerator is in terms of bruises while the denominator is in terms of
crates, indicating a u chart.
The unit size is arbitrary but once determined it cannot be changed without
recomputing all subgroup occurrences-per-unit and control limits. For example,
if the occurrences were accidents and a unit was 100,000 hours worked,
then a month with 250,000 hours worked would be 2.5 units and a month with
50,000 hours worked would be 0.5 units. If the unit size were 200,000 hours
then the two months would have 1.25 and 0.25 units respectively. The equations
for the center line and control limits would ‘‘automatically’’ take into account
the unit size, so the control charts would give identical results regardless of
which unit size is used.
Analysis of ucharts
As with all control charts, a special cause is probably present if there are any
points beyond either the upper or the lower control limit. Analysis of u chart
patterns between the control limits is extremely complicated when the sample
size varies and is usually not done.
Example of u chart
The data in Table 12.7 were obtained by opening randomly selected crates
from each shipment and counting the number of bruises on peaches. There are
250 peaches per crate. Our unit size will be taken as one full crate, i.e., we will
be counting crates rather than the peaches themselves. Normally, samples
consist of one crate per shipment. However, when part-time help is available,
samples of two crates are taken.
Statistical process control (SPC) 413
Table 12.7. Raw data for u chart.
1 1 47 47
2 1 42 42
3 1 55 55
4 1 51 51
5 1 46 46
6 1 61 61
7 1 39 39
8 1 44 44
9 1 41 41
10 1 51 51
11 2 88 44
12 2 101 50.5
13 2 101 50.5
14 1 40 40
15 1 48 48
16 1 47 47
17 1 50 50
18 1 48 48
19 1 57 57
20 1 45 45
21 1 43 43
22 2 105 52.5
23 2 98 49
24 2 100 50
25 2 96 48
TOTALS 32 1,544
Using the above data the center line and control limits are found as follows:
u ?
subgroup count of occurrences
subgroup size in units
These values are shown in the last column of Table 12.7.
u ?
sum of subgroup count of occurrences
sum of subgroup unit sizes ?
32 ? 48:25
which is constant for all subgroups.
n ? 1 unit:
LCL ? u  3 ffiffiffiu
n r ? 48:25  3 ffiffiffiffiffiffiffiffiffiffi 48:25
1 r ? 27:411
UCL ? u ? 3 ffiffiffiu
n r ? 48:25 ? 3 ffiffiffiffiffiffiffiffiffiffi 48:25
1 r ? 69:089
n ? 2 units:
LCL ? 48:25  3 ffiffiffiffiffiffiffiffiffiffi 48:25
2 r ? 33:514
UCL ? 48:25 ? 3 ffiffiffiffiffiffiffiffiffiffi 48:25
2 r ? 62:986
The control limits and the subgroup occurrences-per-unit are shown in
Figure 12.6.
The reader may note that the data used to construct the u chart were the same
as those used for the p chart, except that we considered the counts as being
counts of occurrences (bruises) instead of counts of physical items (bruised peaches).
The practical implications of using a u chart when a p chart should have
been used, or vice versa, are usually not serious. The decisions based on the control
charts will be quite similar in most cases encountered in Six Sigma regardless
of whether a u or a p chart is used.
Statistical process control (SPC) 415
Control charts for counts of occurrences-per-unit ( ccharts)
c charts are statistical tools used to evaluate the number of occurrences-perunit
produced by a process. c charts can be applied to any variable where the
appropriate performance measure is a count of how often a particular event
occurs and samples of constant size are used. c charts answer the question:
‘‘Has a special cause of variation caused the central tendency of this process to
produce an abnormally large or small number of occurrences over the time period
observed?’’ Note that, unlike p or np charts, c charts do not involve counting
physical items. Rather, they involve counting of events. For example, when
using an np chart one would count bruised peaches. When using a c chart one
would count the bruises.
Control limit equations for ccharts
Like all control charts, c charts consist of three guidelines: center line, a lower
control limit, and an upper control limit. The center line is the average number
of occurrences-per-unit and the two control limits are set at plus and minus
three standard deviations. If the process is in statistical control then virtually
all subgroup occurrences-per-unit should be between the control limits and
they should fluctuate randomly about the center line.
Figure 12.6. Completed u control chart.
c ?
sum of subgroup occurrences
number of subgroups ?12:33?
LCL ? c  3 ffiffic p
UCL ? c ? 3 ffiffic p
One way of helping determine whether or not a particular set of data is suitable
for a c chart or an np chart is to examine the equation used to compute the
center line for the control chart. If the unit of measure is the same in both the
numerator and the denominator, then a p chart is indicated, otherwise a c chart
is indicated. For example, if
Center Line ?
number of crates
then the numerator is in terms of bruises while the denominator is in terms of
crates, indicating a c chart.
The unit size is arbitrary but, once determined, it cannot be changed without
recomputing all subgroup occurrences-per-unit and control limits.
Analysis of ccharts
As with all control charts, a special cause is probably present if there are any
points beyond either the upper or the lower control limit. Analysis of c chart
patterns between the control limits is shown later in this chapter.
Example of cchart
The data in Table 12.8 were obtained by opening randomly selected crates
from each shipment and counting the number of bruises. There are 250 peaches
per crate. Our unit size will be taken as one full crate, i.e., we will be counting
crates rather than the peaches themselves. Every subgroup consists of one
crate. If the subgroup size varied, a u chart would be used.
Using the above data the center line and control limits are found as follows:
c ?
sum of subgroup occurrences
number of subgroups ?
30 ? 33:53
LCL ? c  3 ffiffic p
? 33:53  3 ffiffiffiffiffiffiffiffiffiffi 33:53 p ? 16:158
Statistical process control (SPC) 417
UCL ? c ? 3 ffiffiffiffi 8c p
? 33:53 ? 3 ffiffiffiffiffiffiffiffiffiffi 33:53 p ? 50:902
The control limits and the occurrence counts are shown in Figure 12.7.
Selecting the proper control chart for a particular data set is a simple matter if
approached properly. The proper approach is illustrated in Figure 12.8.
To use the decision tree, begin at the left-most node and determine if the data
are measurements or counts. If measurements, then select the control chart
based on the subgroup size. If the data are counts, then determine if the counts
are of occurrences or pieces. An aid in making this determination is to examine
Table 12.8. Raw data for c chart.
1 27
2 32
3 24
4 31
5 42
6 38
7 33
8 35
9 35
10 39
11 41
12 29
13 34
14 34
15 43
Continued at right . . .
16 29
17 33
18 33
19 38
20 32
21 37
22 30
23 31
24 32
25 42
26 40
27 21
28 23
29 39
30 29
the equation for the process average. If the numerator and denominator involve
the same units, then a p or np chart is indicated. If different units of measure
are involved, then a u or c chart is indicated. For example, if the average is in
accidents-per-month, then a c or u chart is indicated because the numerator is
in terms of accidents but the denominator is in terms of time.
Statistical process control (SPC) 419
Figure 12.7. Completed c control chart.
Figure 12.8. Control chart selection decision tree.
The basis of all control charts is the rational subgroup. Rational subgroups
are composed of items which were produced under essentially the same conditions.
The statistics, for example, the average and range, are computed for
each subgroup separately, then plotted on the control chart. When possible,
rational subgroups are formed by using consecutive units. Each subgroup’s
statistics are compared to the control limits, and patterns of variation between
subgroups are analyzed. Note the sharp contrast between this approach and
the random sampling approach used for enumerative statistical methods.
The idea of rational subgrouping becomes a bit fuzzy when dealing with x
charts, or individuals control charts. The reader may well wonder about the
meaning of the term subgrouping when the ‘‘subgroup’’ is a single measurement.
The basic idea underlying control charts of all types is to identify the
capability of the process. The mechanism by which this is accomplished is careful
formation of rational subgroups as defined above. When possible, rational
subgroups are formed by using consecutive units. The measure of process variability,
either the subgroup standard deviation or the subgroup range, is the
basis of the control limits for averages. Conceptually, this is akin to basing the
control limits on short-term variation. These control limits are used to monitor
variation over time.
As far as possible, this approach also forms the basis of establishing control
limits for individual measurements. This is done by forming quasi-subgroups
using pairs of consecutive measurements. These ‘‘subgroups of 2’’ are used to
compute ranges. The ranges are used to compute the control limits for the
individual measurements.
Control charts provide the operational definition of the term special cause. A
special cause is simply anything which leads to an observation beyond a control
limit. However, this simplistic use of control charts does not do justice to their
power. Control charts are running records of the performance of the process
and, as such, they contain a vast store of information on potential improvements.
While some guidelines are presented here, control chart interpretation
is an art that can only be developed by looking at many control charts and probing
the patterns to identify the underlying system of causes at work.
Freak patterns are the classical special cause situation (Figure 12.9). Freaks
result from causes that have a large effect but that occur infrequently. When
investigating freak values look at the cause and effect diagram for items that
meet these criteria. The key to identifying freak causes is timelines in collecting
and recording the data. If you have difficulty, try sampling more frequently.
Drift is generally seen in processes where the current process value is partly
determined by the previous process state. For example, if the process is a plating
bath, the content of the tank cannot change instantaneously, instead it will
change gradually (Figure 12.10). Another common example is tool wear: the
size of the tool is related to its previous size. Once the cause of the drift has
been determined, the appropriate action can be taken. Whenever economically
feasible, the drift should be eliminated, e.g., install an automatic chemical dispenser
for the plating bath, or make automatic compensating adjustments to
correct for tool wear. Note that the total process variability increases when
drift is allowed, which adds cost. When drift elimination is not possible, the control
chart can be modified in one of two ways:
1. Make the slope of the center line and control limits match the natural
process drift. The control chart will then detect departures from the
natural drift.
2. Plot deviations from the natural or expected drift.
Cycles often occur due to the nature of the process. Common cycles include
hour of the day, day of the week, month of the year, quarter of the year, week of
the accounting cycle, etc. (Figure 12.11). Cycles are caused by modifying the process
inputs or methods according to a regular schedule. The existence of this schedule
and its effect on the process may or may not be known in advance. Once
the cycle has been discovered, action can be taken. The action might be to adjust
Statistical process control (SPC) 421
Figure 12.9. Control chart patterns: freaks.
the control chart by plotting the control measure against a variable base. For
example, if a day-of-the-week cycle exists for shipping errors because of the workload,
you might plot shipping errors per 100 orders shipped instead of shipping
errors per day. Alternatively, it may be worthwhile to change the system to
smooth out the cycle. Most processes operate more efficiently when the inputs
are relatively stable and when methods are changed as little as possible.
Figure 12.10. Control chart patterns: drift.
Figure 12.11. Control chart patterns: cycles.
A controlled process will exhibit only ‘‘random looking’’ variation. A pattern
where every nth item is different is, obviously, non-random (Figure 12.12).
These patterns are sometimes quite subtle and difficult to identify. It is sometimes
helpful to see if the average fraction defective is close to some multiple of
a known number of process streams. For example, if the machine is a filler
with 40 stations, look for problems that occur 1/40, 2/40, 3/40, etc., of the time.
When plotting measurement data the assumption is that the numbers exist
on a continuum, i.e., there will be many different values in the data set. In the
real world, the data are never completely continuous (Figure 12.13). It usually
doesn’t matter much if there are, say, 10 or more different numbers. However,
when there are only a few numbers that appear over-and-over it can cause
Statistical process control (SPC) 423
Figure 12.12. Control chart patterns: repeating patterns.
Figure 12.13. Control chart patterns: discrete data.
problems with the analysis. A common problem is that the R chart will underestimate
the average range, causing the control limits on both the average
and range charts to be too close together. The result will be too many ‘‘false
alarms’’ and a general loss of confidence in SPC.
The usual cause of this situation is inadequate gage resolution. The ideal solution
is to obtain a gage with greater resolution. Sometimes the problem occurs
because operators, inspectors, or computers are rounding the numbers. The
solution here is to record additional digits.
The reason SPC is done is to accelerate the learning process and to eventually
produce an improvement. Control charts serve as historical records of the learning
process and they can be used by others to improve other processes. When
an improvement is realized the change should be written on the old control
chart; its effect will show up as a less variable process. These charts are also useful
in communicating the results to leaders, suppliers, customers, and others
interested in quality improvement (Figure 12.14).
Seemingly random patterns on a control chart are evidence of unknown
causes of variation, which is not the same as uncaused variation. There should
be an ongoing effort to reduce the variation from these so-called common causes.
Doing so requires that the unknown causes of variation be identified. One way
of doing this is a retrospective evaluation of control charts. This involves brainstorming
and preparing cause and effect diagrams, then relating the control
chart patterns to the causes listed on the diagram. For example, if ‘‘operator’’ is
a suspected cause of variation, place a label on the control chart points produced
Figure 12.14. Control chart patterns: planned changes.
by each operator (Figure 12.15). If the labels exhibit a pattern, there is evidence to
suggest a problem. Conduct an investigation into the reasons and set up controlled
experiments (prospective studies) to test any theories proposed. If the
experiments indicate a true cause and effect relationship, make the appropriate
process improvements. Keep in mind that a statistical association is not the
same thing as a causal correlation. The observed association must be backed up
with solid subject-matter expertise and experimental data.
Mixture exists when the data from two different cause systems are plotted on
a single control chart (Figure 12.16). It indicates a failure in creating rational
Statistical process control (SPC) 425
Figure 12.15. Control chart patterns: suspected di?erences.
Figure 12.16. Control chart patterns: mixture.
subgroups. The underlying differences should be identified and corrective
action taken. The nature of the corrective action will determine how the control
chart should be modified.
Mixture example #1
The mixture represents two different operators who can be made more
consistent. A single control chart can be used to monitor the new, consistent
Mixture example #2
The mixture is in the number of emergency room cases received on Saturday
evening, versus the number received during a normal week. Separate control
charts should be used to monitor patient-load during the two different time
Run tests
If the process is stable, then the distribution of subgroup averages will be
approximately normal. With this in mind, we can also analyze the patterns on
the control charts to see if they might be attributed to a special cause of variation.
To do this, we divide a normal distribution into zones, with each zone
one standard deviation wide. Figure 12.17 shows the approximate percentage
we expect to find in each zone from a stable process.
Zone C is the area from the mean to the mean plus or minus one sigma, zone
B is from plus or minus one sigma to plus or minus two sigma, and zone A is
from plus or minus two sigma to plus or minus three sigma. Of course, any
point beyond three sigma (i.e., outside of the control limit) is an indication of
an out-of-control process.
Since the control limits are at plus and minus three standard deviations, finding
the one and two sigma lines on a control chart is as simple as dividing the distance
between the grand average and either control limit into thirds, which can
be done using a ruler. This divides each half of the control chart into three
zones. The three zones are labeled A, B, and C as shown on Figure 12.18.
Based on the expected percentages in each zone, sensitive run tests can be
developed for analyzing the patterns of variation in the various zones.
Remember, the existence of a non-random pattern means that a special cause
of variation was (or is) probably present. The averages, np and c control chart
run tests are shown in Figure 12.19.
Note that, when a point responds to an out-of-control test it is marked
with an ‘‘X’’ to make the interpretation of the chart easier. Using this
convention, the patterns on the control charts can be used as an aid in troubleshooting.
Statistical process control (SPC) 427
Figure 12.17. Percentiles for a normal distribution.
Figure 12.18. Zones on a control chart.
Figure 12.19. Tests for out-of-control patterns on control charts.
From ‘‘The Shewhart Control ChartLTests for Special Causes,’’ Journal of Quality
Technology, 16(4), p. 238. Copyright #1986 by Nelson.
Tampering occurs when adjustments are made to a process that is in statistical
control. Adjusting a controlled process will always increase process variability,
an obviously undesirable result. The best means of diagnosing tampering is
to conduct a process capability study (see Chapter 13) and to use a control
chart to provide guidelines for adjusting the process.
Perhaps the best analysis of the effects of tampering is from Deming (1986).
Deming describes four common types of tampering by drawing the analogy of
aiming a funnel to hit a desired target. These ‘‘funnel rules’’ are described by
Deming (1986, p. 328):
1. ‘‘Leave the funnel ?xed, aimed at the target, no adjustment.’’
2. ‘‘At drop k(k ? 1, 2, 3, . . .) the marble will come to rest at point zk,
measured from the target. (In other words, zk is the error at drop k.)
Move the funnel the distance zk from the last position. Memory 1.’’
3. ‘‘Set the funnel at each drop right over the spot zk, measured from the
target. No memory.’’
4. ‘‘Set the funnel at each drop right over the spot (zk) where it last came to
rest. No memory.’’
Rule #1 is the best rule for stable processes. By following this rule, the process
average will remain stable and the variance will be minimized. Rule#2produces
a stable output but one with twice the variance of rule #1. Rule #3
results in a system that ‘‘explodes,’’ i.e., a symmetrical pattern will appear with
a variance that increases without bound. Rule#4 creates a pattern that steadily
moves away from the target, without limit (see figure 12.20).
At first glance, one might wonder about the relevance of such apparently
abstract rules. However, upon more careful consideration, one finds many practical
situations where these rules apply.
Rule #1 is the ideal situation and it can be approximated by using control
charts to guide decision-making. If process adjustments are made only when
special causes are indicated and identified, a pattern similar to that produced
by rule #1 will result.
Rule #2 has intuitive appeal for many people. It is commonly encountered
in such activities as gage calibration (check the standard once and adjust the
gage accordingly) or in some automated equipment (using an automatic gage,
check the size of the last feature produced and make a compensating adjustment).
Since the system produces a stable result, this situation can go unnoticed
indefinitely. However, as shown by Taguchi (1986), increased variance translates
to poorer quality and higher cost.
The rationale that leads to rule#3goes something like this: ‘‘A measurement
was taken and it was found to be 10 units above the desired target. This hap-
Statistical process control (SPC) 429
pened because the process was set 10 units too high. I want the average to equal
the target. To accomplish this I must try to get the next unit to be 10 units too
low.’’ This might be used, for example, in preparing a chemical solution. While
reasonable on its face, the result of this approach is a wildly oscillating system.
A common example of rule #4 is the ‘‘train-the-trainer’’ method. A master
spends a short time training a group of ‘‘experts,’’ who then train others, who
train others, etc. An example is on-the-job training. Another is creating a setup
by using a piece from the last job. Yet another is a gage calibration system
where standards are used to create other standards, which are used to create
still others, and so on. Just how far the final result will be from the ideal depends
on how many levels deep the scheme has progressed.
Short production runs are a way of life with many manufacturing companies.
In the future, this will be the case even more often. The trend in manufacturing
has been toward smaller production runs with product tailored to the specific
needs of individual customers. Henry Ford’s days of ‘‘the customer can have
any color, as long as it’s black’’ have long since passed.
Classical SPC methods, such as XX and R charts, were developed in the era of
mass production of identical parts. Production runs often lasted for weeks,
months, or even years. Many of the ‘‘SPC rules of thumb’’ currently in use
1-50 Rule #1 101-150 Rule #3
51-100 Rule #2 151-200 Rule #4
Figure 12.20. Funnel rule simulation results.
were created for this situation. For example, the rule that control limits not be
calculated until data are available from at least 25 subgroups of 5. This may not
have been a problem in 1930, but it certainly is today. In fact, many entire production
runs involve fewer parts than required to start a standard control chart!
Many times the usual SPC methods can be modified slightly to work with
short and small runs. For example, XX and R control charts can be created
using moving averages and moving ranges (Pyzdek, 1989). However, there
are SPC methods that are particularly well suited to application on short or
small runs.
Variables data, sometimes called continuous data, involve measurements
such as size, weight, pH, temperature, etc. In theory data are variables data if
no two values are exactly the same. In practice this is seldom the case. As a
rough rule of thumb you can consider data to be variables data if at least ten different
values occur and repeat values make up no more than 20% of the data
set. If this is not the case, your data may be too discrete to use standard control
charts. Consider trying an attribute procedure such as the demerit charts
described later in this chapter. We will discuss the following approaches to
SPC for short or small runs:
1. Exact methodLTables of special control chart constants are used to
create X, XX, and R charts that compensate for the fact that a limited number
of subgroups are available for computing control limits. The exact
method is also used to compute control limits when using a code value
chart or stabilized X or XX and R charts (see below). The exact method
allows the calculation of control limits that are correct when only a
small amount of data is available. As more data become available the
exact method updates control limits until, ?nally, no further updates
are required and standard control chart factors can be used (Pyzdek,
2. Code value chartsLControl charts created by subtracting nominal or
other target values from actual measurements. These charts are often
standardized so that measurement units are converted to whole numbers.
For example, if measurements are in thousandths of an inch a reading
of 0.011 inches above nominal would be recorded simply as ‘‘11.’’
Code value charts enable the user to plot several parts from a given process
on a single chart, or to plot several features from a single part on
the same control chart. The exact method can be used to adjust the control
limits when code value charts are created with limited data.
Statistical process control (SPC) 431
3. Stabilized control charts for variablesLStatisticians have known
about normalizing transformations for many years. This approach can
be used to create control charts that are independent of the unit of measure
and scaled in such a way that several di?erent characteristics can
be plotted on the same control chart. Since stabilized control charts are
independent of the unit of measure, they can be thought of as true
process control charts. The exact method adjusts the control limits for stabilized
charts created with limited data.
This procedure, adapted from Hillier (1969) and Proschan and Savage (1960),
applies to short runs or any situation where a small number of subgroups will
be used to set up a control chart. It consists of three stages:
1. ?nding the process (establishing statistical control);
2. setting limits for the remainder of the initial run; and
3. setting limits for future runs.
The procedure correctly compensates for the uncertainties involved when
computing control limits with small amounts of data.
Stage one: ?nd the process
1. Collect an initial sample of subgroups (g). The factors for the recommended
minimum number of subgroups are shown in Appendix Table
15 enclosed in a dark box. If it is not possible to get the minimum number
of subgroups, use the appropriate control chart constant for the number
of subgroups you actually have.
2. Using Table 15 compute the Range chart control limits using the equation
Upper Control Limit for Ranges (UCLR? ? D4F  RR. Compare
the subgroup ranges to the UCLR and drop any out-of-control groups.
Repeat the process until all remaining subgroup ranges are smaller than
3. Using the RR value found in step #2, compute the control limits
for the averages or individuals chart. The control limits are found
by adding and subtracting A2F  RR from the overall average. Drop
any subgroups that have out-of-control averages and recompute.
Continue until all remaining values are within the control limits.
Go to stage two.
Stage two: set limits for remainder of the initial run
1. Using Table 15 compute the control limits for the remainder of the run.
Use the A2S factors for the XX chart and the D4S factors for the R chart;
g ? the number of groups used to compute stage one control limits.
Stage three: set limits for a future run
1. After the run is complete, combine the raw data from the entire run and
perform the analysis as described in stage one above. Use the results of
this analysis to set limits for the next run, following the stage two procedure.
If more than 25 groups are available, use a standard table of control
chart constants.
1. Stage three assumes that there are no special causes of variation between
runs. If there are, the process may go out of control when using the
stage three control limits. In these cases, remove the special causes. If
this isn’t possible, apply this procedure to each run separately (i.e., start
over each time).
2. This approach will lead to the use of standard control chart tables when
enough data are accumulated.
3. The control chart constants for the ?rst stage are A2F and D4F (the ‘‘F’’
subscript stands for First stage); for the second stage use A2S and D4S.
These factors correspond to the A2 and D4 factors usually used, except
that they are adjusted for the small number of subgroups actually available.
Setup approval procedure
The following procedure can be used to determine if a setup is acceptable
using a relatively small number of sample units.
1. After the initial setup, run 3 to 10 pieces without adjusting the process.
2. Compute the average and the range of the sample.
3. Compute T ?
average  target
Use absolute values (i.e., ignore any minus signs). The target value is
usually the speci?cation midpoint or nominal.
Statistical process control (SPC) 433
4. If T is less than the critical T in Table 12.9 accept the setup. Otherwise
adjust the setup to bring it closer to the target. NOTE: there is approximately
1 chance in 20 that an on-target process will fail this test.
Assume we wish to use SPC for a process that involves producing a part in
lots of 30 parts each. The parts are produced approximately once each month.
The control feature on the part is the depth of a groove and we will be measuring
every piece. We decide to use subgroups of size three and to compute the stage
one control limits after the first five groups. The measurements obtained are
shown in Table 12.10.
Using the data in Table 12.10 we can compute the grand average and average
range as
Grand average ? 0:10053
Average range ?RR? ? 0:00334
Table 12.9. Critical value for setup acceptance.
n 3 4 5 6 7 8 9 10
Critical T 0.885 0.529 0.388 0.312 0.263 0.230 0.205 0.186
Table 12.10. Raw data for example of exact method.
1 2 3
1 0.0989 0.0986 0.1031 0.1002 0.0045
2 0.0986 0.0985 0.1059 0.1010 0.0074
3 0.1012 0.1004 0.1000 0.1005 0.0012
4 0.1023 0.1027 0.1000 0.1017 0.0027
5 0.0992 0.0997 0.0988 0.0992 0.0009
From Appendix Table 15 we obtain the first stage constant for the range
chart of D4F ?2.4 in the row for g ?5 groups and a subgroup size of 3. Thus,
UCLR ? D4F  RR ? 2:4  0:00334 ? 0:0080
All of the ranges are below this control limit, so we can proceed to the analysis
of the averages chart. If any R was above the control limit, we would try to
determine why before proceeding.
For the averages chart we get
LCLXX ? grand average  A2F  RR
? 0:10053  1:20  0:00334 ? 0:09652 (rounded)
UCLXX ? grand average ? A2F  RR
? 0:10053 ? 1:20  0:00334 ? 0:10454 (rounded)
All of the subgroup averages are between these limits. Now setting limits for
the remainder of the run we use D4S ? 3:4 and A2S ? 1:47. This gives, after
UCLR ? 0:01136
LCLXX ? 0:09562
UCLXX ? 0:10544
If desired, this procedure can be repeated when a larger number of subgroups
becomes available, say 10 groups. This would provide somewhat better estimates
of the control limits, but it involves considerable administrative overhead.
When the entire run is finished you will have 10 subgroups of 3 per
subgroup. The data from all of these subgroups should be used to compute
stage one and stage two control limits. The resulting stage two control limits
would then be applied to the next run of this part number.
By applying this method in conjunction with the code value charts or
stabilized charts described below, the control limits can be applied to the next
parts produced on this process (assuming the part-to-part difference can be
made negligible). Note that if the standard control chart factors were used the
limits for both stages would be (values are rounded)
UCLR ? 0:00860
LCLXX ? 0:09711
UCLXX ? 0:10395
Statistical process control (SPC) 435
As the number of subgroups available for computing the control limits
increases, the ‘‘short run’’ control limits approach the standard control limits.
However, if the standard control limits are used when only small amounts of
data are available there is a greater chance of erroneously rejecting a process
that is actually in control (Hillier, 1969).
This procedure allows the control of multiple features with a single control
chart. It consists of making a simple transformation to the data, namely
^x ?
X  Target
unit of measure ?12:36?
The resulting ^x values are used to compute the control limits and as plotted
points on the XX and R charts. This makes the target dimension irrelevant for
the purposes of SPC and makes it possible to use a single control chart for
several different features or part numbers.
A lathe is used to produce several different sizes of gear blanks, as is indicated
in Figure 12.21.
Product engineering wants all of the gear blanks to be produced as near as
possible to their nominal size. Process engineering believes that the process
will have as little deviation for larger sizes as it does for smaller sizes. Quality
engineering believes that the inspection system will produce approximately
the same amount of measurement error for larger sizes as for smaller sizes.
Process capability studies and measurement error studies confirm these
assumptions. (I hope you are starting to get the idea that a number of assumptions
are being made and that they must be valid before using code value charts.)
Based on these conclusions, the code value chart is recommended. By using
the code value chart the amount of paperwork will be reduced and more data
will be available for setting control limits.Also, the process history will be easier
to follow since the information won’t be fragmented among several different
charts. The data in Table 12.11 show some of the early results.
Note that the process must be able to produce the tightest tolerance of
0:0005 inches. The capability analysis should indicate its ability to do this;
i.e., Cpk should be at least 1.33 based on the tightest tolerance. It will not be
allowed to drift or deteriorate when the less stringently toleranced parts are produced.
Process control is independent of the product requirements. Permitting
Statistical process control (SPC) 437
Figure 12.21. Some of the gear blanks to be machined.
Table 12.11. Deviation from target in hundred-thousandths.
X R 1 2 3
A 1.0000 1 4 3 25 10.7 22
2 3 3 39 15.0 36
3 16 12 10 12.7 6
B 0.5000 4 21 24 10 18.3 14
5 6 8 4 6.0 4
6 19 7 21 15.7 14
C 2.0000 7 1 11 4 5.3 10
8 1 25 8 11.3 24
9 6 8 7 7.0 2
the process to degrade to its worst acceptable level (from the product perspective)
creates engineering nightmares when the more tightly toleranced parts
come along again. It also confuses and demoralizes operators and others trying
to maintain high levels of quality. In fact, it may be best to publish only the process
performance requirements and to keep the product requirements secret.
The control chart of the data in Table 12.11 is shown in Figure 12.22. Since
only nine groups were available, the exact method was used to compute the control
limits. Note that the control chart shows the deviations on the XX and R
chart axes, not the actual measured dimensions, e.g., the value of Part A, subgroup
#1, sample #1 was +0.00004" from the target value of 1.0000" and it is
shown as a deviation of +4 hundred-thousandths; i.e., the part checked
1.00004". The stage one control chart shows that the process is obviously in statistical
control, but it is producing parts that are consistently too large regardless
of the nominal dimension. If the process were on target, the grand average
would be very close to 0. The setup problem would have been detected by the
second subgroup if the setup approval procedure described earlier in this chapter
had been followed.
Figure 12.22. Code value chart of Table 12.11 data.
This ability to see process performance across different part numbers is one
of the advantages of code value charts. It is good practice to actually identify
the changes in part numbers on the charts, as is done in Figure 12.22.
All control limits, for standard sized runs or short and small runs, are based
on methods that determine if a process statistic falls within limits that might
be expected from chance variation (common causes) alone. In most cases, the
statistic is based on actual measurements from the process and it is in the same
unit of measure as the process measurements. As we saw with code value charts,
it is sometimes useful to transform the data in some way. With code value charts
we used a simple transformation that removed the effect of changing nominal
and target dimensions. While useful, this approach still requires that all measurements
be in the same units of measurement, e.g., all inches, all grams, etc.
For example, all of the variables on the control chart for the different gear
blanks had to be in units of hundred-thousandths of an inch. If we had also
wanted to plot, for example, the perpendicularity of two surfaces on the gear
blank we would have needed a separate control chart because the units would
be in degrees instead of inches.
Stabilized control charts for variables overcome the units of measure problem
by converting all measurements into standard, non-dimensional units.
Such ‘‘standardizing transformations’’ are not new, they have been around for
many years and they are commonly used in all types of statistical analyses. The
two transformations we will be using here are shown in Equations 12.37 and
? XX  grand average?
R ?12:37?
R ?12:38?
As you can see, Equation 12.37 involves subtracting the grand average from
each subgroup average (or from each individual measurement if the subgroup
size is one) and dividing the result by RR.Note that this is not the usual statistical
transformation where the denominator is s. By using RR as our denominator
instead of s we are sacrificing some desirable statistical properties such as normality
and independence to gain simplicity. However, the resulting control
charts remain valid and the false alarm risk based on points beyond the control
limits is identical to standard control charts. Also, as with all transformations,
Statistical process control (SPC) 439
this approach suffers in that it involves plotting numbers that are not in the
usual engineering units people are accustomed to working with. This makes it
more difficult to interpret the results and spot data entry errors.
Equation 12.38 divides each subgroup range by the average range. Since the
numerator and denominator are both in the same unit of measurement, the
unit of measurement cancels and we are left with a number that is in terms of
the number of average ranges, R’s. It turns out that control limits are also in
the same units, i.e., to compute standard control limits we simply multiply R
by the appropriate table constant to determine the width between the control
Hillier (1969) noted that this is equivalent to using the transformations
shown in Equations 12.37 and 12.38 with control limits set at
A2  ? XX  grand average?
R  A2 ?12:39?
for the individuals or averages chart. Control limits are
R  D4 ?12:40?
for the range chart. Duncan (1974) described a similar transformation for attribute
charts, p charts in particular (see below), and called the resulting chart a
‘‘stabilized p chart.’’ We will call charts of the transformed variables data stabilized
charts as well.
Stabilized charts allow you to plot multiple units of measurement on the
same control chart. The procedure described in this chapter for stabilized variables
charts requires that all subgroups be of the same size.* The procedure for
stabilized attribute charts, described later in this chapter allows varying subgroup
sizes. When using stabilized charts the control limits are always fixed.
The raw data are ‘‘transformed’’ to match the scale determined by the control
limits. When only limited amounts of data are available, the constants in
Appendix Table 15 should be used for computing control limits for stabilized
variables charts. As more data become available, the Appendix Table 11 constants
approach the constants in standard tables of control chart factors. Table
12.12 summarizes the control limits for stabilized averages, stabilized ranges,
and stabilized individuals control charts. The values for A2, D3, and D4 can be
found in standard control chart factor tables.
* The procedure for stabilized attribute charts, described later in this chapter, allows varying subgroup sizes.
A circuit board is produced on an electroplating line. Three parameters are
considered important for SPC purposes: lead concentration of the solder plating
bath, plating thickness, and resistance. Process capability studies have been
done using more than 25 groups; thus, based on Table 12.12 the control limits are
A2  XX  A2
for the averages control chart, and
D3  R  D4
for the ranges control chart. The actual values of the constants A2, D3, and D4
depend on the subgroup size; for subgroups of three A2 ? 1:023, D3 ? 0 and
D4 ? 2:574.
Statistical process control (SPC) 441
Table 12.12. Control limits for stabilized charts.
One 25 or less LCL ^A2F None ^A2F 15
Average 0 1 0
Two 25 or less LCL ^A2S None ^A2S 15
Average 0 1 0
More than
LCL ^A2 D3 ^2.66 11
Average 0 1 0
UCL+A 2 D4 +2.66
The capabilities are shown in Table 12.13.
A sample of three will be taken for each feature. The three lead concentration
samples are taken at three different locations in the tank. The results of one
such set of sample measurements is shown in Table 12.14, along with their stabilized
Onthe control chart only the extreme values are plotted. Figure 12.23 shows a
stabilized control chart for several subgroups. Observe that the feature responsible
for the plotted point is written on the control chart. If a long series of
Table 12.13. Process capabilities for example.
A Lead % 10% 1%
B Plating thickness 0.005" 0.0005"
C Resistance 0.1 0.0005
Table 12.14. Sample data for example.
1 11% 0.0050" 0.1000
2 11% 0.0055" 0.1010
3 8% 0.0060" 0.1020
 X 10% 0.0055" 0.1010
R 3% 0.0010" 0.0020
?x  x?= RR 0 1 2
R= RR 3 2 4
largest or smallest values comes from the same feature it is an indication that the
feature has changed. If the process is in statistical control for all features, the feature
responsible for the extreme values will vary randomly.
When using stabilized charts it is possible to have a single control chart
accompany a particular part or lot of parts through the entire production
sequence. For example, the circuit boards described above could have a control
chart that shows the results of process and product measurement for characteristics
at all stages of production. The chart would then show the ‘‘processing
history’’ for the part or lot. The advantage would be a coherent log of the
production of a given part. Table 12.15 illustrates a process control plan that
could possibly use this approach.
A caution is in order if the processing history approach is used. When small
and short runs are common, the history of a given process can be lost among
the charts of many different parts. This can be avoided by keeping a separate
chart for each distinct process; additional paperwork is involved, but it might
be worth the effort. If the additional paperwork burden becomes large, computerized
solutions may be worth investigating.
When data are difficult to obtain, as is usual when small or short runs are
involved, variables SPC should be used if at all possible. A variables
Statistical process control (SPC) 443
Figure 12.23. Stabilized control chart for variables.
measurement on a continuous scale contains more information than a discrete
attributes classification provides. For example, a machine is cutting a
piece of metal tubing to length. The specifications call for the length to be
between 0.990" and 1.010" with the preferred length being 1.000" exactly.
There are two methods available for checking the process. Method #1
involves measuring the length of the tube with a micrometer and recording
the result to the nearest 0.001". Method #2 involves placing the finished
part into a ‘‘go/no-go gage.’’ With method #2 a part that is shorter than
0.990" will go into the ‘‘no-go’’ portion of the gage, while a part that is
longer than 1.010" will fail to go into the ‘‘go’’ portion of the gage. With
method #1 we can determine the size of the part to within 0.001". With
method #2 we can only determine the size of the part to within 0.020";
i.e., either it is within the size tolerance, it’s too short, or it’s too long. If
the process could hold a tolerance of less than 0.020", method #1 would
provide the necessary information to hold the process to the variability it is
capable of holding. Method #2 would not detect a process drift until out
of tolerance parts were actually produced.
Table 12.15. PWB fab process capabilities and SPC plan.
Clean Bath pH 7.5 0.1 3/hr
Rinse contamination 100 ppm 5 ppm 3/hr
Cleanliness quality rating 78 4 3 pcs/hr
Laminate Riston thickness 1.5 min. 0.1mm 3 pcs/hr
Adhesion 7 in.^lbs. 0.2 in.^lbs. 3 pcs/hr
Plating Bath lead % 10% 1% 3/hr
Thickness 0.005" 0.0005" 3 pcs/hr
Resistance 0.1 0.0005 3 pcs/hr
Another way of looking at the two different methods is to consider each part
as belonging to a distinct category, determined by the part’s length. Method
#1allows any part that is within tolerance to be placed into one of twenty categories.
When out of tolerance parts are considered, method #1 is able to place
parts into even more than twenty different categories. Method #1 also tells us
if the part is in the best category, namely within 0.001" of 1.000"; if not, we
know how far the part is from the best category. With method #2 we can
place a given part into only three categories: too short, within tolerance, or too
long. A part that is far too short will be placed in the same category as a part
that is only slightly short. A part that is barely within tolerance will be placed
in the same category as a part that is exactly 1.000" long.
In spite of the disadvantages, it is sometimes necessary to use attributes data.
Special methods must be used for attributes data used to control short run processes.
We will describe two such methods:
. Stabilized attribute control charts.
. Demerit control charts.
When plotting attribute data statistics from short run processes two difficulties
are typically encountered:
1. Varying subgroup sizes.
2. A small number of subgroups per production run.
Item #1 results in messy charts with different control limits for each subgroup,
distorted chart scales that mask significant variations, and chart patterns
that are difficult to interpret because they are affected by both sample size
changes and true process changes. Item #2makes it difficult to track long-term
process trends because the trends are broken up among many different control
charts for individual parts. Because of these things, many people believe that
SPC is not practical unless large and long runs are involved. This is not the
case. In many cases stabilized attribute charts can be used to eliminate these
problems. Although somewhat more complicated than classical control charts,
stabilized attribute control charts offer a way of realizing the benefits of SPC
with processes that are difficult to control any other way.
Stabilized attribute charts may be used if a process is producing part features
that are essentially the same from one part number to the next. Production lot
sizes and sample sizes can vary without visibly affecting the chart.
Statistical process control (SPC) 445
Example one
A lathe is being used to machine terminals of different sizes. Samples (of different
sizes) are taken periodically and inspected for burrs, nicks, tool marks,
and other visual defects.
Example two
A printed circuit board hand assembly operation involves placing electrical
components into a large number of different circuit boards. Although the
boards differ markedly from one another, the hand assembly operation is
similar for all of the different boards.
Example three
A job-shop welding operation produces small quantities of ‘‘one order only’’
items. However, the operation always involves joining parts of similar material
and similar size. The process control statistic is weld imperfections per 100
inches of weld.
The techniques used to create stabilized attribute control charts are all based
on corresponding classical attribute control chart methods. There are four
basic types of control charts involved:
1. Stabilized p charts for proportion of defective units per sample.
2. Stabilized np charts for the number of defective units per sample.
3. Stabilized c charts for the number of defects per unit.
4. Stabilized u charts for the average number of defects per unit.
All of these charts are based on the transformation
Z ?
sample statistic  process average
process standard deviation ?12:41?
In other words, stabilized charts are plots of the number of standard deviations
(plus or minus) between the sample statistic and the long-term process
average. Since control limits are conventionally set at 3 standard deviations,
stabilized control charts always have the lower control limit at 3 and the
upper control limit at +3. Table 12.16 summarizes the control limit equations
for stabilized control charts for attributes.
When applied to long runs, stabilized attribute charts are used to compensate
for varying sample sizes; process averages are assumed to be constant.
However, stabilized attribute charts can be created even if the process average
varies. This is often done when applying this technique to short runs of parts
that vary a great deal in average quality. For example, a wave soldering process
used for several missiles had boards that varied in complexity from less than 100
solder joints to over 1,500 solder joints. Tables 12.17 and 12.18 show how the
situation is handled to create a stabilized u chart. The unit size is 1,000 leads,
set arbitrarily. It doesn’t matter what the unit size is set to, the calculations
will still produce the correct result since the actual number of leads is divided
by the unit size selected. u is the average number of defects per 1,000 leads.
Example four
From the process described in Table 12.17, 10 TOW missile boards of type E
are sampled. Three defects were observed in the sample. Using Tables 12.16
and 12.17 Z is computed for the subgroup as follows:
 ? ffiffiffiffiffiffiffi u=n p ; we get u ? 2 from Table 12.17.
n ?
50  10
1,000 ? 0:5 units
 ? ffiffiffiffiffiffiffiffiffiffi 2=0:5 p ? ffiffi4 p ? 2
u ?
number of defects
number of units ?
0:5 ? 6 defects per unit
Statistical process control (SPC) 447
Table 12.16. Stabilized attribute chart statistics.
Proportion of
defective units
p chart p p ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi p?1  p?=n p ?p  p?=
Number of
defective units
np chart np np ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi np?1  p? p ?np  np?=
Defects per unit c chart c c ffiffic p ?c  c?=
Average defects
per unit
u chart u u ffiffiffiffiffiffiffi u=n p ?u  u?=
Z ?
u  u
6  2
2 ?
2 ? 2
Since Z is between ^3 and +3 we conclude that the process has not gone out
of control; i.e., it is not being influenced by a special cause of variation.
Table 12.18 shows the data for several samples from this process. The resulting
control chart is shown in Figure 12.24. Note that the control chart indicates
that the process was better than average when it produced subgroups 2 and 3
and perhaps 4. Negative Z values mean that the defect rate is below (better
than) the long-term process average. Groups 7 and 8 show an apparent
deterioration in the process with group 7 being out of control. Positive Z values
indicate a defect rate above (worse than) the long-term process average.
The ability to easily see process trends and changes like these in spite of
changing part numbers and sample sizes is the big advantage of stabilized control
charts. The disadvantages of stabilized control charts are:
1. They convert a number that is easy to understand, the number of defects
or defectives, into a confusing statistic with no intuitive meaning.
2. They involve tedious calculation.
Item #1 can only be corrected by training and experience applying the technique.
Item #2 can be handled with computers; the calculations are simple to
perform with a spreadsheet. Table 12.18 can be used as a guide to setting up the
spreadsheet. Inexpensive handheld computers can be used to perform the calculations
right at the process, thus making the results available immediately.
Table 12.17. Data from a wave solder process.
Phoenix A 1,650 1.65 16
B 800 0.80 9
C 1,200 1.20 9
TOW D 80 0.08 4
E 50 0.05 2
F 100 0.10 1
As described above, there are two kinds of data commonly used to perform
SPC: variables data and attributes data. When short runs are involved we can
seldom afford the information loss that results from using attribute data.
However, the following are ways of extracting additional information from
attribute data:
1. Making the attribute data ‘‘less discrete’’ by adding more classi?cation
2. Assigning weights to the categories to accentuate di?erent levels of
Consider a process that involves fabricating a substrate for a hybrid microcircuit.
The surface characteristics of the substrate are extremely important.
The ‘‘ideal part’’ will have a smooth surface, completely free of any visible
Statistical process control (SPC) 449
Table 12.18. Stabilized u chart data for wave solder.
1 E 2 0.05 10 0.50 2.00 3 6.00 2.00
2 A 16 1.65 1 1.65 3.11 8 4.85 ^3.58
3 A 16 1.65 1 1.65 3.11 11 6.67 ^3.00
4 B 9 0.80 1 0.80 3.35 0 0.00 ^2.68
5 F 1 0.10 2 0.20 2.24 1 5.00 1.79
6 E 2 0.05 5 0.25 2.83 2 8.00 2.12
7 C 9 1.20 1 1.20 2.74 25 20.83 4.32
8 D 4 0.08 5 0.40 3.16 5 12.50 2.69
9 B 9 0.80 1 0.80 3.35 7 8.75 ^0.07
10 B 9 0.80 1 0.80 3.35 7 8.75 ^0.07
flaws or blemishes. However, parts are sometimes produced with stains, pits,
voids, cracks and other surface defects. Although undesirable, most of the less
than ideal parts are still acceptable to the customer.
If we were to apply conventional attribute SPC methods to this process the
results would probably be disappointing. Since very few parts are actually
rejected as unacceptable, a standard p chart or stabilized p chart would probably
show a flat line at ‘‘zero defects’’ most of the time, even though the quality
level might be less than the target ideal part. Variables SPC methods can’t be
used because attributes data such as ‘‘stains’’ are not easily measured on a variables
scale. Demerit control charts offer an effective method of applying SPC
in this situation.
To use demerit control charts we must determine how many imperfections
of each type are found in the parts. Weights are assigned to the different categories.
The quality score for a given sample is the sum of the weights times the
frequencies of each category. Table 12.19 illustrates this approach for the substrate
If the subgroup size is kept constant, the average for the demerit control chart
is computed as follows (Burr, 1976),
Average ? DD ?
sum of subgroup demerits
number of subgroups ?12:42?
Figure 12.24. Control chart of Z values from Table 12.18.
Control limits are computed in two steps. First compute the weighted average
defect rate for each category. For example, there might be the following
categories and weights
Major 10
Minor 5
Incidental 1
Three average defect rates, one each for major, minor, and incidental, could
be computed using the following designations:
c1 ? Average number of major defects per subgroup
c2 ? Average number of minor defects per subgroup
c3 ? Average number of incidental defects per subgroup
Statistical process control (SPC) 451
Table 12.19. Demerit scores for substrates.
Attribute Weight Freq. Score Freq. Score Freq. Score
Light stain 1 3 3
Dark stain 5 1 5 1 5
Small blister 1 2 2 1 1
Medium blister 5 1 5
Pit: 0.01^0.05 mm 1 3 3
Pit: 0.06^0.10 mm 5 2 10
Pit: larger than 0.10 mm 10 1 10
The corresponding weights might be W1 ? 10, W2 ? 5, W3 ? 1. Using this
notation we compute the demerit standard deviation for this three category
example as
D ? ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi W2
1 c1 ? W2
2 c2 ? W2
3 c3 p ?12:43?
For the general case the standard deviation is
D ? ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi Xk
i ci vuut
The control limits are
LCL ? DD  3D ?12:45?
UCL ? DD ? 3D ?12:46?
If the Lower Control Limit is negative, it is set to zero.
The above procedure, while correct, may sometimes be too burdensome to
implement effectively. When this is the case a simplified approach may be
used. The simplified approach is summarized as follows:
1. Classify each part in the subgroup into the following classes (weights are
A Preferred quality. All product features at or very near
B Acceptable quality. Some product features have
departed signi?cantly from target quality levels, but
they are a safe distance from the reject limits.
C Marginal quality. One or more product features are in
imminent danger of exceeding reject limits.
D Reject quality. One or more product features fail to
meet minimum acceptability requirements.
2. Plot the total scores for each subgroup, keeping the subgroup sizes
3. Treat the total scores as if they were variables data and prepare an individuals
and moving range control chart or an XX and R chart. These charts
are described in Pyzdek (1989) and in most texts on SPC.
Small runs and short runs are common in modern business environments.
Different strategies are needed to deal with these situations. Advance planning
is essential. Special variables techniques were introduced which compensate
for small sample sizes and short runs by using special tables or mathematically
transforming the statistics and charts. Attribute short run SPC methods were
introduced that make process patterns more evident when small runs are
produced. Demerit and scoring systems were introduced that extract more
information from attribute data.
EWMA charts
Many people erroneously believe that statistics are not needed when automated
manufacturing processes are involved. Since we have measurements
from every unit produced, they reason, sampling methods are inappropriate.
We will simply correct the process when the characteristic is not on target.
This attitude reflects a fundamental misunderstanding of the relationship
between a process and the output of a process. It also shows a lack of appreciation
for the intrinsic variability of processes and of measurements. The fact is,
even if you have a ‘‘complete’’ data record of every feature of every part produced,
you still have only a sample of the output of the process. The process is
future-oriented in time, while the record of measurements is past-oriented.
Unless statistical control is attained, you will be unable to use the data from
past production to predict the variability from the process in the future (refer
to the definition of control in page 321). And without statistical tools you have
no sound basis for the belief that statistical control exists.
Another reason process control should be based on an understanding and
correct use of statistical methods is the effect of making changes without this
understanding. Consider, for example, the following process adjustment rule:
EWMA 453
Measure the diameter of the gear shaft. If the diameter is above the
nominal size, adjust the process to reduce the diameter. If the diameter is
below the nominal size, adjust the process to increase the diameter.
The problem with this approach is described by Deming’s ‘‘funnel rules’’ (see
above). This approach to process control will increase the variability of a statistically
controlled process by 141%, certainly not what the process control analyst
had in mind. The root of the problem is a failure to realize that the part
measurement is a sample from the process and, although it provides information
about the state of the process, the information is incomplete. Only through
using proper statistical methods can the information be extracted, analyzed
and understood.
Afundamental assumption underlying traditional SPC techniques is that the
observed values are independent of one another. Although the SPC tools are
quite insensitive to moderate violations of this assumption (Wheeler, 1991),
automated manufacturing processes often breach the assumption by enough
to make traditional methods fail (Alwan and Roberts, 1989). By using scatter
diagrams, as described in Chapter 14, you can determine if the assumption of
independence is satisfied for your data. If not, you should consider using the
methods described below instead of the traditional SPC methods.
A common complaint about non-standard SPC methods is that they are
usually more complex than the traditional methods (Wheeler, 1991). This is
often true. However, when dealing with automated manufacturing processes
the analysis is usually handled by a computer. Since the complexity of the analysis
is totally invisible to the human operator, it makes little difference. Of
course, if the operator will be required to act based on the results, he or she
must understand how the results are to be used. The techniques described in
this chapter which require human action are interpreted in much the same way
as traditional SPC techniques.
When using traditional SPC techniques the rules are always the same,
1. As long as the variation in the statistic being plotted remains within the
control limits, leave the process alone.
2. If a plotted point exceeds a control limit, look for the cause.
This approach works fine as long as the process remains static. However, the
means of many automated manufacturing processes often drift because of
inherent process factors. In other words, the drift is produced by common
causes. In spite of this, there may be known ways of intervening in the process
to compensate for the drift. Traditionalists would say that the intervention
should be taken in such a way that the control chart exhibits only random variation.
However, this may involve additional cost. Mindlessly applying arbitrary
rules to achieve some abstract result, like a stable control chart, is poor practice.
All of the options should be considered.
One alternative is to allow the drift to continue until the cost of intervention
equals the cost of running off-target. This alternative can be implemented
through the use of a ‘‘common cause chart.’’ This approach, described in
Alwan and Roberts (1989) and Abraham and Whitney (1990), involves creating
a chart of the process mean. However, unlike traditional XX charts, there are no
control limits. Instead, action limits are placed on the chart. Action limits differ
from control limits in two ways
. They are computed based on costs rather than on statistical theory.
. Since the chart shows variation from common causes, violating an action
limit does not result in a search for a special cause. Instead, a prescribed
action is taken to bring the process closer to the target value.
These charts are called ‘‘common cause charts’’ because the changing level of
the process is due to built-in process characteristics. The process mean is
tracked by using exponentially weighted moving averages (EWMA). While
somewhat more complicated than traditional XX charts, EWMA charts have a
number of advantages for automated manufacturing:
. They can be used when processes have inherent drift.
. EWMA charts provide a forecast of where the next process measurement
will be. This allows feed-forward control.
. EWMA models can be used to develop procedures for dynamic process
control, as described later in this section.
When dealing with a process that is essentially static, the predicted value of
the average of every sample is simply the grand average. EWMA charts, on the
other hand, use the actual process data to determine the predicted process
value for processes that may be drifting. If the process has trend or cyclical components,
the EWMA will reflect the effect of these components. Also, the
EWMA chart produces a forecast of what the next sample mean will be; the
traditional XX chart merely shows what the process was doing at the time the
EWMA 455
sample was taken. Thus, theEWMAchart can be used to take preemptive action
to prevent a process from going too far from the target.
If the process has inherent non-random components, an EWMA common
cause chart should be used. This is anEWMAchart with economic action limits
instead of control limits. EWMA control charts, which are described in the
next section, can be used to monitor processes that vary within the action limits.
The equation for computing the EWMA is
EWMA ? ^yt ? l?yt  ^yt? ?12:47?
In this equation ^yt is the predicted value of y at time period t, yt is the actual
value at time period t, and l is a constant between 0 and 1. If l is close to 1,
Equation 12.47 will give little weight to historic data; if l is close to 0 then
current observations will be given little weight. EWMA can also be thought
of as the forecasted process value at time period t ? 1, in other words,
EWMA ? ^yt?1.
Since most people already understand the traditional XX chart, thinking
about the relationship between XX charts and EWMA charts can help you
understand the EWMA chart. It is interesting to note that traditional XX charts
give 100% of the weight to the current sample and 0% to past data. This is
roughly equivalent to setting l ? 1 on an EWMA chart. In other words, the
traditional XX chart can be thought of as a special type of EWMA chart
where past data are considered to be unimportant (assuming run tests are
not applied to the Shewhart chart). This is equivalent to saying that the data
points are all independent of one another. In contrast, the EWMA chart uses
the information from all previous samples. Although Equation 12.47 may
look as though it is only using the results of the most recent data point, in reality
the EWMA weighting scheme applies progressively less weight to each
sample result as time passes. Figure 12.25 compares the weighting schemes of
EWMA and XX charts.
In contrast, as l approaches 0 the EWMA chart begins to behave like a
cusum chart. With a cusum chart all previous points are given equal weight.
Between the two extremes the EWMA chart weights historical data in importance
somewhere between the traditional Shewhart chart and the cusum
chart. By changing the value of l the chart’s behavior can be ‘‘adjusted’’ to
the process being monitored.
In addition to the weighting, there are other differences between the
EWMA chart and the XX chart. The ‘‘forecast’’ from the XX chart is always the
same: the next data point will be equal to the historical grand average. In
other words, the XX chart treats all data points as coming from a process that
doesn’t change its central tendency (implied when the forecast is always the
grand average).*
When using an XXchart it is not essential that the sampling interval be kept
constant. After all, the process is supposed to behave as if it were static.
However, the EWMA chart is designed to account for process drift and, therefore,
the sampling interval should be kept constant when using EWMA charts.
This is usually not a problem with automated manufacturing.
Krishnamoorthi (1991) describes amold line that produces green sand molds
at the rate of about one per minute. The molds are used to pour cylinder blocks
for large size engines. Application of SPC to the process revealed that the process
had an assignable cause that could not be eliminated from the process.
The mold sand, which was partly recycled, tended to increase and decrease in
EWMA 457
Figure 12.25. XX versus EWMA weighting.
*We aren’t saying this situation actually exists,we are just saying that the XX treats the process as if this were true. Studying the
patterns of variation will often reveal clues to making the process more consistent, even if the process variation remains
within the control limits.
temperature based on the size of the block being produced and the number of
blocks in the order. Sand temperature is important because it affects the compactability
percent, an important parameter. The sand temperature could not
be better controlled without adding an automatic sand cooler, which was not
deemed economical. However, the effect of the sand temperature on the compactability
percent could be made negligible by modifying the amount of
water added to the sand so feed-forward control was feasible.
Although Krishnamoorthi doesn’t indicate thatEWMAcharts were used for
this process, it is an excellent application for EWMA common cause charts.
The level of the sand temperature doesn’t really matter, as long as it is known.
The sand temperature tends to drift in cycles because the amount of heated
sand depends on the size of the casting and how many are being produced.A traditional
control chart for the temperature would indicate that sand temperature
is out-of-control, which we already know. What is really needed is a method to
predict what the sand temperature will be the next time it is checked, then the
operator can add the correct amount of water so the effect on the sand compactability
percent can be minimized. This will produce an in-control control chart
for compactability percent, which is what really matters.
The data in Table 12.20 show the EWMA calculations for the sand temperature
data. Using a spreadsheet program, Microsoft Excel for Windows, the
optimal value of l, that is the value which provided the ‘‘best fit’’ in the sense
that it produced the smallest sum of the squared errors, was found to be close
to 0.9. Figure 12.26 shows the EWMA common cause chart for this data, and
the raw temperature data as well. TheEWMAis a forecast of what the sand temperature
will be the next time it is checked. The operator can adjust the rate of
water addition based on this forecast.
Although it is not always necessary to put control limits on theEWMAchart,
as shown by the above example, it is possible to do so when the situation calls
for it. Three sigma control limits for the EWMA chart are computed based on
EWMA ? 2 l
?2  l?   ?12:48?
For the sand temperature example above, l ? 0:9 which gives
EWMA ? 2 0:9
?2  0:9?  ? 0:822
EWMA 459
Table 12.20. Data for EWMA chart of sand temperature.
125 125.00* 0.00
123 125.00 ^2.00**
118 123.20*** ^5.20
116 118.52 ^2.52
108 116.25 ^8.25
112 108.83 3.17
101 111.68 ^10.68
100 102.07 ^2.07
98 100.21 ^2.21
102 98.22 3.78
111 101.62 9.38
107 110.6 ^3.06
112 107.31 4.69
112 111.53 0.47
122 111.95 10.05
140 121.00 19.00
125 138.10 ^13.10
130 126.31 3.69
136 129.63 6.37
130 135.36 ^5.36
Continued on next page . . .
112 130.54 ^18.54
115 113.85 1.15
100 114.89 ^14.89
113 101.49 11.51
111 111.85 ^0.85
128 111.08 16.92
122 126.31 ^4.31
142 122.43 19.57
134 140.64 ^6.04
130 134.60 ^4.60
131 130.46 0.54
104 130.95 ^26.95
84 106.69 ^22.69
86 86.27 ^0.27
99 86.03 12.97
90 97.70 ^7.70
91 90.77 0.23
90 90.98 ^0.98
101 90.10 10.90
* The starting EWMA is either the target, or, if there is no target, the ?rst
** Error = Actual observation ^ EWMA. E.g., ^2 ?123 ^ 125.
*** Other than the ?rst sample, all EWMAs are computed as EWMA = last EWMA +
error. E.g., 123.2 ?125 + 0.9 (^2).
Table 12.20 (cont.)
2 is estimated using all of the data. For the sand temperature data  ? 15:37 so
 EWMA ? 15:37  ffiffiffiffiffiffiffiffi 0:82 p ? 13:92. The 3 control limits for the EWMA
chart are placed at the grand average plus and minus 41.75. Figure 12.27 shows
the control chart for these data. The EWMA line must remain within the control
limits. Since the EWMA accounts for ‘‘normal drift’’ in the process center
EWMA 461
Figure 12.26. EWMA chart of sand temperature.
Figure 12.27. EWMA control chart of sand temperature.
line, deviations beyond the control limits imply assignable causes other than
those accounted for by normal drift. Again, since the e?ects of changes in temperature
can be ameliorated by adjusting the rate of water input, the EWMA
control chart may not be necessary.
The choice of l is the subject of much literature. A value l of near 0 provides
more ‘‘smoothing’’ by giving greater weight to historic data, while a l value
near 1 gives greater weight to current data. Most authors recommend a value in
the range of 0.2 to 0.3. The justification for this range of l values is probably
based on applications of the EWMA technique in the field of economics,
where EWMA methods are in widespread use. Industrial applications are less
common, although the use of EWMA techniques is growing rapidly.
Hunter (1989) proposes a EWMA control chart scheme where l = 0.4. This
value of l provides a control chart with approximately the same statistical properties
as a traditional XX chart combined with the run tests described in the
AT&T Statistical Quality Control Handbook (commonly called the Western
Electric Rules). It also has the advantage of providing control limits that are
exactly half as wide as the control limits on a traditional XX chart. Thus, to compute
the control limits for an EWMA chart when l is 0.4 you simply compute
the traditional XX chart (or X chart) control limits and divide the distance
between the upper and lower control limits by two. The EWMA should remain
within these limits.
As mentioned above, the optimal value of l can be found using some spreadsheet
programs. The sum of the squared errors is minimized by changing the
value of l. If your spreadsheet doesn’t automatically find the minimum, it can
be approximated manually by changing the cell containing l or by setting up a
range of l values and watching what happens to the cell containing the sum of
the squared errors. A graph of the error sum of the squares versus different l
values can indicate where the optimum l lies.
Minitab has a built-inEWMAanalysis capability. We will repeat our analysis
for the sand temperature data. Choose Stat > Control Charts > EWMA and
you will see a dialog box similar to the one shown in Figure 12.28. Entering the
weight of 0.9 and a subgroup size of 1, then clicking OK, produces the chart in
Figure 12.29.
You may notice that the control limits calculated with Minitab are different
than those calculated in the previous example. The reason is that Minitab’s esti-
EWMA 463
Figure 12.29. Minitab EWMA chart.
Figure 12.28. Minitab EWMA dialog box.
mate of sigma is based on the average moving range. This method gives a sigma
value of 7.185517, substantially less than the estimate of 15.37 obtained by simply
calculating sigma combining all of the data. Minitab’s approach removes
the effect of the process drift. Whether or not this effect should be removed
from the estimate of sigma is an interesting question. In most situations we
probably want to remove it so our control chart will be more sensitive, allowing
us to detect more special causes for removal. However, as this example illustrates,
the situation isn’t always clear cut. In the situation described by the example
we might actually want to include the variation from drift into the control
limit calculations to prevent operator tampering.
In many cases an individuals control chart (I chart) will give results comparable
to the EWMA control chart. When this is the case it is usually best to opt
for the simpler I chart. An I chart is shown in Figure 12.30 for comparison with
the EWMA chart. The results are very similar to the EWMA chart from
Figure 12.30. I chart for sand temperature.
Whether using a EWMA common cause chart without control limits or an
EWMA control chart, it is a good idea to keep track of the forecast errors
using a control chart. The special cause chart is a traditional X chart, created
using the difference between the EWMA forecast and the actual observed
values. Figure 12.31 shows the special cause chart of the sand temperature data
analyzed above. The chart indicates good statistical control.
SPC and automatic process control
As SPC has grown in popularity its use has been mandated with more and
more processes. When this trend reached automated manufacturing processes
there was resistance from process control analysts who were applying a different
approach with considerable success (Palm, 1990). Advocates of SPC
attempted to force the use of traditional SPC techniques as feedback mechanisms
for process control. This inappropriate application of SPC was correctly
denounced by process control analysts. SPC is designed to serve a purpose
fundamentally different than automatic process control (APC). SPC advocates
correctly pointed out that APC was not a cure-all and that many process controllers
added variation by making adjustments based on data analysis that was
statistically invalid.
EWMA 465
Figure 12.31. Special cause control chart of EWMA errors.
Both SPC and APC have their rightful place in Six Sigma. APC attempts to
dynamically control a process to minimize variation around a target value.
This requires valid statistical analysis, which is the domain of the statistical
sciences. SPC makes a distinction between special causes and common causes
of variation. If APC responds to all variation as if it were the same it will result
in missed opportunities to reduce variation by attacking it at the source. A process
that operates closer to the target without correction will produce less variation
overall than a process that is frequently returned to the target via APC.
However, at times APC must respond to common cause variation that can’t be
economically eliminated, e.g., the mold process described above. Properly used,
APC can greatly reduce variability in the output.
Hunter (1986) shows that there is a statistical equivalent to the PID control
equation commonly used. The PID equation is
u?t? ? Ke?t? ?
TI ?1
e?s?ds ? KTD
dt   ?12:49?
The ‘‘PID’’ label comes from the fact that the first term is a proportional
term, the second an integral term and the third a derivative term. Hunter modified
the basic EWMA equation by adding two additional terms. The result is
the empirical control equation.
^yt?1 ? ^yt ? let ? l2Pet ? l3ret ?12:50?
The term ret means the first difference of the errors et i.e., ret ? et  et1.
Like the PID equation, the empirical control equation has a proportional, an
integral and a differential term. It can be used by APC or the results can be
plotted on a common cause chart and reacted to by human operators, as
described above. A special cause chart can be created to track the errors in the
forecast from the empirical control equation. Such an approach may help to
bring SPC and APC together to work on process improvement.
^ ^ ^
Process Capability Analysis
Process capability analysis is a two-stage process that involves:
1. Bringing a process into a state of statistical control for a reasonable
period of time.
2. Comparing the long-term process performance to management or engineering
Process capability analysis can be done with either attribute data or continuous
data if and only if the process is in statistical control, and has been for a reasonable
period of time (Figure 13.1).*
Application of process capability methods to processes that are not in statistical
control results in unreliable estimates of process capability and should
never be done.
How to perform a process capability study
This section presents a step-by-step approach to process capability analysis
(Pyzdek, 1985).
*Occasional freak values from known causes can usually be ignored.
Copyright 2003 by The McGraw-Hill Companies, Inc. Click Here for Terms of Use.
1. Select a candidate for the study
This step should be institutionalized.Agoal of any organization should
be ongoing process improvement. However, because a company has
only a limited resource base and can’t solve all problems simultaneous-
Figure 13.1. Process control concepts illustrated.
From Continuing Process Control and Process Capability Improvement, p. 4a. Copyright 1983.
Used by permission of the publisher, Ford Motor Company, Dearborn, Michigan.
ly, it must set priorities for its e?orts. The tools for this include Pareto
analysis and ?shbone diagrams.
2. De?ne the process
It is all too easy to slip into the trap of solving the wrong
problem. Once the candidate area has been selected in step 1,
de?ne the scope of the study. A process is a unique combination
of machines, tools, methods, and personnel engaged in adding
value by providing a product or service. Each element of the
process should be identi?ed at this stage. This is not a trivial exercise.
The input of many people may be required. There are likely
to be a number of con?icting opinions about what the process
actually involves.
3. Procure resources for the study
Process capability studies disrupt normal operations and require
signi?cant expenditures of both material and human resources.
Since it is a project of major importance, it should be managed as
such. All of the usual project management techniques should be
brought to bear. This includes planning, scheduling, and management
status reporting.
4. Evaluate the measurement system
Using the techniques described in Chapter 9, evaluate the measurement
system’s ability to do the job. Again, be prepared to spend the
time necessary to get a valid means of measuring the process before
going ahead.
5. Prepare a control plan
The purpose of the control plan is twofold: 1) isolate and control as
many important variables as possible and, 2) provide a mechanism for
tracking variables that cannot be completely controlled. The object of
the capability analysis is to determine what the process can do if it is
operated the way it is designed to be operated. This means that such
obvious sources of potential variation as operators and vendors will be
controlled while the study is conducted. In other words, a single welltrained
operator will be used and the material will be from a single
There are usually some variables that are important, but that are not
controllable. One example is the ambient environment, such as temperature,
barometric pressure, or humidity. Certain process variables
may degrade as part of the normal operation; for example, tools wear
and chemicals are used. These variables should still be tracked using
logsheets and similar tools. See page 74, Information systems requirements.
Process capability analysis (PCA) 469
6. Select a method for the analysis
The SPC method will depend on the decisions made up to this point. If
the performance measure is an attribute, one of the attribute charts
will be used. Variables charts will be used for process performance measures
assessed on a continuous scale. Also considered will be the skill
level of the personnel involved, need for sensitivity, and other resources
required to collect, record, and analyze the data.
7. Gather and analyze the data
Use one of the control charts described in Chapter 12, plus common
sense. It is usually advisable to have at least two people go over the
data analysis to catch inadvertent errors in entering data or performing
the analysis.
8. Track down and remove special causes
A special cause of variation may be obvious, or it may take months of
investigation to ?nd it. The e?ect of the special cause may be good or
bad. Removing a special cause that has a bad e?ect usually involves
eliminating the cause itself. For example, if poorly trained operators
are causing variability the special cause is the training system (not the
operator) and it is eliminated by developing an improved training system
or a process that requires less training. However, the ‘‘removal’’
of a bene?cial special cause may actually involve incorporating the special
cause into the normal operating procedure. For example, if it is discovered
that materials with a particular chemistry produce better
product the special cause is the newly discovered material and it can
be made a common cause simply by changing the speci?cation to assure
that the new chemistry is always used.
9. Estimate the process capability
One point cannot be overemphasized: the process capability cannot be
estimated until a state of statistical control has been achieved! After
this stage has been reached, the methods described later in this chapter
may be used. After the numerical estimate of process capability has
been arrived at, it must be compared to management’s goals for the process,
or it can be used as an input into economic models. Deming’s allor-
none rules (Deming 1986, 409?) provide a simple model that can be
used to determine if the output from a process should be sorted 100%
or shipped as-is.
10. Establish a plan for continuous process improvement
Once a stable process state has been attained, steps should be taken to
maintain it and improve upon it. SPC is just one means of doing this.
Far more important than the particular approach taken is a company
environment that makes continuous improvement a normal part of
the daily routine of everyone.
Process capability analysis (PCA) 471
Statistical analysis of process capability data
This section presents several methods of analyzing the data obtained from a
process capability study.
1. Collect samples from 25 or more subgroups of consecutively produced
units. Follow the guidelines presented in steps 1^10 above.
2. Plot the results on the appropriate control chart (e.g., c chart). If all
groups are in statistical control, go to the step #3. Otherwise identify
the special cause of variation and take action to eliminate it. Note that a
special cause might be bene?cial. Bene?cial activities can be ‘‘eliminated’’
as special causes by doing them all of the time. A special cause is
‘‘special’’ only because it comes and goes, not because its impact is either
good or bad.
3. Using the control limits from the previous step (called operation control
limits), put the control chart to use for a period of time. Once you are
satis?ed that su?cient time has passed for most special causes to have
been identi?ed and eliminated, as veri?ed by the control charts, go to
the step #4.
4. The process capability is estimated as the control chart centerline. The
centerline on attribute charts is the long-term expected quality level
of the process, e.g., the average proportion defective. This is the level
created by the common causes of variation.
If the process capability doesn’t meet management requirements, take
immediate action to modify the process for the better. ‘‘Problem solving’’ (e.g.,
studying each defective) won’t help, and it may result in tampering. Whether it
meets requirements or not, always be on the lookout for possible process
improvements. The control charts will provide verification of improvement.
1. Collect samples from 25 or more subgroups of consecutively produced
units, following the 10-step plan described above.
2. Plot the results on the appropriate control chart (e.g., XX and R chart). If
all groups are in statistical control, go to the step #3. Otherwise identify
the special cause of variation and take action to eliminate it.
3. Using the control limits from the previous step (called operation control
limits), put the control chart to use for a period of time. Once you are
satis?ed that su?cient time has passed for most special causes to have
been identi?ed and eliminated, as veri?ed by the control charts, estimate
process capability as described below.
The process capability is estimated from the process average and standard
deviation, where the standard deviation is computed based on the average
range or average standard deviation. When statistical control has been achieved,
the capability is the level created by the common causes of process variation.
The formulas for estimating the process standard deviation are:
R chart method:
^  ?
d2 ?13:1?
S chart method:
^  ?
c4 ?13:2?
The values d2 and c4 are constants from Table 11 in the Appendix.
Process capability indexes
Only now can the process be compared to engineering requirements.* One
way of doing this is by calculating ‘‘Capability Indexes.’’ Several popular capability
indexes are given in Table 13.1.
Table 13.1. Process capability indices.
CP ?
engineering tolerance
6 ^  ?13:3?
CR ? 100 
6 ^ 
engineering tolerance ?13:4?
CM ?
engineering tolerance
8 ^  ?13:5?
ZU ?
upper specification  
^  ?13:6?
ZL ?

 lower specification
^  ?13:7?
Continued on next page . . .
*Other sources of requirements include customers and management.
Process capability analysis (PCA) 473
ZMIN: ? Minimum fZL; ZUg ?13:8?
3 ?13:9?
Cpm ?
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi 1 ? ?  T?2
^ 2 r ?13:10?
Table 13.1LContinued . . .
Interpreting capability indexes
Perhaps the biggest drawback of using process capability indexes is that they
take the analysis a step away from the data. The danger is that the analyst will
lose sight of the purpose of the capability analysis, which is to improve quality.
To the extent that capability indexes help accomplish this goal, they are worthwhile.
To the extent that they distract from the goal, they are harmful. The
analyst should continually refer to this principle when interpreting capability
CPLThis is one of the first capability indexes used. The ‘‘natural tolerance’’
of the process is computed as 6. The index simply makes a direct comparison
of the process natural tolerance to the engineering requirements.
Assuming the process distribution is normal and the process
average is exactly centered between the engineering requirements, a
CP index of 1 would give a ‘‘capable process.’’ However, to allow a bit
of room for process drift, the generally accepted minimum value for
CP is 1.33. In general, the larger CP is, the better. For a Six Sigma process,
i.e., a process that produces 3.4 defects per million opportunities
including a 1.5 sigma shift, the value of CP would be 2.
The CP index has two major shortcomings. First, it can’t be used
unless there are both upper and lower specifications. Second, it does
not account for process centering. If the process average is not exactly
centered relative to the engineering requirements, the CP index will
give misleading results. In recent years, the CP index has largely been
replaced by CPK (see below).
CRLThe CR index is equivalent to the CP index. The index simply makes a
direct comparison of the process to the engineering requirements.
Assuming the process distribution is normal and the process average
is exactly centered between the engineering requirements, a CR index
of 100% would give a ‘‘capable process.’’ However, to allow a bit of
room for process drift, the generally accepted maximum value for CR
is 75%. In general, the smaller CR is, the better. The CR index suffers
from the same shortcomings as the CP index. For a Six Sigma process,
i.e., a process that produces 3.4 defects per million opportunities including
a 1.5 sigma shift, the value of CR would be 50%.
CMLThe CM index is generally used to evaluate machine capability studies,
rather than full-blown process capability studies. Since variation will
increase when other sources of process variation are added (e.g., tooling,
fixtures, materials, etc.), CM uses an 8 sigma spread rather than a 6
sigma spread to represent the natural tolerance of the process. For a
machine to be used on a Six Sigma process, a 10 sigma spread would
be used.
ZULThe ZU index measures the process location (central tendency) relative
to its standard deviation and the upper requirement. If the distribution
is normal, the value of ZU can be used to determine the percentage
above the upper requirement by using Table 2 in the Appendix. The
method is the same as described in Equations 9.11 and 9.12, using the
Z statistic, simply use ZU instead of using Z.
In general, the bigger ZU is, the better. A value of at least +3 is
required to assure that 0.1% or less defective will be produced. A value
of +4 is generally desired to allow some room for process drift. For a
Six Sigma process ZU would be +6.
ZLLThe ZL index measures the process location relative to its standard
deviation and the lower requirement. If the distribution is normal, the
value of ZL can be used to determine the percentage below the lower
requirement by using Table 2 in the Appendix. The method is the
same as described in Equations 9.11 and 9.12, using the Z transformation,
except that you use ZL instead of using Z.
In general, the bigger ZL is, the better. A value of at least +3 is
required to assure that 0.1% or less defective will be produced. A value
of +4 is generally desired to allow some room for process drift. For a
Six Sigma process ZL would be +6.
ZMINLThe value of ZMIN is simply the smaller of the ZL or the ZU values. It is
used in computing CPK. For a Six Sigma process ZMIN would be +6.
CPKLThe value of CPK is simply ZMIN divided by 3. Since the smallest value
represents the nearest specification, the value ofCPK tells you if the pro-
cess is truly capable of meeting requirements. A CPK of at least +1 is
required, and +1.33 is preferred. Note that CPK is closely related to CP,
the difference between CPK and CP represents the potential gain to be
had from centering the process. For a Six Sigma processCPK would be 2.
Example of capability analysis using normally
distributed variables data
Assume we have conducted a capability analysis using X-bar and R charts
with subgroups of 5. Also assume that we found the process to be in statistical
control with a grand average of 0.99832 and an average range of 0.02205. From
the table of d2 values (Appendix Table 11), we find d2 is 2.326 for subgroups of
5. Thus, using Equation 13.1,
^  ?
2:326 ? 0:00948
Before we can analyze process capability, we must know the requirements.
For this process the requirements are a lower specification of 0.980 and an
upper specification of 1.020 (1:000  0:020). With this information, plus the
knowledge that the process performance has been in statistical control, we can
compute the capability indexes for this process.
CP ?
engineering tolerance
6 ^  ?
1:020  0:9800
6  0:00948 ? 0:703
CR ? 100 
6 ^ 
engineering tolerance ? 100 
6  0:00948
0:04 ? 142:2%
CM ?
engineering tolerance
8 ^  ?
8  0:00948 ? 0:527
ZU ?
upper specification  XX
^  ?
1:020  0:99832
0:00948 ? 2:3
ZL ?
X  lower specification
^  ?
0:99832  0:980
0:00948 ? 1:9
ZMIN ? Minimum f1:9; 2:3g ? 1:9
3 ?
3 ? 0:63
Process capability analysis (PCA) 475
Assuming that the target is precisely 1.000, we compute
Cpm ?
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi 1 ? ? XX  T?2
^ 2 s ?
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi 1 ? ?0:99832  1:000?2
0:009482 r ? 0:692
CPL(0.703) Since the minimum acceptable value for this index is 1, the 0.703
result indicates that this process cannot meet the requirements.
Furthermore, since the CP index doesn’t consider the centering process,
we know that the process can’t be made acceptable by merely
adjusting the process closer to the center of the requirements. Thus,
we can expect the ZL, ZU, and ZMIN values to be unacceptable too.
CRL(142.2%) This value always gives the same conclusions as the CP index.
The number itself means that the ‘‘natural tolerance’’ of the process
uses 142.2% of the engineering requirement, which is, of course, unacceptable.
CML(0.527) The CM index should be 1.33 or greater. Obviously it is not. If
this were a machine capability study the value of the CM index would
indicate that the machine was incapable of meeting the requirement.
ZUL(+2.3) We desire a ZU of at least +3, so this value is unacceptable. We
can use ZU to estimate the percentage of production that will exceed
the upper specification. Referring to Table 2 in the Appendix we find
that approximately 1.1% will be oversized.
ZLL(+1.9)We desire a ZL of at least +3, so this value is unacceptable.Wecan
use ZL to estimate the percentage of production that will be below the
lower specification. Referring to Table 2 in the Appendix we find that
approximately 2.9% will be undersized. Adding this to the 1.1% oversized
and we estimate a total reject rate of 4.0%. By subtracting this
from 100% we get the projected yield of 96.0%.
ZMINL(+1.9) The smaller of ZL and ZU. Since neither of these two results
were acceptable, ZMIN cannot be acceptable.
CPKL(0.63) The value of CPK is only slightly smaller than that of CP. This
indicates that we will not gain much by centering the process. The
actual amount we would gain can be calculated by assuming the process
is exactly centered at 1.000 and recalculating ZMIN. This gives a predicted
total reject rate of 3.6% instead of 4.0%.
Minitab has a built-in capability analysis feature, which will be demonstrated
here using the rod diameter data. The output is shown in Figure 13.2. The summary,
which is called a ‘‘Six Pack’’ in Minitab, provides a compact picture of
the most important statistics and analysis. The control charts tell you if the process
is in statistical control (it is). If it’s out of control, stop and find out why.
The histogram and normal probability plot tell you if the normality assumption
is justified. If not, you can’t trust the capability indices. Consider using
Minitab’s non-normal capability analysis (see ‘‘Example of non-normal capability
analysis using Minitab’’ below). The ‘‘within’’ capability indices are based
on within-subgroup variation only, called short-term variability. The Cp and
Cpk values are both unacceptable. The ‘‘overall’’ capability indices are based on
total variation, called long-term variability. Total variation includes variation
within subgroups and variation between subgroups. The Pp and Ppk values are
both unacceptable. The Capability Plot in the lower right of the six pack graphically
compares within variability (short-term) and overall variability (longterm)
to the specifications. Ideally, for a Six Sigma process, the process variability
(Process Tolerance) will be about half of the specifications. However, the
Process capability analysis (PCA) 477
Figure 13.2. Minitab capability analysis for normally distributed data.
capability plot for the example shows that the process tolerance is actually
wider than the specifications.
What is missing in the six pack is an estimate of the process yield. There is an
option in the six pack to have this information (and a great deal more) stored
in the worksheet. Alternatively, you can run Minitab’s Capability Analysis
(Normal) procedure and get the information along with a larger histogram
(see Figure 13.3). The PPM levels confirm what the capability and performance
indices told us, this process just ain’t up to snuff!
Minitab has a built-in capability to perform process capability analysis for
non-normal data which will be demonstrated with an example. The process
involved is technical support by telephone. A call center has recorded the total
time it takes to ‘‘handle’’ 500 technical support calls. Handle time is a total
cycle time metric which includes gathering preliminary information, addressing
the customer’s issues, and after call work. It is a CTQ metric that also impacts
the shareholder. It has been determined that the upper limit on handle time is
Figure 13.3. An alternative Minitab capability analysis.
45 minutes. We assume that the instructions specified in the ‘‘how to perform a
process capability study’’ approach have been followed and that we have completed
the first six steps and have gathered the data. We are, therefore, at the
‘‘analyze the data’’ step.
Phase 19Check for special causes
To begin we must determine if special causes of variation were present during
our study. A special cause is operationally defined as points beyond one of the
control limits. Some authors recommend that individuals control charts be
used for all analysis, so we’ll try this first, see Figure 13.4.
There are 12 out-of-control points in the chart shown in Figure 13.4, indicating
that special causes are present. However, a closer look will show that there’s
something odd about the chart. Note that the lower control limit (LCL) is
18:32. Since we are talking about handle time, it is impossible to obtain any
result that is less than zero. A reasonable process owner might argue that if the
LCL is in the wrong place (which it obviously is), then the upper control limit
(UCL)may be as well. Also, the data appear to be strangely cut-off near the bottom.
Apparently the individuals chart is not the best way to analyze data like
Process capability analysis (PCA) 479
Figure 13.4. Individuals process behavior chart for handle time.
But what can be done? Since we don’t know if special causes were present, we
can’t determine the proper distribution for the data. Likewise, if we don’t
know the distribution of the data we can’t determine if special causes are present
because the control limits may be in the wrong place. This may seem to be a classic
case of ‘‘which came first, the chicken or the egg?’’ Fortunately there is a
way out. The central limit theorem tells us that stable distributions produce normally
distributed averages, even when the individuals data are not normally distributed.
Since ‘‘stable’’ means no special causes, then a process with nonnormal
averages would be one that is influenced by special causes, which is precisely
what we are looking for. We created subgroups of 10 in Minitab (i.e.,
observations 1^10 are in subgroup 1, observations 11^20 are in subgroup 2,
etc.) and tested the normality of the averages. The probability plot in Figure
13.5 indicates that the averages are normally distributed.
Figure 13.6 shows the control chart for the process using averages instead of
individuals. The chart indicates that the process is in statistical control. The process
average is stable at 18.79 minutes. The LCL is comfortably above zero at
5.9 minutes; any average below this is an indication that things are better than
normal and we’d want to know why in case we can do it all of the time. Any aver-
Figure 13.5. Normality test of subgroups of n ? 10.
age above 31.67 minutes indicates worse than normal behavior and we’d like to
find the reason and fix it. Averages between these two limits are normal for
this process.
Phase 29Examine the distribution
Now that stability has been determined, we can trust the histogram to give us
an accurate display of the distribution of handle times. The histogram shows
the distribution of actual handle times, which we can compare to the upper specification
limit of 45 minutes. The couldn’t be done with the control chart in
Figure 13.6 because it shows averages, not individual times. Figure 13.7 shows
the histogram of handle time with the management upper requirement of 45
minutes drawn in. Obviously a lot of calls exceed the 45 minute requirement.
Since the control chart is stable, we know that this is what we can expect from
this process. There is no point in asking why a particular call took longer than
45 minutes. The answer is ‘‘It’s normal for this process.’’ If management doesn’t
like the answer they’ll need to sponsor one or more Six Sigma projects to
improve the process.
Process capability analysis (PCA) 481
Figure 13.6. Averages of handle time (n ? 10 per subgroup).
Phase 39Predicting the long-termdefect rate for the process
The histogram makes it visually clear that the process distribution is nonnormal.
This conclusion can be tested statistically with Minitab by going to
Stats > Basic Statistics > Normality test. Minitab presents the data in a chart
specially scaled so that normally distributed data will plot as a straight line
(Figure 13.8). The vertical axis is scaled in cumulative probability and the horizontal
in actual measurement values. The plot shows that the data are not even
close to falling on the straight line, and the P-value of 0 confirms that the data
are not normal.*
To make a prediction about the defect rate we need to find a distribution that
fits the data reasonably well. Minitab offers an option that performs capability
analysis using the Weibull rather than the normal distribution. Choose Stat >
Quality Tools > Capability Analysis (Weibull) and enter the column name for
the handle time data. The output is shown in Figure 13.9.
Minitab calculates process performance indices rather than process capability
indices (i.e., Ppk instead of Cpk). This means that the denominator for the
Figure 13.7. Histogram of handle time.
*The null hypothesis is that the data are normally distributed. The P-value is the probability of obtaining the observed results
if the null hypothesis were true. In this case, the probability is 0.
Process capability analysis (PCA) 483
Figure 13.8. Normality test of handle time.
Figure 13.9. Capability analysis of handle times based on the Weibull distribution.
indices is the overall standard deviation rather than the standard deviation
based on only within-subgroup variability. This is called the long-term process
capability, which Minitab labels as ‘‘Overall (LT) Capability.’’ When the
process is in statistical control, as this one is, there will be little difference in
the estimates of the standard deviation. When the process is not in statistical
control the short-term capability estimates have no meaning, and the longterm
estimates are of dubious value as well. Process performance indices are
interpreted in exactly the same way as their process capability counterparts.
Minitab’s analysis indicates that the process is not capable (Ppk < 1). The
estimated long-term performance of the process is 41,422 defects per million
calls. The observed performance is even worse, 56,000 defects per million
calls. The difference is a reflection of lack of fit. The part of the Weibull
curve we’re most interested in is the tail area above 45, and the curve appears
to drop off more quickly than the actual data. When this is the case it is better
to estimate the long-term performance using the actual defect count rather
than Minitab’s estimates.
Rolled throughput yield and sigma level
The rolled throughput yield (RTY) summarizes defects-per-millionopportunities
(DPMO) data for a process or product. DPMO is the same as
the parts-per-million calculated by Minitab. RTY is a measure of the overall
process quality level or, as its name suggests, throughput. For a process,
throughput is a measure of what comes out of a process as a function of
what goes into it. For a product, throughput is a measure of the quality of
the entire product as a function of the quality of its various features.
Throughput combines the results of the capability analyses into a measure of
overall performance.
To compute the rolled throughput yield for an N-step process (or N-characteristic
product), use the following equation:
Rolled Throughput Yield
? 1 
1,000,000   1 
1,000,000     1 
1,000,000  ?13:11?
Where DPMOx is the defects-per-million-opportunities for step x in the process.
For example, consider a 4-step process with the following DPMO levels
at each step (Table 13.2) (dpu is defects-per-unit).
Figure 13.10 shows the Excel spreadsheet and formula for this example. The
meaning of the RTY is simple: if you started 1,000 units through this 4-step process
you would only get 979 units out the other end. Or, equivalently, to get
1,000 units out of this process you should start with ?1,000=0:979? ? 1 ? 1,022
units of input. Note that the RTY is worse than the worst yield of any process
or step. It is also worse than the average yield of 0.995. Many a process owner is
lulled into complacency by reports showing high average process yields. They
are confused by the fact that, despite high average yields, their ratio of end-ofthe-
line output to starting input is abysmal. Calculating RTY may help open
their eyes to what is really going on. The effect of declining RTYs grows exponentially
as more process steps are involved.
Estimating process yield 485
Table 13.2. Calculations used to ?nd RTY.
PROCESS STEP DPMO dpu=DPMO/1,000,000 1 ^ dpu
1 5,000 0.005000 0.9950
2 15,000 0.015000 0.9850
3 1,000 0.001000 0.9990
4 50 0.000050 0.99995
Rolled Throughput Yield ? 0:995  0:985  0:999  0:99995 ? 0:979
Figure 13.10. Excel spreadsheet for RTY.
The sigma level equivalent for this 4-step process RTY is 3.5 (see Appendix,
Table 18). This would be the estimated ‘‘process’’ sigma level. Also see
‘‘Normalized yield and sigma level’’ below. Use the RTY worksheet below to
document the RTY.
In Chapter 9 we discussed that, if a Poisson distribution is assumed for
defects, then the probability of getting exactly x defects on a unit from a process
with an average defect rate of  is P?x? ? ?xe?=x!, where e ? 2:71828.
Recall that RTY is the number of units that get through all of the processes or
process steps with no defects, i.e., x = 0. If we let  ? dpu then the RTY can
be calculated as the probability of getting exactly 0 defects on a unit with an
average defect rate of dpu, or RTY ? edpu. However, this approach can only
be used when all of the process steps have the same dpu. This is seldom the
case. If this approach is used for processes with unequal dpu’s, the calculated
RTY will underestimate the actual RTY. For the example presented in Table
13.2 we obtain the following results using this approach:
dpu ?
NXdpu ?
0:005 ? 0:015 ? 0:001 ? 0:00005 ? ?? 0:005263
edpu ? e0:005263 ? 0:994751
Note that this is considerably better than the 0.979 RTY calculated above.
Since the individual process steps have greatly different dpu’s, the earlier estimate
should be used.
RTY worksheet
RTY Capability
RTY Actual
Project RTY Goal
Things to consider:
& How large are the gaps between the actual RTY, the capability RTY, and the
project’s goal RTY?
& Does actual process performance indicate a need for a breakthrough project?
& Would we need a breakthrough project if we operated up to capability?
& Would focusing on a subset of CTXs achieve the project’s goals at lower cost?
Normalized yield and sigma level
To compute the normalized yield, which is a kind of average, for an Nprocess
or N-product department or organization, use following equation:
Normalized Yield
? ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffififfiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffififfiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi 1 
1,000,000   1 
1,000,000     1 
1,000,000   Ns ?13:12?
For example, consider a 4-process organization with the following DPMO
levels for each process:
PROCESS DPMO DPMO/1,000,000 1-(DPMO/1,000,000)
Billing 5,000 0.005000 0.9950000
Shipping 15,000 0.015000 0.9850000
Manufacturing 1,000 0.001000 0.9990000
Receiving 50 0.000050 0.9999500
Normalized Yield ? ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffififfiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi 0:995  0:985  0:999  0:99995 4 p ? 0:99472
Figure 13.11 shows the Excel spreadsheet for this example.
The sigma level equivalent of this 4-process organization’s normalized yield
is 4.1 (see Appendix, Table 18). This would be the estimated ‘‘organization’’
sigma level. Normalized yield should be considered a handy accounting device
for measuring overall system quality. Because it is a type of average it is not
Estimating process yield 487
Figure 13.11. Excel spreadsheet for calculating normalized yield.
necessarily indicative of any particular product or process yield or of how the
organization’s products will perform in the field. To calculate these refer to
‘‘Rolled throughput yield and sigma level’’ above.
Assuming every step has an equal yield, it is possible to ‘‘backsolve’’ to find
the normalized yield required in order to get a desired RTY for the entire process,
see Equation 13.13.
Yn ? ffiffiffiffiffiffiffiffiffi RTY Np
? RTY1=N ?13:13?
where Yn is the yield for an individual process step and N is the total number of
If the process yields are not equal, then Yn is the required yield of the worst
step in the process. For example, for a 10-step process with a desired RTY
of 0.999 the worst acceptable yield for any process step is Yn ? RTY1=10 ? 0:999 ? ?1=10? 0:9999. If all other yields are not 100% then the worst-step yield
must be even higher.
Unfortunately, finding the RTY isn’t always as straightforward as described
above. In the real world you seldom find a series of process steps all neatly feeding
into one another in a nice, linear fashion. Instead, you have different supplier
streams, each with different volumes and different yields. There are steps
that are sometimes taken and sometimes not. There are test and inspection
stations, with imperfect results. There is rework and repair. The list goes on
and on. In such cases it is sometimes possible to trace a particular batch of inputs
through the process, monitoring the results after each step. However, this is
often exceedingly difficult to control. The production and information systems
are not designed to provide the kind of tracking needed to get accurate results.
The usual outcome of such attempts is questionable data and disappointment.
High-end simulation software offers an alternative. With simulation you can
model the individual steps, then combine the steps into a process using the software.
The software will monitor the results as it ‘‘runs’’ the process as often as
necessary to obtain the accuracy needed. Figure 13.12 shows an example. Note
that the Properties dialog box is for step 12 in the process (‘‘Right Med?’’). The
model is programmed to keep track of the errors encountered as a Med Order
works its way through the process. Statistics are defined to calculate dpu and
RTY for the process as a whole (see the Custom Statistics box in the lower
right section of Figure 13.12). Since the process is non-linear (i.e., it includes
feedback loops) it isn’t a simple matter to determine which steps would have
the greatest impact on RTY. However, the software lets the Black Belt test multiple
what-if scenarios to determine this. It can also link to Minitab or Excel to
allow detailed data capture and analysis.
Estimating process yield 489
Figure 13.12. Finding RTY using simulation software (iGrafx Process for
Six Sigma, Corel Corporation).
^ ^ ^
Statistical Analysis of Cause
and E?ect
Many statistical tests are only valid if certain underlying assumptions are
met. In most cases, these assumptions are stated in the statistical textbooks
along with the descriptions of the particular statistical technique. This chapter
describes some of the more common assumptions encountered in Six Sigma
project work and how to test for them. However, the subject of testing underlying
assumptions is a big one and you might wish to explore it further with a
Master Black Belt.
Continuous versus discrete data
Data come in two basic flavors: Continuous and Discrete. These data types
are discussed elsewhere in this book. To review the basic idea, continuous data
are numbers that can be expressed to any desired level of precision, at least in
theory. For example, using a mercury thermometer I can say that the temperature
is 75 degrees Fahrenheit. With a home digital thermometer I could say it’s
75.4 degrees. A weather bureau instrument could add additional decimal places.
Discrete data can only assume certain values. For example, the counting numbers
can only be integers. Some survey responses force the respondent to choose
a particular number from a list (pick a rating on a scale from 1 to 10).
Copyright 2003 by The McGraw-Hill Companies, Inc. Click Here for Terms of Use.
Some statistical tests assume that you are working with either continuous or
discrete data. For example, ANOVA assumes that continuous data are being
analyzed, while chi-square and correspondence analysis assume that your data
are counts. In many cases the tests are insensitive to departures from the datatype
assumption. For example, expenditures can only be expressed to two decimal
places (dollars and cents), but they can be treated as if they are continuous
data. Counts can usually be treated as continuous data if there are many different
counts in the data set. For example, if the data are defect counts ranging
from 10 to 30 defects with all 21 counts showing up in the data (10, 11, 12, . . .,
28, 29, 30).
In some cases, however, the data type matters. For example, if discrete data
are plotted on control charts intended for continuous data the control limit
calculations will be incorrect. Run tests and other non-parametric tests will
also be affected by this. The problem of ‘‘discretized’’ data is often caused by
rounding the data to too few decimal places when they are recorded. This
rounding can be human caused, or it might be a computer program not
recording or displaying enough digits. The simple solution is to record more
digits. The problem may be caused by an inadequate measurement system.
This situation can be identified by a measurement system analysis (see
Chapter 10). The problem can be readily detected by creating a dot plot of
the data.
Let’s say you want to determine if operator experience has an impact on
the defects. One way to analyze this is to use a technique such as regression
analysis to regress X ? years of experience on Y ? defects. Another would
be to perform a chi-square analysis on the defects by experience level. To do
this you need to put the operators into discrete categories, then analyze
the defects in each category. This can be accomplished by ‘‘discretizing’’ the
experience variable. For example, you might create the following discrete
Testing common assumptions 491
Experience (years) Experience Category
Less than 1 New
1 to 2 Moderately experienced
3 to 5 Experienced
More than 5 Very experienced
The newly classified data are now suitable for chi-square analysis or other
techniques that require discrete data.
Independence assumption
Statistical independence means that two values are not related to one
another. In other words, knowing what one value is provides no information
as to what the other value is. If you throw two dice and I tell you that one of
them is a 4, that information doesn’t help you predict the value on the other
die. Many statistical techniques assume that the data are independent. For
example, if a regression model fits the data adequately, then the residuals will
be independent. Control charts assume that the individual data values are independent;
i.e., knowing the diameter of piston #100 doesn’t help me predict
the diameter of piston #101, nor does it tell me what the diameter of piston
#99 was. If I don’t have independence, the results of my analysis will be
wrong. I will believe that the model fits the data when it does not. I will tamper
with controlled processes.
Independence can be tested in a variety of ways. If the data are normal (testing
the normality assumption is discussed below) then the run tests described
for control charts can be used.
A scatter plot can also be used. Let y ? Xt1 and plot X vs. Y. You will see
random patterns if the data are independent. Software such as Minitab offer several
ways of examining independence in time series data. Note: lack of independence
in time series data is called autocorrelation.
If you don’t have independence you have several options. In many cases the
best course of action is to identify the reason why the data are not independent
and fix the underlying cause. If the residuals are not independent, add terms to
the model. If the process is drifting, add compensating adjustments.
If fixing the root cause is not a viable option, an alternative is to use a statistical
technique that accounts for the lack of independence. For example, the
EWMA control chart or a time series analysis that can model autocorrelated
data. Another is to modify the technique to work with your autocorrelated
data, such as using sloped control limits on the control chart. If data are cyclical
you can create uncorrelated data by using a sampling interval equal to the cycle
length. For example, you can create a control chart comparing performance on
Monday mornings.
Normality assumption
Statistical techniques such as t-tests, Z-tests, ANOVA, and many others
assume that the data are at least approximately normal. This assumption is
easily tested using software. There are two approaches to testing normality:
graphical and statistical.
One graphical approach involves plotting a histogram of the data, then superimposing
a normal curve over the histogram. This approach works best if you
have at least 200 data points, and the more the merrier. For small data sets the
interpretation of the histogram is difficult; the usual problem is seeing a lack of
fit when none exists. In any case, the interpretation is subjective and two people
often reach different conclusions when viewing the same data. Figure 14.1
Testing common assumptions 493
Figure 14.1. Histograms with normal curves for di?erent sample sizes.
shows four histograms for normally distributed data with mean?10, sigma?1
and sample sizes ranging from 30 to 500.
An alternative to the histogram/normal curve approach is to calculate a
‘‘goodness-of-fit’’ statistic and a P-value. This gives an unambiguous acceptance
criterion; usually the researcher rejects the assumption of normality if
P<0.05. However, it has the disadvantage of being non-graphical. This violates
the three rules of data analysis:
To avoid violating these important rules, the usual approach is to supplement
the statistical analysis with a probability plot. The probability plot is scaled so
that normally distributed data will plot as a straight line. Figure 14.2 shows the
probability plots that correspond to the histograms and normal curves in
Figure 14.1. The table below Figure 14.2 shows that the P-values are all comfortably
above 0.05, leading us to conclude that the data are reasonably close to
the normal distribution.
Figure 14.2. Normal probability plots and goodness of ?t tests.
N 30 100 200 500
P-Value 0.139 0.452 0.816 0.345
When data are not normal, the following steps are usually pursued:
* Do nothing. Often the histogram or probability plot shows that the
normal model ?ts the data well ‘‘where it counts.’’ If the primary interest
is in the tails, for example, and the curve ?ts the data well there,
then proceed to use the normal model despite the fact that the
P-value is less than 0.05. Or if the model ?ts the middle of the distribution
well and that’s your focus, go with it. Likewise, if you have a
very large sample you may get P-values greater than 0.05 even though
the model appears to ?t well everywhere. I work with clients who
routinely analyze data sets of 100,000+ records. Samples this large
will ?ag functionally and economically unimportant departures from
normality as ‘‘statistically signi?cant,’’ but it isn’t worth the time or
the expense to do anything about it.
* Transform the data. It is often possible to make the data normal by
performing a mathematical operation on the data. For example, if
the data distribution has very long tails to the high side, taking the
logarithm often creates data that are normally distributed. Minitab’s
control chart feature o?ers the Box-Cox normalizing power transformation
that works with many data distributions encountered in Six
Sigma work. The downside to transforming is that data have to be
returned to the original measurement scale before being presented to
non-technical personnel. Some statistics can’t be directly returned to
their original units; for example, if you use the log transform then
you can’t ?nd the mean of the original data by taking the inverse log
of the mean of the transformed data.
* Use averages. Averages are a special type of transformation because
averages of subgroups always tend to be normally distributed, even if
the underlying data are not. Sometimes the subgroup sizes required
to achieve normality can be quite small.
* Fit another statistical distribution. The normal distribution isn’t the
only game in town. Try ?tting other curves to the data, such as the
Weibull or the exponential. Most statistics packages, such as
Minitab, have the ability to do this. If you have a knack for programming
spreadsheets, you can use Excel’s solver add-in to evaluate the
?t of several distributions.
Testing common assumptions 495
* Use a non-parametric technique. There are statistical methods, called
non-parametric methods, that don’t make any assumptions about
the underlying distribution of the data. Rather than evaluating the differences
of parameters such as the mean or variance, non-parametric
methods use other comparisons. For example, if the observations are
paired they may be compared directly to see if the after is di?erent
than the before. Or the method might examine the pattern of points
above and below the median to see if the before and after values are
randomly scattered in the two regions. Or ranks might be analyzed.
Non-parametric statistical methods are discussed later in this chapter.
Equal variance assumption
Many statistical techniques assume equal variances. ANOVA tests the
hypothesis that the means are equal, not that variances are equal. In addition
to assuming normality, ANOVA assumes that variances are equal for each
treatment. Models fitted by regression analysis are evaluated partly by looking
for equal variances of residuals for different levels of Xs and Y.
Minitab’s test for equal variances is found in Stat > ANOVA > Test for
Equal Variances. You need a column containing the data and one or more columns
specifying the factor level for each data point. If the data have already
passed the normality test, use the P-value from Bartlett’s test to test the equal
variances assumption. Otherwise, use the P-value from Levene’s test. The test
shown in Figure 14.3 involved five factor levels and Minitab shows a confidence
interval bar for sigma of each of the five samples; the tick mark in the center of
the bar represents the sample sigma. These are the data from the sample of 100
analyzed earlier and found to be normally distributed, so Bartlett’s test can be
used. The P-value from Bartlett’s test is 0.182, indicating that we can expect
this much variability from populations with equal variances 18.2% of the time.
Since this is greater than 5%, we fail to reject the null hypothesis of equal variances.
Had the data not been normally distributed we would’ve used Levene’s
test, which has a P-value of 0.243 and leads to the same conclusion.
Scatter plots
DefinitionLAscatter diagram is a plot of one variable versus another. One
variable is called the independent variable and it is usually shown on
the horizontal (bottom) axis. The other variable is called the dependent
variable and it is shown on the vertical (side) axis.
UsageLScatter diagrams are used to evaluate cause and effect relationships.
The assumption is that the independent variable is causing a change in
the dependent variable. Scatter plots are used to answer such questions
as ‘‘Does vendor A’s material machine better than vendor B’s?’’ ‘‘Does
the length of training have anything to do with the amount of scrap an
operator makes?’’ and so on.
1. Gather several paired sets of observations, preferably 20 or more. A
paired set is one where the dependent variable can be directly tied to
the independent variable.
2. Find the largest and smallest independent variable and the largest and
smallest dependent variable.
3. Construct the vertical and horizontal axes so that the smallest and
largest values can be plotted. Figure 14.4 shows the basic structure of a
scatter diagram.
4. Plot the data by placing a mark at the point corresponding to each X^Y
pair, as illustrated by Figure 14.5. If more than one classi?cation is
used, you may use di?erent symbols to represent each group.
Regression and correlation analysis 497
Figure 14.3. Output from Minitab’s test for equal variances.
Figure 14.4. Layout of a scatter diagram.
Figure 14.5. Plotting points on a scatter diagram.
From Pyzdek’s Guide to SPCVolume One: Fundamentals, p. 66.
Copyright#1990 by Thomas Pyzdek.
The orchard manager has been keeping track of the weight of peaches on a
day by day basis. The data are provided in Table 14.1.
1. Organize the data into X^Y pairs, as shown in Table 14.1. The independent
variable, X, is the number of days the fruit has been on the
tree. The dependent variable, Y, is the weight of the peach.
2. Find the largest and smallest values for each data set. The largest and
smallest values from Table 14.1 are shown in Table 14.2.
Regression and correlation analysis 499
Table 14.1. Raw data for scatter diagram.
From Pyzdek’s Guide to SPCVolume One: Fundamentals, p. 67.
Copyright#1990 by Thomas Pyzdek.
1 75 4.5
2 76 4.5
3 77 4.4
4 78 4.6
5 79 5.0
6 80 4.8
7 80 4.9
8 81 5.1
9 82 5.2
10 82 5.2
11 83 5.5
12 84 5.4
13 85 5.5
14 85 5.5
15 86 5.6
16 87 5.7
17 88 5.8
18 89 5.8
19 90 6.0
20 90 6.1
3. Construct the axes. In this case, we need a horizontal axis that allows us
to cover the range from 75 to 90 days. The vertical axis must cover the
smallest of the small weights (4.4 ounces) to the largest of the weights
(6.1 ounces). We will select values beyond these minimum requirements,
because we want to estimate how long it will take for a peach to reach
6.5 ounces.
4. Plot the data. The completed scatter diagram is shown in Figure 14.6.
. Scatter diagrams display di?erent patterns that must be interpreted;
Figure 14.7 provides a scatter diagram interpretation guide.
Table 14.2. Smallest and largest values.
From Pyzdek’s Guide to SPCVolume One: Fundamentals, p. 68.
Copyright#1990 by Thomas Pyzdek.
Days on tree (X) 75 90
Weight of peach (Y) 4.4 6.1
Figure 14.6. Completed scatter diagram.
From Pyzdek’s Guide to SPCVolume One: Fundamentals, p. 68.
Copyright#1990 by Thomas Pyzdek.
. Be sure that the independent variable, X, is varied over a su?ciently large
range. When X is changed only a small amount, you may not see a correlation
with Y, even though the correlation really does exist.
. If you make a prediction for Y, for an X value that lies outside of the range
you tested, be advised that the prediction is highly questionable and
should be tested thoroughly. Predicting a Y value beyond the X range
actually tested is called extrapolation.
. Keep an eye out for the e?ect of variables you didn’t evaluate. Often, an
uncontrolled variable will wipe out the e?ect of your X variable. It is also
possible that an uncontrolled variable will be causing the e?ect and you
will mistake the X variable you are controlling as the true cause. This
problem is much less likely to occur if you choose X levels at random. An
example of this is our peaches. It is possible that any number of variables
changed steadily over the time period investigated. It is possible that
these variables, and not the independent variable, are responsible for the
weight gain (e.g., was fertilizer added periodically during the time period
Regression and correlation analysis 501
Figure 14.7. Scatter diagram interpretation guide.
From Pyzdek’s Guide to SPCVolume One: Fundamentals, p. 69.
Copyright#1990 by Thomas Pyzdek.
. Beware of ‘‘happenstance’’ data! Happenstance data are data that were collected
in the past for apurposedi?erentthanconstructing a scatter diagram.
Since little or no control was exercised over important variables, you may
?nd nearly anything. Happenstance data should be used only to get ideas
for further investigation, never for reaching ?nal conclusions. One
common problem with happenstance data is that the variable that is truly
important is not recorded. For example, records might show a correlation
between the defect rate and the shift. However, perhaps the real cause of
defects is the ambient temperature, which also changes with the shift.
. If there is more than one possible source for the dependent variable, try
using di?erent plotting symbols for each source. For example, if the
orchard manager knew that some peaches were taken from trees near a
busy highway, he could use a di?erent symbol for those peaches. He
might ?nd an interaction, that is, perhaps the peaches from trees near the
highway have a di?erent growth rate than those from trees deep within
the orchard.
Although it is possible to do advanced analysis without plotting the scatter
diagram, this is generally bad practice. This misses the enormous learning
opportunity provided by the graphical analysis of the data.
Correlation and regression
Correlation analysis (the study of the strength of the linear relationships
among variables) and regression analysis (modeling the relationship between
one or more independent variables and a dependent variable) are activities of
considerable importance in Six Sigma. A regression problem considers the frequency
distributions of one variable when another is held fixed at each of several
levels. A correlation problem considers the joint variation of two variables,
neither of which is restricted by the experimenter. Correlation and regression
analyses are designed to assist the analyst in studying cause and effect. They
may be employed in all stages of the problem-solving and planning process. Of
course, statistics cannot by themselves establish cause and effect. Proving
cause and effect requires sound scientific understanding of the situation at
hand. The statistical methods described in this section assist the analyst in
performing this task.
A linear model is simply an expression of a type of association between two
variables, x and y. A linear relationship simply means that a change of a given
size in x produces a proportionate change in y. Linear models have the form:
y ? a ? bx ?14:1?
where a and b are constants. The equation simply says that when x changes by
one unit, y will change by b units. This relationship can be shown graphically.
In Figure 14.8, a ? 1 and b ? 2. The term a is called the intercept and b is
called the slope. When x ? 0, y is equal to the intercept. Figure 14.8 depicts a
perfect linear fit, e.g., if x is known we can determine y exactly. Of course, perfect
fits are virtually unknown when real data are used. In practice we must
deal with error in x and y. These issues are discussed below.
Many types of associations are non-linear. For example, over a given range of
x values, y might increase, and for other x values, y might decrease. This curvilinear
relationship is shown in Figure 14.9.
Here we see that y increases when x increases and is less than 1, and decreases
as x increases when x is greater than 1. Curvilinear relationships are valuable
in the design of robust systems. A wide variety of processes produces such relationships.
It is often helpful to convert these non-linear forms to linear form for analysis
using standard computer programs or scientific calculators. Several such transformations
are shown in Table 14.3.
Regression and correlation analysis 503
Figure 14.8. Scatter diagram of a linear relationship.
Fit the straight line YT ? b0 ? b1XT using the usual linear regression procedures
(see below). In all formulas, substitute YT for Y and XT for X. A simple
method for selecting a transformation is to simply program the transformation
into a spreadsheet and run regressions using every transformation. Then select
the transformation which gives the largest value for the statistic R2.
There are other ways of analyzing non-linear responses. One common
method is to break the response into segments that are piecewise linear, and
then to analyze each piece separately. For example, in Figure 14.9 y is roughly
linear and increasing over the range 0 < x < 1 and linear and decreasing over
the range x > 1. Of course, if the analyst has access to powerful statistical software,
non-linear forms can be analyzed directly.
When conducting regression and correlation analysis we can distinguish two
main types of variables. One type we call predictor variables or independent
variables; the other, response variables or dependent variables. By predictor
independent variables we usually mean variables that can either be set to a
desired variable (e.g., oven temperature) or else take values that can be observed
but not controlled (e.g., outdoors ambient humidity). As a result of changes
that are deliberately made, or simply take place in the predictor variables, an
effect is transmitted to the response variables (e.g., the grain size of a composite
material). We are usually interested in discovering how changes in the predictor
variables affect the values of the response variables. Ideally, we hope that a
small number of predictor variables will ‘‘explain’’ nearly all of the variation in
the response variables.
Figure 14.9. Scatter diagram of a curvilinear relationship.
In practice, it is sometimes difficult to draw a clear distinction between independent
and dependent variables. In many cases it depends on the objective of
the investigator. For example, an analyst may treat ambient temperature as a
predictor variable in the study of paint quality, and as the response variable in
a study of clean room particulates. However, the above definitions are useful
in planning Six Sigma studies.
Another idea important to studying cause and effect is that of the data space
of the study The data space of a study refers to the region bounded by the
range of the independent variables under study. In general, predictions based
on values outside the data space studied, called extrapolations, are little more
than speculation and not advised. Figure 14.10 illustrates the concept of data
space for two independent variables. Defining the data space can be quite tricky
when large numbers of independent variables are involved.
Regression and correlation analysis 505
Table 14.3. Some linearizing transformations.
(Source: Experimental Statistics, NBS Handbook 91, pp. 5^31.)
YT XT b0 b1
Y ? a ?
Y 1
a b
Y ? a ? bX
X a b
Y ?
a ? bX
X a b
Y ? abX logY X log a log b
Y ? aebx logY X loga blog e
Y ? aXb log Y log X loga b
Y ? a ? bXn
where n is known
Y Xn a b
While the numerical analysis of data provides valuable information, it
should always be supplemented with graphical analysis as well. Scatter diagrams
are one very useful supplement to regression and correlation analysis.
Figure 14.11 illustrates the value of supplementing numerical analysis with
scatter diagrams.
Figure 14.10. Data space.
Figure 14.11. Illustration of the value of scatter diagrams.
(Source: The Visual Display of Quantitative Information, Edward R. Tufte, pp. 13^14.)
Continued on next page . . .
Statistics for Processes IIV
n ? 11
X ? 9:0
Y ? 7:5
best fit line: Y ? 3 ? 0:5X
standard error of slope: 0:118
t ? 4:24
XX  XX ? 110:0
regression SS ? 27:50
residual SS ? 13:75
r ? 0:82
r2 ? 0:67
Regression and correlation analysis 507
Figure 14.11. Illustration of the value of scatter diagrams.
(Source: The Visual Display of Quantitative Information, Edward R. Tufte, pp. 13^14.)
Figure 14.119Continued . . .
In other words, although the scatter diagrams clearly show four distinct processes,
the statistical analysis does not. In Six Sigma, numerical analysis alone
is not enough.
If all data fell on a perfectly straight line it would be easy to compute the slope
and intercept given any two points. However, the situation becomes more complicated
when there is ‘‘scatter’’ around the line. That is, for a given value of x,
more than one value of y appears. When this occurs, we have error in the
model. Figure 14.12 illustrates the concept of error.
The model for a simple linear regression with error is:
y ? a ? bx ? " ?14:2?
where " represents error. Generally, assuming the model adequately ?ts the
data, errors are assumed to follow a normal distribution with a mean of 0 and a
constant standard deviation. The standard deviation of the errors is known as
the standard error. We discuss ways of verifying our assumptions below.
Figure 14.12. Error in the linear model.
When error occurs, as it does in nearly all ‘‘real-world’’ situations, there are
many possible lines which might be used to model the data. Some method
must be found which provides, in some sense, a ‘‘best-fit’’ equation in these
everyday situations. Statisticians have developed a large number of such methods.
The method most commonly used in Six Sigma finds the straight line that
minimizes the sum of the squares of the errors for all of the data points. This
method is known as the ‘‘least-squares’’ best-fit line. In other words, the leastsquares
best-fit line equation is y0i
? a ? bxi where a and b are found so that
the sum of the squared deviations from the line is minimized. The best-fit
equations for a and b are:
b ? P?Xi  XX??Yi  YY?
P?Xi  XX?2 ?14:3?
a ? YY  b XX ?14:4?
where the sum is taken over all n values. Most spreadsheets and scienti?c calculators
have a built-in capability to compute a and b. As stated above, there
are many other ways to compute the slope and intercept (e.g., minimize the
sum of the absolute deviations, minimize the maximum deviation, etc.); in certain
situations one of the alternatives may be preferred. The reader is advised
to consult books devoted to regression analysis for additional information
(see, for example, Draper and Smith (1981)).
The reader should note that the fit obtained by regressing x on y will not in
general produce the same line as would be obtained by regressing y on x. This
is illustrated in Figure 14.13.
When weight is regressed on height the equation indicates the average weight
(in pounds) for a given height (in inches). When height is regressed on weight
the equation indicates the average height for a given weight. The two lines intersect
at the average height and weight.
These examples show how a single independent variable is used to model the
response of a dependent variable. This is known as simple linear regression. It
is also possible to model the dependent variable in terms of two or more independent
variables; this is known as multiple linear regression. The mathematical
model for multiple linear regression has additional terms for the additional
independent variables. Equation 14.5 shows a linear model when there are two
independent variables.
^y ? a ? b1x1 ? b2x2 ? " ?14:5?
Regression and correlation analysis 509
where x1, x2 are independent variables, b1 is the coe?cient for x1 and b2 is the
coe?cient for x2.
Example of regression analysis
A restaurant conducted surveys of 42 customers, obtaining customer ratings
on staff service, food quality, and overall satisfaction with their visit to the restaurant.
Figure 14.14 shows the regression analysis output from a spreadsheet
regression function (Microsoft Excel).
The data consist of two independent variables, staff and food quality, and a
single dependent variable, overall satisfaction. The basic idea is that the quality
of staff service and the food are causes and the overall satisfaction score is an
effect. The regression output is interpreted as follows:
Multiple RLthe multiple correlation coefficient. It is the correlation
between y and ^ y. For the example: multiple R?0.847, which indicates
that y and ^ y are highly correlated, which implies that there is an association
between overall satisfaction and the quality of the food and
Figure 14.13. Least-squares lines of weight vs. height and height vs. weight.
R squareLthe square of multiple R, it measures the proportion of total variation
about the mean YY explained by the regression. For the example:
R2?0.717, which indicates that the fitted equation explains 71.7% of
the total variation about the average satisfaction level.
AdjustedRsquareLameasure ofR2 ‘‘adjusted for degrees of freedom.’’ The
equation is
Adjusted R2 ? 1  ?1  R2?
n  1
n  p   ?14:6?
where p is the number of parameters (coe?cients for the xs) estimated
in the model. For the example: p ? 2, since there are two x terms.
Some experimenters prefer the adjusted R2 to the unadjusted R2,
while others see little advantage to it (e.g., Draper and Smith, 1981, p.
Standard errorLthe standard deviation of the residuals. The residual is the
difference between the observed values of y and the predicted values
based on the regression equation.
ObservationsLrefer to the number of cases in the regression analysis, or n.
Regression and correlation analysis 511
Figure 14.14. Regression analysis output.
ANOVA,or ANalysis Of VArianceLatable examining the hypothesis that
the variation explained by the regression is zero. If this is so, then the
observed association could be explained by chance alone. The rows
and columns are those of a standard one-factor ANOVA table (see
Chapter 17). For this example, the important item is the column labeled
‘‘Significance F.’’ The value shown, 0.00, indicates that the probability
of getting these results due to chance alone is less than 0.01; i.e., the association
is probably not due to chance alone. Note that the ANOVA
applies to the entire model, not to the individual variables.
The next table in the output examines each of the terms in the linear model
separately. The intercept is as described above, and corresponds to our term a
in the linear equation. Our model uses two independent variables. In our terminology
staff?b1, food? b2. Thus, reading from the coefficients column, the
linear model is: y ? 1:188 ? 0:902  staff score + 0.379  food score. The
remaining columns test the hypotheses that each coefficient in the model is
actually zero.
Standard error columnLgives the standard deviations of each term, i.e., the
standard deviation of the intercept?0.565, etc.
t Stat columnLthe coefficient divided by the standard error, i.e., it shows
how many standard deviations the observed coefficient is from zero.
P-valueLshows the area in the tail of a t distribution beyond the computed t
value. For most experimental work, a P-value less than 0.05 is accepted
as an indication that the coefficient is significantly different than zero.
All of the terms in our model have significant P-values.
Lower 95% and Upper 95% columnsLa 95% confidence interval on the
coefficient. If the confidence interval does not include zero, we will
fail to reject the hypothesis that the coefficient is zero. None of the
intervals in our example include zero.
As mentioned earlier, a correlation problem considers the joint variation of
two variables, neither of which is restricted by the experimenter. Unlike regression
analysis, which considers the effect of the independent variable(s) on a
dependent variable, correlation analysis is concerned with the joint variation
of one independent variable with another. In a correlation problem, the analyst
has two measurements for each individual item in the sample. Unlike a regression
study where the analyst controls the values of the x variables, correlation
studies usually involve spontaneous variation in the variables being studied.
Correlation methods for determining the strength of the linear relationship
between two or more variables are among the most widely applied statistical
techniques. More advanced methods exist for studying situations with more
than two variables (e.g., canonical analysis, factor analysis, principal components
analysis, etc.), however, with the exception of multiple regression, our discussion
will focus on the linear association of two variables at a time.
In most cases, the measure of correlation used by analysts is the statistic r,
sometimes referred to as Pearson’s product-moment correlation. Usually x and
y are assumed to have a bivariate normal distribution. Under this assumption r
is a sample statistic which estimates the population correlation parameter 
One interpretation of r is based on the linear regression model described earlier,
namely that r2 is the proportion of the total variability in the y data which can
be explained by the linear regression model. The equation for r is:
r ?
sxsy ?
nPxy PxPy
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffififfiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi ?nPx2  ?Px?2
?nPy2  ?Py?2
 p ?14:7?
and, of course, r2 is simply the square of r. r is bounded at1 and +1. When the
assumptions hold, the signi?cance of r is tested by the regression ANOVA.
Interpreting r can become quite tricky, so scatter plots should always be used
(see above). When the relationship between x and y is non-linear, the ‘‘explanatory
power’’ of r is difficult to interpret in precise terms and should be discussed
with great care. While it is easy to see the value of very high correlations such
as r ? 0:99, it is not so easy to draw conclusions from lower values of r, even
when they are statistically significant (i.e., they are significantly different than
0.0). For example, r ? 0:5 does not mean the data show half as much clustering
as a perfect straight-line fit. In fact, r ? 0 does not mean that there is no relationship
between the x and y data, as Figure 14.15 shows. When r > 0, y tends to
increase when x increases. When r < 0, y tends to decrease when x increases.
Although r ? 0, the relationship between x and y is perfect, albeit non-linear.
At the other extreme, r ? 1, a ‘‘perfect correlation,’’ does not mean that there
is a cause and effect relationship between x and y. For example, both x and y
might be determined by a third variable, z. In such situations, z is described as
a lurking variable which ‘‘hides’’ in the background, unknown to the experimenter.
Lurking variables are behind some of the infamous silly associations,
such as the association between teacher’s pay and liquor sales (the lurking variable
is general prosperity).*
Regression and correlation analysis 513
*It is possible to evaluate the association of x and y by removing the effect of the lurking variable. This can be done using
regression analysis and computing partial correlation coefficients. This advanced procedure is described in most texts on
regression analysis.
Establishing causation requires solid scientific understanding. Causation
cannot be ‘‘proven’’ by statistics alone. Some statistical techniques, such as
path analysis, can help determine if the correlations between a number of variables
are consistent with causal assumptions. However, these methods are
beyond the scope of this book.
Chi-square, tables
In Six Sigma, there are many instances when the analyst wants to compare
the percentage of items distributed among several categories. The things might
be operators, methods, materials, or any other grouping of interest. From each
of the groups a sample is taken, evaluated, and placed into one of several categories
(e.g., high quality, marginal quality, reject quality). The results can be presented
as a table with m rows representing the groups of interest and k
columns representing the categories. Such tables can be analyzed to answer the
question ‘‘Do the groups differ with regard to the proportion of items in the
categories?’’ The chi-square statistic can be used for this purpose.
Figure 14.15. Interpreting r ? 0 for curvilinear data.
The following example is from Natrella (1963).
Rejects of metal castings were classified by cause of rejection for three different
weeks, as given in the following tabulation. The question to be answered
is: Does the distribution of rejects differ from week to week?
Sand Misrun Shift Drop Corebreak Broken Other Total
Week 1 97 8 18 8 23 21 5 180
Week 2 120 15 12 13 21 17 15 213
Week 3 82 4 0 12 38 25 19 180
Total 299 27 30 33 82 63 39 573
Chi-square (	2) is computed by first finding the expected frequencies in each
cell. This is done using the equation:
Frequency expected ? fe ?
Row sum  column sum
overall sum
For example, for week 1, the frequency expected of sand rejects is (180299)/
573?93.93. The table below shows the frequency expected for the remainder
of the cells.
Sand Misrun Shift Drop Corebreak Broken Other
Week 1 93.93 8.48 9.42 10.37 25.76 19.79 12.25
Week 2 111.15 10.04 11.15 12.27 30.48 23.42 14.50
Week 3 93.93 8.48 9.42 10.37 25.76 19.79 12.25
The next step is to compute 	2 as follows:
	2 ? X over all cells
(Frequency expected  Frequency observed?2
Frequency expected
? ?93:93  97?2
93:93 ?  ??12:25  19?2
12:25 ? 45:60
Analysis of categorical data 515
Next choose a value for ; we will use  ? 0:10 for this example. The degrees
of freedom for the 	2 test are ?k  1??m  1? ? 12. Referring to Table 4 in the
Appendix we find the critical value of 	2 ? 18:55 for our values. Since our computed
value of 	2 exceeds the critical value, we conclude that the weeks differ
with regard to proportions of various types of defectives.
Logistic regression
Logistic regression, like least squares regression, investigates the relationship
between a response variable and one or more predictors. However, linear regression
is used when response variables are continuous, while logistic regression
techniques are used with categorical response variables. We will look at three
different types of logistic regression, based on the type of response variable
being analyzed (see Table 14.4.)
The basic idea behind logistic regression is very simple, as shown in Figure
14.16. X is a hypothetical ‘‘cause’’ of a response. X can be either continuous or
categorical. Y is an event that we are interested in and it must be categorical. A
model can have multiple Xs, but only one response variable. For example, Y
Table 14.4. Types of logistic regression analysis.
Binary 2 two levels Go/not-go, pass/fail,
buy/doesn’t buy, yes/no,
recovers/dies, male/female
Ordinal 3 or more natural ordering of the
Nominal 3 or more no natural ordering of
the levels
Black/white/Hispanic, black
hair/brown hair/blonde hair,
might be whether a prospect purchased a magazine or not, and Xs might be the
age and race of the prospect. The model would produce a prediction of the probability
of a magazine being purchased based on the age and race of the prospect,
which might be used to prioritize a list for telemarketing purposes.
Figure 14.16 illustrates a direct modeling of the proportion responding versus
a predictor variable. The problem is that in the real world the response pattern
can take on a wide variety of forms and a simple model of the proportion
responding as a function of predictors isn’t flexible enough to take on all of the
various shapes. The solution to this is to use a mathematical function, called
the logit, that makes it possible to develop versatile models. The formula for
the logit is shown in Equation 14.8. Although it looks intimidating, it is really
very similar to the equation for a linear regression. Notice that e is raised to a
power that is just a linear function of the Xs. In fact, the power term is just the
multiple linear regression model. However, where linear regression can only
model straight-line functions, the logit takes on a wide variety of curve shapes
as the estimates of the parameters vary. Figure 14.17 shows logit curves for a
few values of 
, with a held constant at 0 (changing  would result in shifting
the curves left or right).
P?x? ?
1 ? e?
nxn ?14:8?
Analysis of categorical data 517
Figure 14.16 Logistic regression.
Odds ratios
When the logit link is used (it’s the default in most software packages, including
Minitab), logistic regression evaluates the odds of some event of interest
happening versus the odds of it not happening. This is done via odds ratios.
‘‘Odds’’ and probabilities are similar, but not identical. In a standard deck of
cards there are 13 different card values, ace, king, queen, etc. The odds of a randomly
selected card being an ace is 12-to-1, i.e., there are 12 non-aces to 1 ace.
The probability of selecting an ace is 1-in-13, i.e., there are 13 choices of which
1 is an ace. In most statistical analyses used in Six Sigma work we use probabilities,
but logistic regression uses odds for its calculations.
Consider a Six Sigma project involving a web site. The goal of the project is to
make it easier for customers to find what they are looking for. A survey was
administered to people who visited the web site and the results in Table 14.5
Figure 14.17. Plot of the logit for  ? 0, 
Table 14.5. Odds ratio example.
Old 50 169
New 26 46
were obtained. The Black Belt wants to know if the design change had an impact
on the customer’s ability to find an answer to their question.
The odds ratio for these data is calculated as follows:
Odds of finding answer with old design ? 50=169 ? 0:296
Odds of finding answer with new design ? 26=46 ? 0:565
Odds ratio ? 0:565=0:296 ? 1:91
It can be seen that the odds of the customer ?nding the answer appears to be
91% better with the new design than with the old design. However, to interpret
this result properly we must know if this improvement is statistically signi
?cant. We can determine this by using binary logistic regression.
Note: another way to analyze these data is to use chi-square. Logistic regression,
in addition to providing a predictive model, will sometimes work when
chi-square analysis will not.
Minitab’s binary logistic regression function is located in the
Stat>Regression menu. The data must be arranged in one of the formats
Minitab accepts. Minitab’s Binary Logistic Regression dialog box (Figure
14.18), shows the input for this problem in columns C1, C2, C3, and C4.
Column C4 is a code value that is 0 if the customer visited after the change, 1
Interpreting Minitab’s binary logistic regression output
There is a great deal of information displayed in Figure 14.19; let’s take a
closer look at it. At the top we see that Minitab used the logit link in the
analysis, which is its default. Next Minitab summarizes the response information,
which matches the input in Table 14.5L(odds ratio example). Next we
see the predictive model coefficients. The coefficient labeled ‘‘Constant’’
(0.5705) is the value for  in Equation 14.8, and the coefficient labeled
‘‘WhenCode’’ is the coefficient for 
. The P column is the test for significance
and P < 0.05 is the critical value. Since P < 0.05 for both the constant and the
WhenCode, we conclude that the constant is not zero and that when the data
were taken (before or after the design change) made a difference.
In the WhenCode row we have three additional columns: odds ratio, 95%
confidence interval lower limit and 95% confidence interval upper limit. The
odds ratio is the 1.91 we calculated directly earlier. The 95% confidence interval
Analysis of categorical data 519
on the odds ratio goes from 1.07 to 3.40. If the design change made no difference,
the expected value of the odds ratio would be 1.00. Since the interval doesn’t
include 1.00 we conclude (at 95% confidence) that the design change made a difference.
This conclusion is confirmed by the P-value of 0.029 for the test that
all slopes are equal (testing for equal slopes is equivalent to testing the null
hypothesis that the design change had no effect).
Had we had a covariate term (an X on a continuous scale) Minitab would’ve
performed a goodness of fit test by dividing the data into 10 groups and performing
a chi-square analysis of the resulting table.
Next Minitab compares the predicted probabilities with the actual
responses. The data are compared pairwise, predicted: found and not found vs.
actual: found and not found. A pair is ‘‘concordant’’ if actual and predicted categories
are the same, ‘‘discordant’’ if they are different, and ‘‘tied’’ otherwise.
Table 14.6 shows the classifications for our example.
The total number of found times not found pairs is 76  215 ? 16340.
The total number of concordant pairs is 169  26 = 4394. The total number of
Figure 14.18. Minitab’s Binary Logistic Regression dialog box.
discordant pairs is 50  46 ? 2300. The remaining 16340  4394  2300 ? 9646 pairs are ties. The model correctly discriminated between and classified
the concordant pairs, or27%. It incorrectly classified the discordant pairs, or 14%.
Somers’ D, Goodman-Kruskal Gamma, and Kendall’s Tau-a are summaries
of the table of concordant and discordant pairs. The numbers have the same
Analysis of categorical data 521
Figure 14.19. Output from Minitab binary logistic regression.
Table 14.6. Concordant and discordant results.
Not found 169 Concordant
Found 50 Discordant
Found 26 Concordant
Not Found 46 Discordant
numerator: the number of concordant pairs minus the number of discordant
pairs. The denominators are the total number of pairs with Somers’ D, the
total number of pairs excepting ties with Goodman-Kruskal Gamma, and the
number of all possible observation pairs for Kendall’s Tau-a. These measures
most likely lie between 0 and 1 where larger values indicate a better predictive
ability of the model. The three summary measures of fit range between 0.05
and 0.31. This isn’t especially impressive, but the P-value and the concordance/
discordance analysis indicate that it’s better than randomly guessing.
The main conclusion is found in the odds ratio and P-value. The new design is
better than the original design. The mediocre predictability of the model indicates
that there’s more to finding the correct answer than the different web
designs. In this case it would probably pay to continue looking for ways to
improve the process, only 36% of the customers find the correct answer (a process
sigma that is less than zero!).
If the response variable has more than two categories, and if the categories
have a natural order, then use ordinal logistic regression. Minitab’s procedure
for performing this analysis assumes parallel logistic regression lines. You may
also want to perform a nominal logistic regression, which doesn’t assume parallel
regression lines, and compare the results. An advantage to using ordinal logistic
regression is that the output includes estimated probabilities for the
response variables as a function of the factors and covariates.
Ordinal logistic regression example
A call center conducted a survey of its customers to determine the impact of
various call center variables on overall customer satisfaction. Customers were
asked to read a statement, then to respond by indicating the extent of their
agreement with the statement. The two survey items we will analyze are:
Q3: The technical support representative was professional. (X)
Q17: I plan to use XXX in the future, should the need arise. (Y)
Customers were asked to choose one of the following responses to each question:
1. I strongly disagree with the statement.
2. I disagree with the statement.
3. I neither agree nor disagree with the statement.
4. I agree with the statement.
5. I strongly agree with the statement.
The results are shown in Table 14.7. Table 14.8 presents the first part of the
Minitab worksheet for the dataLnote that this is the same information as in
Table 14.7, just rearranged. There is one row for each combination of responses
to Q3 and Q17.
Minitab’s dialog box for this example is shown in Figure 14.20. The storage
dialog box allows you to tell Minitab to calculate the probabilities for the
various responses. I also recommend telling Minitab to calculate the number
of occurrences so that you can cross check your frequencies with Minitab’s to
Analysis of categorical data 523
Table 14.7. Survey response cross-tabulation.
Q3 RESPONSE 1 2 3 4 5
1 7 6 7 12 9
2 5 2 8 18 3
3 4 2 20 42 10
4 7 5 24 231 119
5 0 2 14 136 303
Table 14.8. Table 14.7 data reformatted for Minitab.
1 7 1
2 5 1
3 4 1
4 7 1
5 0 1
1 6 2
2 2 2
Etc. Etc. Etc.
assure that you have the data in the correct format. When you tell Minitab
to store results, the information is placed in new columns in your active
worksheet, not in the session window. Note the data entries for the response,
frequency, model, and factors.
Minitab’s session window output is shown in Figure 14.21. For simplicity
only part of the output is shown. The goodness-of-fit statistics (concordance,
discordance, etc.) have been omitted, but the interpretation is the same as for
binary logistic regression. Minitab needs to designate one of the response values
as the reference event. Unless you specifically choose a reference event,
Minitab defines the reference event based on the data type:
* For numeric factors, the reference event is the greatest numeric value.
* For date/time factors, the reference event is the most recent date/
* For text factors, the reference event is the last in alphabetical order.
A summary of the interpretation follows:
* The odds of a reference event is the ratio of P(event) to P(not event).
* The estimated coe?cient can also be used to calculate the odds ratio,
or the ratio between two odds. Exponentiating the parameter estimate
of a factor yields the ratio of P(event)/P(not event) for a certain factor
level compared to the reference level.
Figure 14.20. Ordinal Logistic Regression Minitab dialog boxes.
You can change the default reference event in the Options subdialog box. For
our example, category 5 (strongly agree) is the reference event. The odds ratios
are calculated as the probability of the response being a 5 versus the probability
that it is not a 5. For factors, the smallest numerical value is the reference
event. For the example, this is a Q3 response of 1.
The odds ratios and their confidence intervals are given near the bottom of
the table. A negative coefficient and an odds ratio less than 1 indicate that
higher responses to Q17 tend to be associated with higher responses to Q3.
Odds ratios whose confidence intervals do not include 1.00 are statistically
significant. For the example, this applies to responses of 4 or 5 to Q3, i.e., a
customer who chooses a 4 or 5 in response to Q3 is more likely to choose a
5 in response to Q17.
The statistical probabilities stored by Minitab are plotted in Figure 14.22.
The lines for Q3?4 and Q3?5, the factor categories with significant odds
ratios, are shown as bold lines. Note that the gap between these two lines and
the other lines is greatest for Q17?5.
Analysis of categorical data 525
Figure 14.21. Minitab ordinal logistic regression session window output.
Nominal logistic regression, as indicated in Table 14.4, is used when the
response is categorical, there are two or more response categories, and there is
no natural ordering of the response categories. It can also be used to evaluate
whether the parallel line assumption of ordinal logistic regression is reasonable.
Example of nominal logistic regression
Upon further investigation the Master Black Belt discovered that the Black
Belt working on the web site redesign project described in the binary logistic
regression example section above had captured additional categories. Rather
than just responding that the answer to their question was found or not found,
there were several other response categories (Figures 14.23 and 14.24). Since
the various not found subcategories have no natural order, nominal logistic
regression is the correct procedure for analyzing these data.
The result of Minitab’s analysis, shown in Figure 14.25, shows that only the
odds ratio for found and worked versus not found is significant. The confidence
interval for all other found subcategories compared with found and worked
includes 1.00. The family P-value is a significance test for all comparisons simultaneously.
Since we are making four comparisons, the significance level is
higher than that of each separate test.
Figure 14.22. Minitab stored results.
Analysis of categorical data 527
Figure 14.23. Minitab’s Nominal Logistic Regression dialog box.
Figure 14.24. Minitab nominal logistic regression output.
Comparison with chi-square
If a chi-square analysis is performed on the web redesign data Minitab produces
the output shown in Figure 14.26. Note that the chi-square procedure
prints a warning that there are two cells with less than the recommended minimum
expected frequency of 5.0. It also gives a P-value of 0.116, which is greater
than the critical value of 0.05, leading to a somewhat different conclusion than
the logistic regression analysis. The chi-square test only lets us look at the significance
of the overall result, which is analogous to the ‘‘family P-value’’ test
performed in the nominal logistic regression analysis. However, in this case we
are primarily concerned with the improved odds of finding the correct answer
with the new web design vs. the old web design, which is provided by logit 4 of
the logistic regression.
The most commonly used statistical tests (t-tests, Z-tests, ANOVA, etc.)
are based on a number of assumptions (see testing assumptions above).
Non-parametric tests, while not assumption-free, make no assumption of a
specific distribution for the population. The qualifiers (assuming . . .) for
non-parametric tests are always much less restrictive than for their para-
Figure 14.25. Interpretation of Minitab nominal logistic regression output.
metric counterparts. For example, classical ANOVA requires the assumptions
of mutually independent random samples drawn from normal distributions
that have equal variances, while the non-parametric counterparts
require only the assumption that the samples come from any identical continuous
distributions. Also, classical statistical methods are strictly valid
only for data measured on interval or ratio scales, while non-parametric statistics
apply to frequency or count data and to data measured on nominal
or ordinal scales. Since interval and ratio data can be transformed to nominal
or ordinal data, non-parametric methods are valid in all cases where
classical methods are valid; the reverse is not true. Ordinal and nominal
data are very common in Six Sigma work. Nearly all customer and employee
surveys, product quality ratings, and many other activities produce ordinal
and nominal data.
Non-parametric methods 529
Figure 14.26. Chi-square analysis of web design data.
So if non-parametric methods are so great, why do we ever use parametric
methods? When the assumptions hold, parametric tests will provide greater
power than non-parametric tests. That is, the probability of rejecting H0 when
it is false is higher with parametric tests than with a non-parametric test using
the same sample size. However, if the assumptions do not hold, then nonparametric
tests may have considerably greater power than their parametric
It should be noted that non-parametric tests perform comparisons using
medians rather than means, ranks rather than measurements, and signs of difference
rather than measured differences. In addition to not requiring any distributional
assumptions, these statistics are also more robust to outliers and
extreme values.
The subject of non-parametric statistics is a big one and there are many entire
books written about it. We can’t hope to cover the entire subject in a book
about Six Sigma. Instead, we briefly describe the non-parametric tests performed
by Minitab (Figure 14.27). Minitab’s non-parametric tests cover a reasonably
wide range of applications to Six Sigma work, as shown in Table 14.9.
Figure 14.27. Minitab’s non-parametric tests.
Non-parametric methods 531
Table 14.9. Applications for Minitab’s non-parametric tests.*
1-sample sign Performs a one-sample sign test of the
median and calculates the
corresponding point estimate and
confidence interval.
*1-sample Z-test
*1-sample t-test
1-sample Wilcoxon Performs a one-sample Wilcoxon
signed rank test of the median and
calculates the corresponding point
estimate and confidence interval.
*1-sample Z-test
*1-sample t-test
Mann-Whitney Performs a hypothesis test of the
equality of two population medians and
calculates the correspondingpoint
estimate and confidence interval.
*2-sample t-test
Kruskal-Wallis Kruskal-Wallis performs a hypothesis
test of the equality of population
medians for a one-way design (two or
more populations). This test is a
generalization of the procedure used by
the Mann-Whitney test.
See also: Mood’s median test.
*One-way ANOVA
Mood’s median test Performs a hypothesis test of the
equality of population medians in a
one-way design. Sometimes called a
median test or sign scores test.
Mood’s median test is robust against
outliers and errors in data, and is
particularly appropriate in the
preliminary stages of analysis.
Mood’s median test is more robust
against outliers than the Kruskal-
Wallis test, but is less powerful (the
confidence interval is wider, on the
average) for analyzing data from
many distributions, including data
from the normal distribution.
See also: Kruskal-Wallis test.
*One-way ANOVA
Continued next page . . .
Friedman Performs a non-parametric analysis of a
randomized block experiment.
Randomized block experiments are a
generalization of paired experiments.
The Friedman test is a generalization
of the paired sign test with a null
hypothesis of treatments having no
effect. This test requires exactly one
observation per treatment-block
*2-way ANOVA
*Paired sign test
Runs tests Test whether or not the data order is
random. Use Minitab’s Stat >Quality
Tools >Run Chart to generate a run
Pairwise averages Pairwise averages calculates and stores
the average for each possible pair of
values in a single column, including
each value with itself. Pairwise averages
are also called Walsh averages. Pairwise
averages are used, for example, for the
Wilcoxon method.
Pairwise differences Pairwise differences calculates and
stores the differences between all
possible pairs of values formed from
two columns. These differences are
useful for non-parametric tests and
confidence intervals. For example, the
point estimate given by Mann-Whitney
can be computed as the median of the
Pairwise slopes Pairwise slopes calculates and stores
the slope between all possible pairs of
points, where a row in y-x columns
defines a point in the plane. This
procedure is useful for finding robust
estimates of the slope of a line through
the data.
*Simple linear regression
Table 14.99Continued.
Continued next page . . .
Non-parametric methods 533
Levene’s test Test for equal variances. This method
considers the distances of the
observations from their sample median
rather than their sample mean. Using
the sample median rather than the
sample mean makes the test more
robust for smaller samples.
*Bartlett’s test
Non-parametric Dist
Analyzes times-to-failure when no
distribution can be found to fit the
(censored) data. Tests for the equality
of survival curves.
*Parametric Dist
Hazard plotsLnonparametric
If data are right censored, plots empirical
hazard function or actuarial estimates.
If data are arbitrarily censored, plots
actuarial estimates.
*Hazard plotsL
parametric distribution
*#All Rights Reserved. 2000 Minitab, Inc. Used by permission.
Table 14.99Continued.
Guidelines on when to use non-parametric tests
Use non-parametric analysis when any of the following are true (Gibbons,
1. The data are counts or frequencies of di?erent types of outcomes.
2. The data are measured on a nominal scale.
3. The data are measured on an ordinal scale.
4. The assumptions required for the validity of the corresponding parametric
procedure are not met or cannot be veri?ed.
5. Theshapeof the distributionfromwhichthesampleisdrawnisunknown.
6. The sample size is small.
7. The measurements are imprecise.
8. There are outliers and/or extreme values in the data, making the median
more representative than the mean.
Use a parametric procedure when both of the following are true:
1. The data are collected and analyzed using an interval or ratio scale of
2. All of the assumptions required for the validity of that parametric procedure
can be veri?ed.
^ ^ ^
Managing Six Sigma
The dictionary de?nes the word project as follows:
1. A plan or proposal; a scheme. See synonyms at plan.
2. An undertaking requiring concerted e?ort.
Under the synonym plan we find:
1. A scheme, program, or method worked out beforehand for the accomplishment
of an objective: a plan of attack.
2. A proposed or tentative project or course of action.
3. A systematic arrangement of important parts.
Although truly dramatic improvement in quality often requires transforming
the management philosophy and organization culture, the fact is that,
sooner or later, projects must be undertaken to make things happen. Projects
are the means through which things are systematically changed, projects are
the bridge between the planning and the doing.
Frank Gryna makes the following observations about projects (Juran and
Gryna, 1988, pp. 22.18^22.19):
. An agreed-upon project is also a legitimate project. This legitimacy puts
the project on the o?cial priority list. It helps to secure the needed bud-
*Some of the material in this chapter is from The Six Sigma Project Planner, by Thomas Pyzdek.#2003 by McGraw-Hill.
Copyright 2003 by The McGraw-Hill Companies, Inc. Click Here for Terms of Use.
gets, facilities, and personnel. It also helps those guiding the project to
secure attendance at scheduled meetings, to acquire requested data, to
secure permission to conduct experiments, etc.
. The project provides a forum of converting an atmosphere of defensiveness
or blame into one of constructive action.
. Participation in a project increases the likelihood that the participant will
act on the ?ndings.
. All breakthrough is achieved project by project, and in no otherway.
The last item represents both good news and bad news. The bad news is that
few projects are truly successful; the good news is that companies can and do
become proficient at implementing projects without the need for mystical
powers. What is needed is effective project management.
Project management is a system for planning and implementing change that
will produce the desired result most efficiently. There are a number of tools
and techniques that have been found useful in project management. Brief
descriptions of the major project management methods are provided here.
Techniques specific to project management are covered in greater detail elsewhere
in this chapter. Many of these tools are used in a wide variety of quality
improvement and quality control situations in addition to project management;
additional information on each of these more general techniques is found elsewhere
in this book; consult the index for details.
Project planLTheproject plan shows the ‘‘why’’ and the ‘‘how’’ of a project.
A good project plan will include a statement of the goal, a cost/ benefit analysis,
a feasibility analysis, a listing of the major steps to be taken, a timetable for completion,
and a description of the resources required (including human
resources) to carry out the project. The plan will also identify objective measures
of success that will be used to evaluate the effectiveness of the proposed
changes; these are sometimes called the ‘‘deliverables’’ of the project.
Gantt chartLA Gantt chart shows the relationships among the project
tasks, along with time constraints. See below for a discussion of Gantt charts.
Milestone chartsLA Gantt chart modified to provide additional information
on project status. See below for a discussion of milestone charts.
Pareto analysisLPareto analysis is a technique that helps one to rank
opportunities to determine which of many potential projects should be pursued
first. It can also be used sequentially to determine which step to take next. The
Useful project management tools and techniques 535
Pareto principle has been described by Juran as separating the ‘‘vital few’’ from
the ‘‘trivial many.’’ It is the ‘‘why’’ and the ‘‘benefit’’ of the project plan. See
Chapter 8 for additional discussion.
BudgetLAbudget is an itemized summary of estimated or intended expenditures
for a given project along with proposals for financing them. Project budgets
present management with a systematic plan for the expenditure of the
organization’s resources, such as money or time, during the course of the project.
The resources spent include time of personnel, money, equipment utilization
and so on. The budget is the ‘‘cost’’ portion of the project plan. Also see
Process decision program chart (PDPC)LThe PDPC technique is used to
develop contingency plans. It is modeled after reliability engineering methods
such as failure mode, effects, and criticality analysis (FMECA) and fault tree
analysis (FTA). The emphasis of PDPC is the impact of problems on project
plans. PDPCs are accompanied by specific actions to be taken should the problems
occur to mitigate the impact of the problems. PDPCs are useful in the
planning of projects in developing a project plan with a minimum chance of
encountering serious problems. Also see Chapter 8.
Quality function deployment (QFD)LTraditionally, QFD is a system for
design of a product or service based on customer demands, a system that
moves methodically from customer requirements to requirements for the
products or services.QFDprovides the documentation for the decision-making
process. QFD can also be used to show the ‘‘whats’’ and ‘‘hows’’ of a project.
Used in this way QFD becomes a powerful project planning tool. Also see
Chapter 3.
Matrix chartLA matrix chart is a simplified application of QFD (or, perhaps,
QFD is an elaborate application of matrix charts). This chart is constructed
to systematically analyze the correlations between two groups of
ideas. When applied to project management the two ideas might be, for example
1) what is to be done? 2) who is to do it? Also see Chapter 8.
Arrow diagramsLArrow diagrams are simple network representations of
project flows. They show which tasks must be completed in the project and
the order in which the tasks must be completed. See Chapter 8. Arrow diagrams
are a simplification of PERT-type systems (see below).
There are several reasons why one should plan carefully before starting a
project (Ruskin and Estes, 1995, p. 44):
1. The plan is a simulation of prospective project work, which allows ?aws
to be identi?ed in time to be corrected.
2. The plan is a vehicle for discussing each person’s role and responsibilities,
thereby helping direct and control the work of the project.
3. The plan shows how the parts ?t together, which is essential for coordinating
related activities.
4. The plan is a point of reference for any changes of scope, thereby helping
project managers deal with their customers.
5. The plan helps everyone know when the objectives have been reached
and therefore when to stop.
The project plan shows the ‘‘why’’ and the ‘‘how’’ of a project.Agood project
plan will include the following elements:
. statement of the goal
. cost/bene?t analysis
. feasibility analysis
. listing of the major steps to be taken
. timetable for completion
. description of the resources required (including human resources) to
carry out the project
The plan will also identify objective measures of success that will be used to
evaluate the effectiveness of the proposed changes; these are sometimes called
the ‘‘deliverables’’ of the project.
Most projects important enough to have a significant impact on quality
are too large to tackle all at once. Instead, large projects must be broken
down into smaller projects and, in turn, into specific work elements and
tasks. The process of going from project objectives to tasks is called decomposition.
Project decomposition begins with the preparation of a preliminary
plan. A preliminary project plan will identify, in broad high-level terms, the
objectives of the project and constraints in terms of time and resources.
The work to be performed should be described and precedence relationships
should be sketched out. Preliminary budgets and schedules will be
developed. Finally, subplans will be developed for each subproject for the
. Control plans
^ Quality control plans
^ Cost control plans
^ Schedule control plans
Project planning 537
. Sta?ng plans
. Material plans
. Reporting plans
. Other plans as deemed necessary
These subplans are developed in parallel for the various subprojects.
Also see cross-functional collaboration, below.
Projects should be selected consistent with the organization’s overall
strategy and mission. Because of this global perspective most projects
involve the efforts of several different functional areas. Not only do individual
quality projects tend to cut across organizational boundaries, different
projects are often related to one another. To effectively manage this complexity
it is necessary to integrate the planning and execution of projects
(For additional details on teams see Chapter 5.)
Teams are chartered by senior leadership, generally the only group with the
necessary authority to designate cross-functional responsibilities and allow
access to interdepartmental resources. The team facilitator should ask senior
leadership to put the problem statement in writing. The problem statement
should be specific enough to help the team identify the scope of the project
and the major stakeholders. Problems of gargantuan proportions should be subdivided
into smaller projects.
There are six steps in the chartering process:
1. Obtaining a problem statement
2. Identifying the principal stakeholders
3. Creating a macro ?ow chart of the process
4. Selecting the team members
5. Training the team
6. Selecting the team leader
The official authorization for the project should be summarized in a document
like that shown in the Six Sigma project charter below.
Six Sigma Project Charter.
Project Name/
Sponsoring Organization
Key Leadership Name Phone Number Mail Station
Project Black Belt
Project Green Belt
Team Members Title/Role Phone Number Mail Station
Support Personnel
Financial Adviser
Key Stakeholders Title Phone Number Mail Station
Date Chartered Project Start Date Target Completion Date
Revision Number: Date
Sponsor Signature:
Project charter 539
Continued next page . . .
Project Name/Number
ProjectMission Statement
Problem Statement (‘‘What’s wrong with the status quo?’’)
Business NeedAddressed by Project (‘‘What is the ‘Burning Platform’ for this project?’’)
Project Scope (Product or Service Created by this Project (Deliverables))
Resources Authorized for Project (include Charge Number)
Six Sigma Phase Status (DMAIC projects)
Six Sigma Stage Summary
Completion Status
Project Completion Barriers Encountered (Top 3)
# Issue Lessons Learned
Several problems with projects appear repeatedly:
. Projects have little or no impact on the organization’s success, even if successful,
no one will really care.
. Missions overlap the missions of other teams. E.g., Team A’s mission is to
reduce solder rejects, Team B’s mission is to reduce wave solder rejects,
Team C’s mission is to reduce circuit board assembly problems.
. Projects improve processes that are scheduled for extensive redesign or
discontinuation. For example, working on improving work ?ow for a production
process that is to be relocated to another factory.
. Studying a huge system (‘‘patient admitting’’), rather than a manageable
process (‘‘outpatient surgery preadmission’’).
. Studying symptoms (‘‘touch-up of defective solder joints’’) rather than
root causes (‘‘wave solder defects’’)
. Project deliverables are unde?ned. E.g., ‘‘Study TQM’’ rather than
‘‘Reduce waiting time in Urgent Care.’’
Ruskin and Estes (1995) define work breakdown structures (WBS) as a process
for defining the final and intermediate products of a project and their relationships.
Defining project tasks is typically complex and accomplished by a
series of decompositions followed by a series of aggregations. For example, a
software project to develop an SPC software application would disaggregate
the customer requirements into very specific analytic requirements (e.g., the
customer’s requirement that the product create X-bar charts would be decomposed
into analytic requirements such as subroutines for computing subgroup
means and ranges, plotting data points, drawing lines, etc.). Aggregation would
involve linking the various modules to produce an X-bar chart displayed on
the screen.
The WBS can be represented in a tree diagram, as shown in Figure 15.1.
Preliminary requirements WBSLis a statement of the overall requirements
for the project as expressed by the customer (e.g., the deliverables or
‘‘product’’), and subsidiary requirements as expressed by management (e.g.,
billing, reports required).
Detailed plan WBSLbreaks down the product into subproducts.
Requirements are listed for each subproduct (e.g., tooling, staff). The subproducts
are, in turn, broken down into their subproducts, etc., until a reasonable
limit is reached. All work begins at the lowest level. Detailed plans for
each subsystem include control plans for quality, cost and schedule, staffing
plans, materials plans, reporting plans, contingency plans, and work authorization
plans. In addition, the overall detailed plan covers objectives, con-
Project charter 541
straints, precedence relationships, timetables, budgets, and review and reporting
Typical subsystem WBSLare created, i.e., the process just described is
performed for each subsystem. Subsystems are then built.
Integration WBSLdetail how the various subsystems will be assembled
into the product deliverables. This usually involves integrating into larger subsystems,
then still larger subsystems, etc., to the highest level of integration.
Validation WBSLplans explain how the various system integrations will be
measured and tested to assure that the final requirements will be met.
Figure 15.1. WBS of a spacecraft system.
From Ruskin, A.M. and Estes, W.E. What Every Engineer Should Know About Project
Management, Second Edition. Copyright#1995 by Marcel Dekker, Inc.
Reprinted with permission.
The project plan is itself an important feedback tool. It provides details on
the tasks that are to be performed, when they are to be performed, and how
much resource is to be consumed. The plan should also include explicit provisions
for feedback. Typical forms of feedback are:
. Status reportsLFormal, periodic written reports, often with a standardized
format, telling what the project is based on, and where it is supposed
to be relative to the plan. Where project performance does not match
planned performance, the reports include additional information as to
the cause of the problem and what is being done to bring the project into
alignment with the plan. Remedial action may, at times, involve revising
the plan. When the project is not meeting the plan due to obstacles which
the project team cannot overcome, the status report will request senior
management intervention.
. Management reviewsLThese are meetings, scheduled in advance, where
the project leader will have the opportunity to interact with key members
of the management team. The chief responsibility for these meetings is
management’s. The purpose is to brief management on the status of the
project, review the project charter and project team mission, discuss
those management activities likely to have an impact on the progress of
the team, etc. This is the appropriate forum for addressing systems barriers
encountered by the team: while the team must work within existing
systems, management has the authority to change the systems. At times a
minor system change can dramatically enhance the ability of the team to
. Budget reviewsLWhile budget reports are included in each status report,
a budget review is a formal evaluation of actual resource utilization with
respect to budgeted utilization. Budget review may also involve revising
budgets, either upward or downward, based on developments since the
original budget approval. Among those unschooled in the science of statistics
there is an unfortunate tendency to react to every random tick in budget
variances as if they were due to a special cause of variation. Six Sigma
managers should coach ?nance and management personnel on the principles
of variation to preclude tampering with the budgeting process (also
see below).
. Customer auditsLThe ‘‘customer’’ in this context means the principal
stakeholder in the project. This person is the ‘‘owner’’ of the process
being modi?ed by the project. The project deliverables are designed to
meet the objectives of this customer, and the customer should play an
active role in keeping the project on track to the stated goals.
Feedback loops 543
. Updating plans and timetablesLThe purpose of feedback is to provide
information to form a basis for modifying future behavior. Since that
behavior is documented in the project plans and schedules, these documents
must be modi?ed to ensure that the appropriate action is taken.
Remember, in the PDCA cycle, plans change ?rst.
. Resource redirectionLThe modi?cations made to the plans and timetables
will result in increasing or decreasing resource allocation to the
project, or accelerating or decelerating the timetable for resource
utilization. The impact of these resource redirections on other projects
should be evaluated by management in view of the organization’s overall
There are a wide variety of tools and techniques available to help the project
manager develop a realistic project timetable, to use the timetable to time the
allocation of resources, and to track progress during the implementation of the
project plan. We will review two of the most common here: Gantt charts and
PERT-type systems.
Gantt charts
Gantt chartLA Gantt chart shows the relationships among the project
tasks, along with time constraints. The horizontal axis of a Gantt chart shows
the units of time (days, weeks, months, etc.). The vertical axis shows the activities
to be completed. Bars show the estimated start time and duration of the
various activities. Figure 15.2 illustrates a simple Gantt chart.
Figure 15.2. Gantt chart.
Milestone chartsLGantt charts are often modified in a variety of ways to
provide additional information. One common variation is shown in Figure
15.3. The milestone symbol represents an event rather than an activity; it does
not consume time or resources. When Gantt charts are modified in this way
they are sometimes called ‘‘milestone charts.’’
Gantt charts and milestone charts can be modified to show additional information,
such as who is responsible for a task, why a task is behind schedule,
remedial action planned or already taken, etc.
Typical DMAIC project tasks and responsibilities
Although every project is unique, most Six Sigma projects which use the
DMAIC framework have many tasks in common, at least at a general level.
Many people find it helpful if they have a generic ‘‘template’’ they can use to
plan their project activities. This is especially true when the Black Belt or
Green Belt is new and has limited project management experience. Table 15.1
can be used as a planning tool by Six Sigma teams. It shows typical tasks, responsibilities
and tools for each major phase of a typical Six Sigma project.
PERT-CPM-type project management systems
While useful, Gantt charts and their derivatives provide limited project schedule
analysis capabilities. The successful management of large-scale projects
requires more rigorous planning, scheduling and coordinating of numerous
interrelated activities. To aid in these tasks, formal procedures based on the
Performance measures 545
Figure 15.3. Enhanced Gantt chart (milestone chart).
Table 15.1. Typical DMAIC project tasks and responsibilities.
Charter Project
&Identify opportunity for improvement Black Belt
&Identify sponsor Black Belt
&Estimate savings Black Belt
&Draft project charter Black Belt, sponsor
&Sponsor project review (weekly) Sponsor, Black Belt
&Team selection Sponsor, Black Belt
&Complete project charter Black Belt
&Team training Black Belt, Green Belt
&Review existing process documentation Team member, process expert
&De?ne project objectives and plan Team
&Present objectives and plan to management Green Belt
&De?ne and map as-is process Team, process expert
&Review and re-de?ne problem, if necessary Team
&Identify CTQs Green Belt, Black Belt
&Collect data on subtasks and cycle time Team
&Validate measurement system Black Belt, process operator
&Prepare baseline graphs on subtasks/cycle time Black Belt, Green Belt
&Analyze impacts, e.g., subtasks, Pareto . . . Black Belt, Green Belt
&Use subteams to analyze time and value, risk
&Benchmark other companies Team member
Continued next page . . .
use of networks and network techniques were developed beginning in the late
1950s. The most prominent of these procedures have been PERT (Program
Evaluation and Review Technique) and CPM (Critical PathMethod). The two
approaches are usually referred to as PERT-type project management systems.
The most important difference between PERT and CPM is that originally the
time estimates for the activities were assumed deterministic in CPM and were
probabilistic in PERT. Today, PERT and CPM actually comprise one technique
and the differences are mainly historical.
Project scheduling by PERT-CPM consists of four basic phases: planning,
scheduling, improvement, and controlling The planning phase involves break-
Performance measures 547
&Discuss subteams’ preliminary ?ndings Team
&Consolidate subteams’ analyses/?ndings Team
&Present recommendations to process owners and
Sponsor, team
&Review recommendations/formulate pilot Team, Black Belt
&Prepare for improved process pilot Team, process owner
&Test improved process (run pilot) Process operator
&Analyze pilot and results Black Belt, Green Belt
&Develop implementation plan Team, process owner
&Prepare ?nal presentation Team
&Present ?nal recommendations to management team Green Belt
&De?ne control metrics Black Belt, Green Belt, process
&Develop metrics collection tool Black Belt
&Roll-out improved process Process owner
&Roll-out control metrics Process owner
&Monitor process monthly using control metrics Process owner, Black Belt
Table 15.19Continued.
ing the project into distinct activities. The time estimates for these activities are
then determined and a network (or arrow) diagram is constructed with each
activity being represented by an arrow.
PERT-type systems are used to:
. Aid in planning and control of projects
. Determine the feasibility of meeting speci?ed deadlines
. Identify the most likely bottlenecks in a project
. Evaluate the e?ects of changes in the project requirements or schedule
. Evaluate the e?ects of deviating from schedule
. Evaluate the e?ect of diverting resources from the project, or redirecting
additional resources to the project.
The ultimate objective of the scheduling phase is to construct a time chart
showing the start and finish times for each activity as well as its relationship
to other activities in the project. The schedule must identify activities that are
‘‘critical’’ in the sense that they must be completed on time to keep the project
on schedule.
It is vital not to merely accept the schedule as a given. The information
obtained in preparing the schedule can be used to improve the project schedule.
Activities that the analysis indicates to be critical are candidates for improvement.
Pareto analysis can be used to identify those critical elements that are
most likely to lead to significant improvement in overall project completion
time. Cost data can be used to supplement the time data, and the combined
time/cost information analyzed using Pareto analysis.
The final phase in PERT-CPM project management is project control. This
includes the use of the network diagram and Gantt chart for making periodic
progress assessments
The following is based on an example from Hillier and Lieberman (1980).
Let’s say that we wish to use PERT on a project for constructing a house. The
activities involved, and their estimated completion times, are presented in
Table 15.2.
Now, it is important that certain of these activities be done in a particular
order. For example, one cannot put on the roof until the walls are built. This is
called a precedence relationship, i.e., the walls must precede the roof. The network
diagram graphically displays the precedence relationships involved in constructing
a house. A PERT network for constructing a house is shown in
Figure 15.4 (incidentally, the figure is also an arrow diagram).
There are two time-values of interest for each event: its earliest time of completion
and its latest time of completion. The earliest time for a given event is
the estimated time at which the event will occur if the preceding activities are
started as early as possible. The latest time for an event is the estimated time
the event can occur without delaying the completion of the project beyond its
earliest time. Earliest times of events are found by starting at the initial event
and working forward, successively calculating the time at which each event
will occur if each immediately preceding event occurs at its earliest time and
each intervening activity uses only its estimated time.
Slacktime for an event is the difference between the latest and earliest times
for a given event. Thus, assuming everything else remains on schedule, the
slack for an event indicates how much delay in reaching the event can be tolerated
without delaying the project completion.
Events and activities with slack times of zero are said to lie on the critical path
for the project. A critical path for a project is defined as a path through the network
such that the activities on this path have zero slack. All activities and
events having zero slack must lie on a critical path, but no others can. Figure
15.5 shows the activities on the critical path for the housing construction project
as thick lines.
Performance measures 549
Table 15.2. Activities involved in constructing a house.
Excavate 2
Foundation 4
Rough wall 10
Rough electrical work 7
Rough exterior plumbing 4
Rough interior plumbing 5
Wall board 5
Flooring 4
Interior painting 5
Interior ?xtures 6
Roof 6
Exterior siding 7
Exterior painting 9
Exterior ?xtures 2
Project managers can use the network and the information obtained from the
network analysis in a variety of ways to help them manage their projects. One
way is, of course, to pay close attention to the activities that lie on the critical
path. Any delay in these activities will result in a delay for the project.
However, the manager should also consider assembling a team to review the
network with an eye towards modifying the project plan to reduce the total
Figure 15.4. Project network for constructing a house.
Source: Based on Introduction to Operations Research, 3rd Edition, Hillier and Lieberman.
Copyright#1980 by Holden-Day, Inc., San Francisco, California.
time needed to complete the project. The manager should also be aware that the
network times are based on estimates. In fact, it is likely that the completion
times will vary. When this occurs it often happens that a new critical path
appears. Thus, the network should be viewed as a dynamic entity which should
be revised as conditions change.
Primary causes of slippage include poor planning and poor management of
the project. Outside forces beyond the control of the project manager will
often play a role. However, it isn’t enough to be able to simply identify ‘‘outside
forces’’ as the cause and beg forgiveness. Astute project managers will anticipate
as many such possibilities as possible and prepare contingency plans to deal
with them. The PDPC technique is useful in this endeavor. Schedule slippage
should also be addressed rigorously in the schedule control plan, which was
mentioned earlier as a primary deliverable from the project planning process.
Performance measures 551
Figure 15.5. Critical path for house construction example.
The control plan should make provision for reviews conducted at intervals frequent
enough to assure that any unanticipated problems are identified before
schedule slippage becomes a problem.
Resources are those assets of the firm, including the time of employees, that
are used to accomplish the objectives of the project. The project manager should
define, negotiate, and secure resource commitments for the personnel, equipment,
facilities, and services needed for the project. Resource commitments
should be as specific as possible. Generally, resource utilization is specified in
the project budget (see below).
The following items should be defined and negotiated:
. What will be furnished?
. By whom?
. When?
. How will it be delivered?
. How much will it cost?
^ Who will pay?
^When will payment be made?
Resource conflicts
Of course, there are always other opportunities for utilizing resources. On
large projects, conflicts over resource allocation are inevitable. It is best if
resource conflicts can be resolved between those managers directly involved.
However, in some cases, resource conflicts must be addressed by higher levels
of management. Senior managers should view resource conflicts as potential
indications that the management system for allocating resources must be modified
or redesigned. Often, such conflicts create ill will among managers and
lead to lack of support, or even active resistance to the project. Too many such
conflicts can lead to resentment towards quality improvement efforts in general.
Most project schedules can be compressed, if one is willing to pay the additional
costs. For the analysis here, costs are defined to include direct elements
only. Indirect costs (administration, overhead, etc.) will be considered in the
final analysis. Assume that a straight-line relationship exists between the cost
of performing an activity on a normal schedule, and the cost of performing the
activity on a crash schedule. Also assume that there is a crash time beyond
which no further time saving is possible, regardless of cost. Figure 15.6 illustrates
these concepts.
For a given activity the cost-per-unit-of-time saved is found as
crash cost  normal cost
normal time  crash time ?15:1?
When deciding which activity on the critical path to improve, one should
begin with the activity that has the smallest cost-per-unit-of-time saved. The project
manager should be aware that once an activity time has been reduced there
may be a new critical path. If so, the analysis should proceed using the updated
information, i.e., activities on the new critical path should be analyzed.
The data for the house construction example are shown in Table 15.3, with
additional data for costs and crash schedule times for each activity.
Activities shown in bold are on the critical path; only critical path activities
are being considered since only they can produce an improvement in overall
project duration. Thus, the first activity to consider improving would be foundation
work, which costs $800 per day saved on the schedule (identified with
Performance measures 553
Figure 15.6 Cost-time relationship for an activity.
an asterisk [*] in Table 15.3). If additional resources could be directed towards
this activity it would produce the best ‘‘bang for the buck’’ in terms of reducing
the total time of the project. Next, assuming the critical path doesn’t change,
would be excavation, then exterior painting, etc.
As activities are addressed one by one, the time it takes to complete the project
will decline, while the direct costs of completing the project will increase.
Figure 15.7 illustrates the cost-duration relationship graphically.
Conversely, indirect costs such as overhead, etc., are expected to increase as
projects take longer to complete. When the indirect costs are added to the direct
costs, total costs will generally follow a pattern similar to that shown in Figure
To optimize resource utilization, the project manager will seek to develop a
project plan that produces the minimum cost schedule. Of course, the organization
will likely have multiple projects being conducted simultaneously,
which places additional constraints on resource allocation.
Table 15.3. Schedule costs for activities involved in constructing a house.
Normal Schedule Crash Schedule
ACTIVITY Time (days) Cost Time (days) Cost Slope
Excavate 2 1000 1 2000 1000
Foundation 4 1600 3 2400 800*
Rough wall 10 7500 6 14000 1625
Rough electrical work 7 7000 4 14000 2333
Rough exterior plumbing 4 4400 3 6000 1600
Rough interior plumbing 5 3750 3 7500 1875
Wall board 5 3500 3 7000 1750
Flooring 4 3200 2 5600 1200
Interior painting 5 3000 3 5500 1250
Interior ?xtures 6 4800 2 11000 1550
Roof 6 4900 2 12000 1775
Exterior siding 7 5600 3 12000 1600
Exterior painting 9 4500 5 9000 1125
Exterior ?xtures 2 1800 1 3200 1400
Project information should be collected on an ongoing basis as the project
progresses. Information obtained should be communicated in a timely fashion
to interested parties and decision-makers. The people who receive the
information can often help the project manager to maintain or recover the
schedule. There are two types of communication involved: feedback and
Performance measures 555
Figure 15.7. Direct costs as a function of project duration.
Figure 15.8. Total costs as a function of project duration.
feedforward. Feedback is historical in nature and includes such things as
performance to schedule, cost variances (relative to the project budget),
and quality variances (relative to the quality plan). The reader will recall
that initial project planning called for special control plans in each of
these three areas. Feedforward is oriented towards the future and is primarily
concerned with heading off future variances in these three areas.
Information reporting formats commonly fall into one of the following
. formal, written reports
. informal reports and correspondence
. presentations
. meetings
. guided tours of the project, when feasible
. conversations
The principles of effective communication should be kept constantly in
mind. The choice of format for the communication should consider the nature
of the audience and their needs and the time and resources available.
Audiences can be characterized along five dimensions (Ruskin and Estes,
1. Audience diversity
2. Audience sophistication
3. Audience familiarity with the subject matter
4. Audience size and geographic location
5. Audience need to know
The report or presentation should be planned to avoid wasting the time of the
audience members, or the time of those preparing the report or presentation.
Objectives should be stated and the steps necessary to meet the objectives
should be clearly stated in the plan. It may help to consider the communication
as a ‘‘lesson’’ and the plan as a ‘‘lesson plan.’’ Provision should be made for assuring
that the objectives were, in fact, met.
Project communication is a process and, like all processes, it can be
improved. The tools of Six Sigma are designed for just this purpose.
Measurements can be analyzed using the quality control tools described in
Chapters 9 and 10 and used to improve the process of project management.
The PDCA cycle also applies to project management.
Relevant stakeholders
Large quality improvement projects impact large numbers of people within
the organization. Those impacted are known as ‘‘stakeholders’’ in the project.
As far as is practicable, the interests of stakeholders should be aligned with the
objectives of the project. If this is not the case, when stakeholders act according
to their own interests they will be acting to sabotage the project, intentionally
or unintentionally.
Identifying project stakeholders begins with obtaining a project charter.
Once the project charter has been finalized, the project team should prepare a
list of potential stakeholders and their roles. If the project will have significant
impact on hourly employees, they should be involved as well. If the workers
are unionized, the union should be informed. Sell all stakeholders on the merits
of the project. People resist changes unless they see the value in them and the
urgency to take action. Stakeholders must be identified and their needs analyzed
so that an action plan can be created to meet the needs and gain commitment.
To avoid problems, the project team must constantly communicate with the
Stakeholder focus groups are a method that allows group members to
evaluate the potential impact of a plan by identifying the stakeholders affected
by or having influence over the project plan. The focus group approach is a
highly structured method in which the project team first identifies the stakeholders
and their assumptions, then brings those identified together to elicit
their responses to the proposed project (see Chapter 3 for a discussion of the
focus group technique). The team then rates these assumptions for importance
to the stakeholders and importance to the plan. A stakeholder satisfaction
plan may be developed to assure the support of key individuals and
As soon as possible the project manager should arrange a short, informal
meeting with all of these individuals identified as being impacted, including
one executive who sits on the Six Sigma Council (but not the entire council).
The project manager and process owner are letting the stakeholders know that
a project is about to be undertaken in ‘‘their’’ area, with the permission and
direction of the senior executives. This meeting also represents an informal
invitation for the middle managers to challenge the decision to conduct the
project. It is important to allow the managers about a week to attempt to reverse
the leadership’s decision to pursue the project. If concrete information suggests
that tampering or sabotage is occurring, the project manager or process owner
should immediately bring it to the attention of the senior executives who
approved the project charter. The senior leadership should resolve the issue
If a week or so passes without clear opposition to the project, the project
manager should proceed with the implementation of the project plan. Of
course, the lines of communication should remain open throughout the implementation
of the project plan.
Performance measures 557
In this section we will provide an overview of budgeting as it applies to project
The project manager must know where he stands in terms of expenditures.
Once he is informed that a given amount of future expense is allocated to him
for a particular project, it is his job to run the project so that this allowance is
not exceeded. The process of allocating resources to be expended in the future
is called budgeting. Budgets should be viewed as forecasts of future events, in
this case the events are expenditures. A listing of these expenditures, broken
out into specific categories, is called the budget.
Ruskin and Estes (1995) list the following types of project-related budgets:
Direct labor budgets are usually prepared for each work element in the project
plan, then aggregated for the project as a whole. Control is usually maintained
at the work element level to assure the aggregate budget allowance is
not exceeded. Budgets may be in terms of dollars or some other measure of
value, such as direct labor hours expended.
Support services budgets need to be prepared because, without budgets, support
services tend to charge based on actual costs, without allowances for errors,
rework, etc. The discipline imposed by making budget estimates and being
held to them often leads to improved efficiency and higher quality.
Purchased items budgets cover purchased materials, equipment, and services.
The budgets can be based on negotiated or market prices. The issues mentioned
for support services also apply here.
Budgets allocate resources to be used in the future. No one can predict the
future with certainty. Thus, an important element in the budgeting process is
tracking actual expenditures after the budgets have been prepared. The following
techniques are useful in monitoring actual expenditures versus budgeted
Expenditure reports which compare actual expenditures to budgeted expenditures
are periodically submitted to the budget authority, e.g., finance, sponsor.
Expenditure audits are conducted to verify that charges to the project are
legitimate and that the work charged was actually performed. In most large
organizations with multiple projects in work at any given time it is possible to
find projects being charged for work done on other projects, for work not yet
done, etc. While these charges are often inadvertent, they must still be identified.
Variance reporting compares actual expenditures directly to budgeted expenditures.
The term ‘‘variance’’ is used here in the accounting sense, not the statistical
sense. In accounting, a variance is simply a comparison of a planned
amount with an actual amount. An accounting variance may or may not indicate
a special cause of variation; statistical techniques are required to make this
determination. The timing of variance reporting varies depending on the need
for control. The timing of variance reports should be determined in advance
and written into the project plan.
Variance tables: Variance reports can appear in a variety of formats. Most
common are simple tables that show the actual/budgeted/variances by budget
item, overall for the current period, and cumulatively for the project. Since it is
unlikely that variances will be zero, an allowance is usually made, e.g., 5% over
or under is allowed without the need for explanations. For longer projects,
historical data can be plotted on control charts and used to set allowances.
Variance graphs:When only tables are used it is difficult to spot patterns. To
remedy this tables are often supplemented with graphs. Graphs generally show
the budget variances in a time-ordered sequence on a line chart. The allowance
lines can be drawn on the graph to provide a visual guide to the eye.
The project manager should review the variance data for patterns which
contain useful information. Ideally, the pattern will be a mixture of positive
and negative but minor variances. Assuming that this pattern is accompanied
by an on-schedule project, this indicates a reasonably good budget, i.e., an
accurate forecasting of expenditures. Variances should be evaluated separately
for each type of budget (direct labor, materials, etc.). However, the variance
report for the entire project is the primary source of information concerning
the status of the project in terms of resource utilization. Reports are received
and analyzed periodically. For most quality improvement projects, monthly
or weekly reports are adequate. Budget variance analysis* should include the
Trends: Occasional departures from budget are to be expected. Of greater
concern is a pattern that indicates a fundamental problem with the budget.
Trends are easier to detect from graphic reports.
Performance measures 559
*This is not to be confused with the statistical technique Analysis of Variance (ANOVA).
Overspending: Since budgeted resources are generally scarce, overspending
represents a serious threat to the project and, perhaps, to the organization itself.
When a project overspends its budget, it depletes the resources available for
other activities and projects. The project team and team leader and sponsors
should design monitoring systems to detect and correct overspending before it
threatens the project or the organization. Overspending is often a symptom of
other problems with the project, e.g., paying extra in an attempt to ‘‘catch up’’
after falling behind schedule, additional expenses for rework, etc.
Underspending is potentially as serious as overspending. If the project budget
was prepared properly then the expenses reflect a given schedule and quality
level. Underspending may reflect ‘‘cutting corners’’ or allowing suppliers an
allowance for slower delivery. The reasons for any significant departure from
the plan should be explained.
Management support and organizational
Most organizations still have a hierarchical, command-and-control organizational
structure, sometimes called ‘‘smoke stacks’’ or ‘‘silos.’’ The functional
specialists in charge of each smoke stack tend to focus on optimizing their own
functional area, often to the detriment of the organization as a whole. In addition,
the hierarchy gives these managers a monopoly on the authority to act on
matters related to their functional specialty. The combined effect is both a
desire to resist change and the authority to resist change, which often creates
insurmountable roadblocks to quality improvement projects.
It is important to realize that organizational rules are, by their nature, a barrier
to change. The formal rules take the form of written standard operating procedures
(SOPs). The very purpose of SOPs is to standardize behavior. The
quality profession has (in this author’s opinion) historically overemphasized
formal documentation, and it continues to do so by advocating such approaches
as ISO 9000 and ISO 14000. Formal rules are often responses to past problems
and they often continue to exist long after the reason for their existence has
passed. In an organization that is serious about its written rules even senior
leaders find themselves helpless to act without submitting to a burdensome
rule-changing process. The true power in such an organization is the bureaucracy
that controls the procedures. If the organization falls into the trap of creating
written rules for too many things, it can find itself moribund in a fastchanging
external environment. This is a recipe for disaster.
Restrictive rules need not take the form of management limitations on itself,
procedures that define hourly work in great detail also produce barriers, e.g.,
union work rules. Projects almost always require that work be done differently
and such procedures prohibit such change. Organizations that tend to be excessive
in SOPs also tend to be heavy on work rules. The combination is often
deadly to quality improvement efforts.
Organization structures preserve the status quo in other ways besides formal,
written restrictions in the form of procedures and rules. Another effective
method of limiting change is to require permission from various departments,
committees, councils, boards, experts, etc. Even though the organization may
not have a formal requirement, that ‘‘permission’’ be obtained, the effect may
be the same, e.g., ‘‘You should run that past accounting’’ or ‘‘Ms. Reimer and
Mr. Evans should be informed about this project.’’When permission for vehicles
for change (e.g., project budgets, plan approvals) is required from a group that
meets infrequently it creates problems for project planners. Plans may be rushed
so they can be presented at the next meeting, lest the project be delayed for
months. Plans that need modificationsmaybe put on hold until the next meeting,
months away. Or, projects may miss the deadline and be put off indefinitely.
Modern organizations do not exist as islands. Powerful external forces take
an active interest in what happens within the organization. Government bodies
have created a labyrinth of rules and regulations that the organization
must negotiate to utilize its human resources without incurring penalties or
sanctions. The restrictions placed on modern businesses by outside regulators
are challenging to say the least. When research involves people, ethical and
legal concerns sometimes require that external approvals be obtained. The
approvals are contingent on such issues as informed consent, safety, cost and
so on.
Many industries have ‘‘dedicated’’ agencies to deal with. For example, the
pharmaceutical industry must deal with the Food and Drug Administration
(FDA). These agencies must often be consulted before undertaking projects.
For example, a new treatment protocol involving a new process for treatment
of pregnant women prior to labor may involve using a drug in a new way (e.g.,
administered on an outpatient basis instead of on an inpatient basis).
Many professionals face liability risks that are part of every decision. Often
these fears create a ‘‘play it safe’’ mentality that acts as a barrier to change. The
fear is even greater when the project involves new and untried practices and
Project management implementation 561
Perhaps the most significant change, and therefore the most difficult, is to
change ourselves. It seems to be a part of human nature to resist changing oneself.
By and large, we worked hard to get where we are, and our first impulse is
to resist anything that threatens our current position. Forsha (1992) provides
the process for personal change shown in Figure 15.9.
The adjustment path results in preservation of the status quo. The action
path results in change. The well-known PDCA cycle can be used once a commitment
to action has been made by the individual. The goal of such change is
continuous self-improvement.
Within an organizational context, the individual’s reference group plays a
part in personal resistance to change. A reference group is the aggregation of
people a person thinks of when they use the word ‘‘we.’’ If ‘‘we’’ refers to the
company, then the company is the individual’s reference group and he or she
Figure 15.9. The process of personal change.
From The Pursuit of Quality Through Personal Change, by H.I. Forsha. Copyright#1992
by ASQ Quality Press, Milwaukee, WI. Used by permission.
feels connected to the company’s success or failure. However, ‘‘we’’ might refer
to the individual’s profession or trade group, e.g., ‘‘We doctors,’’ ‘‘We engineers,’’
‘‘We union members.’’ In this case the leaders shown on the formal organization
chart will have little influence on the individual’s attitude towards the
success or failure of the project. When a project involves external reference
groups with competing agendas, the task of building buy-in and consensus is
daunting indeed.
Strategy #1: command people to act as you wishLWith this approach the
senior leadership simply commands people to act as the leaders wish. The implication
is that those who do not comply will be subjected to disciplinary action.
People in less senior levels of an organization often have an inflated view of the
value of raw power. The truth is that even senior leaders have limited power to
rule by decree. Human beings by their nature tend to act according to their
own best judgment. Thankfully, commanding that they do otherwise usually
has little effect. The result of invoking authority is that the decision-maker
must constantly try to divine what the leader wants them to do in a particular
situation. This leads to stagnation and confusion as everyone waits on the
leader. Another problem with commanding as a form of ‘‘leadership’’ is the
simple communication problem. Under the best of circumstances people will
often simply misinterpret the leadership’s commands.
Strategy #2: change the rules by decreeLWhen rules are changed by decree
the result is again confusion. What are the rules today? What will they be tomorrow?
This leads again to stagnation because people don’t have the ability to
plan for the future. Although rules make it difficult to change, they also provide
stability and structure that may serve some useful purpose. Arbitrarily changing
the rules based on force (which is what ‘‘authority’’ comes down to) instead of
a set of guiding principles does more harm than good.
Strategy #3: authorize circumventing of the rulesLHere the rules are
allowed to stand, but exceptions are made for the leader’s ‘‘pet projects.’’ The
result is general disrespect for and disregard of the rules, and resentment of the
people who are allowed to violate rules that bind everyone else. An improvement
is to develop a formal method for circumventing the rules, e.g., deviation
request procedures. While this is less arbitrary, it adds another layer of complexity
and still doesn’t change the rules that are making change difficult in the
first place.
Strategy #4: redirect resources to the projectLLeaders may also use their
command authority to redirect resources to the project. A better way is to develop
a fair and easily understood system to assure that projects of strategic impor-
Project management implementation 563
tance are adequately funded as a matter of policy. In our earlier discussion of
project scheduling we discussed ‘‘crash scheduling’’ as a means of completing
projects in a shorter time frame. However, the assumption was that the basis
for the allocation was cost or some other objective measure of the organization’s
best interest. Here we are talking about political clout as the basis of
the allocation.
Strategy #1: transform the formal organization and the organization’s
cultureLBy far the best solution to the problems posed by organizational roadblock
is to transform the organization to one where these roadblocks no longer
exist. As discussed earlier, this process can’t be implemented by decree. As the
leader helps project teams succeed, he will learn about the need for transformation.
Using his persuasive powers the leader-champion can undertake the exciting
challenge of creating a culture that embraces change instead of fighting it.
Strategy #2: mentoringLIn Greek mythology, Mentor was an elderly man,
the trusted counselor of Odysseus, and the guardian and teacher of his son
Telemachus. Today the term, ‘‘mentor’’ is still used to describe a wise and
trusted counselor or teacher. When this person occupies an important position
in the organization’s hierarchy, he or she can be a powerful force for eliminating
roadblocks. Modern organizations are complex and confusing. It is often difficult
to determine just where one must go to solve a problem or obtain a needed
resource. The mentor can help guide the project manager through this maze by
clarifying lines of authority. At the same time, the mentor’s senior position
enables him to see the implications of complexity and to work to eliminate
unnecessary rules and procedures.
Strategy #3: identify informal leaders and enlist their supportLBecause of
their experience, mentors often know that the person whose support the project
really needs is not the one occupying the relevant box on the organization
chart. The mentor can direct the project leader to the person whose opinion
really has influence. For example, a project may need the approval of, say, the
vice-president of engineering. The engineering VP may be balking because his
senior metallurgist hasn’t endorsed the project.
Strategy #4: find legitimate ways around people, procedures, resource constraints
and other roadblocksLIt may be possible to get approvals or resources
through means not known to the project manager. Perhaps a minor change in
the project plan can bypass a cumbersome procedure entirely. For example,
adding an engineer to the team might automatically place the authority to
approve process experiments within the team rather than in the hands of the
engineering department.
Short-term (tactical) plans
Conceptually, project plans are subsets of bigger plans, all of which are
designed to carry out the organization’s mission. The project plan must be
broken down further. The objective is to reach a level where projects are
‘‘tiny.’’ A tiny project is reached when it is possible to easily answer two questions:
1. Is the project complete?
2. Is the project done correctly?
For example, a software development team concluded that a tiny computer
module had the following characteristics: 1) it implemented only one concept;
2) it could be described in 6 lines or less of easily understood pseudo-code
(English like descriptions of what the program would do); and 3) the programming
would fit on a single sheet of paper. By looking at the completed programming
for the module, the team felt that it could answer the two questions.
On Six Sigma projects, tactical plans are created by developing work breakdown
structures. The process of creating work breakdown structures was discussed
above. Tactical planning takes place at the bottom-most level of the
work breakdown structures. If the project team doesn’t agree that the bottom
level is tiny, then additional work breakdown must take place.
Creating WBS employs the tree diagram technique. Tree diagrams are
described in Chapter 8. Tree diagrams are used to break down or stratify ideas
in progressively greater detail. The objective is to partition a big idea or problem
into its smaller components. By doing this, you will make the idea easier to
understand, or the problem easier to solve. The basic idea behind this is that, at
some level, a problem’s solution becomes relatively easy to find. This is the
tiny level. Work takes place on the smallest elements in the tree diagram.
Tactical plans are still project plans, albeit for tiny projects. As such, they
should include all of the elements of any well-designed project plan.
Contingency plans should be prepared to deal with unexpected but potentially
damaging events. The process decision program chart (PDPC) is a useful
tool for identifying possible events that might be encountered during the project.
The emphasis of PDPC is the impact of the ‘‘failures’’ (problems) on project
schedules. Also, PDPC seeks to describe specific actions to be taken to
prevent the problems from occurring in the first place, and to mitigate the
impact of the problems if they do occur. An enhancement to classical PDPC is
to assign subjective probabilities to the various problems and to use these to
help assign priorities. The amount of detail that should go into contingency
plans is a judgment call. The project manager should consider both the seriousness
of the potential problem and the likelihood of its occurring. See
Chapter 8 for additional information on PDPC.
Project management implementation 565
Cross-functional collaboration
This section will address the impact of organizational structures on management
of Six Sigma projects.
Six Sigma projects are process-oriented and most processes that have significant
impact on quality cut across several different departments. Modern organizations,
however, are hierarchical, i.e., they are defined by superior/
subordinate relationships. These organizations tend to focus on specialized
functions (e.g., accounting, engineering). But adding value for the customer
requires that several different functions work together. The ideal solution is
the transformation of the organization into a structure designed to produce
value without the need for a hierarchical structure. However, until that is
accomplished, Six Sigma project managers will need to deal with the conflicts
inherent in doing cross-functional projects in a hierarchical environment.
Project managers ‘‘borrow’’ people from many departments for their projects,
which creates matrix organizational structures. The essential feature of a
matrix organization is that some people have two or more bosses or project
customers. These people effectively report to multiple bosses, e.g., the project
manager and their own boss. Ruskin and Estes refer to people with more than
one boss as multi-bossed individuals, and their bosses and customers as multiple
bosses. Somewhere in the organization is a common boss, who resolves conflicts
between multiple bosses when they are unable to do so on their own. Of course,
multiple bosses can prevent conflicts by cooperation and collaboration before
problems arise.
Often multi-bossed individuals are involved with several projects, further
complicating the situation. Collaborative planning between the multiple bosses
is necessary to determine how the time of multi-bossed individuals, and other
resources, will be shared. Figure 15.10 illustrates the simplest multi-bossing
structure where the multi-bossed individual has just two multiple bosses, and
the common boss is directly above the multiple bosses on the organizational
hierarchy. For additional discussion of more complex matrix structures see
Ruskin and Estes (1995, pp. 169^182).
The matrix structure raises a number of questions regarding project planning
and coordination. What should be planned? Who should organize planning
activities? Who should participate in the planning? These issues were addressed
earlier in this chapter, especially in the section entitled ‘‘Relevant stakeholders.’’
Good communication is helpful in preventing problems. Perhaps the most
important communication is frequent, informal updates of all interested parties
by the project manager. More formal status reports should also be specified in
the project plan and sent to people with an interest in the project. The project
manager should determine who gets what information, which is often tricky
due to the multi-boss status of the project manager. Some managers may not
want ‘‘too much’’ information about their department’s ‘‘business’’ shared
with their peers from other areas. Other managers may be offended if they
receive less information than everyone else. The project manager’s best diplomatic
skills may be required to find the right balance.
Status reports invariably indicate that the project plan is less than perfect.
The process by which the plans will be adjusted should be understood in
advance. The process should specify who will be permitted to make adjustments,
when the adjustments will be allowed and how much authority the
bosses and project manager have in making adjustments.
Negotiated agreements should be documented, while generating the minimum
possible amount of additional red tape and paperwork. The documentation
will save the project manager a great deal of time in resolving disputes
down the road regarding who agreed to what.
Continuous review and enhancement of quality
The project management system can be improved, just as any system can be
improved. The Six Sigma Black Belt’s arsenal of tools and principles offer the
means. Address -such issues as cycle time, supplier management, customer
service, etc., just as you would for any other critical management system.
Continuous improvement principles, tools, and techniques are described in
detail in Chapter 8.
Project management implementation 567
Figure 15.10. Multi-boss collaborative planning
Projects have customers, usually internal. The techniques described in
Chapter 3 can be used to evaluate the satisfaction of the customers whose
needs are being addressed by the project.
The records of the project provide the raw data for process improvement.
These records, combined with customer satisfaction analysis, tell management
where the project planning and implementation process can be
improved. The project manager should include recommendations for process
improvement in the final project report. Organizational management, in
particular common bosses, should aggregate the recommendations from
several projects to identify systemic problems. Project schedules and budgets
should be compared to actual results to evaluate the accuracy of these forecasts.
The results can be analyzed using the techniques described in Chapters
11^14. Where special causes of variation in the results are present, they should
be identified and corrected quickly. Common causes of variation indicate
systemic problems.
Documentation and procedures
Project records provide information that is useful both while the project is
underway, and afterwards. Project records serve three basic purposes:
. cost accounting requirements
. legal requirements
. learning
Project records should be organized and maintained as if they were part of a
single database, even if it isn’t possible to keep all of the records in a single
location. There should be one ‘‘official’’ copy of the documentation, and a
person designated to act as the caretaker of this information while the project
is active. Upon completion of the project, the documentation should be sent
to the organization’s archives. Large quality improvement projects are expensive,
time-consuming undertakings. The process involved is complex and
often confusing. However, a great deal can be learned from studying the
‘‘project process.’’ To do this requires that there be data. Archives of a number
of projects can be used to identify common problems and patterns between
the various projects. For example, project schedules may be consistently too
optimistic or too pessimistic.
The following records should be kept:
. statement of work
. plans and schedules for projects and subprojects
. correspondence (written and electronic)
. written agreements
. meeting minutes
. action items and responsibilities
. budgets and ?nancial reports
. cost-bene?t analyses
. status reports
. presentation materials
. documentation of changes made to plans and budgets
. procedures followed or developed
. notes of signi?cant lessons learned
It is good practice for the project team to have a final meeting to perform a
‘‘post mortem’’ of the project. The meeting should be conducted soon after the
project’s completion, while memories are still fresh. The meeting will cover
the lessons learned from conducting the project, and recommendations for
improving the process. The minutes from these meetings should be used to educate
project managers.
The author believes that former guidelines for record retention are now
outdated. In the past, record storage involved warehousing costs, insurance,
aging of the paper, protecting the records from damage, etc. Furthermore,
using the records required that one sift through endless boxes of disorganized
material thrown in haphazardly. Today it is possible to reduce mountains of
paper to electronic form. Low-cost software can automatically catalog the
information and provide the ability to search the entire database quickly and
easily. There seems to be little reason not to store complete project information
When people speak of ‘‘reports,’’ they usually mean formal written reports
or, perhaps, formal verbal presentations. These forms of communication have
certain advantages. They are relatively self-contained and complete and thus
useful to personnel not intimately involved with the project. Their form lends
itself to long-term retention. They usually include additional background
materials. Finally, formal written reports are usually designed to address the
concerns of all parties. However, formal written reports also have some
drawbacks. Their preparation takes considerable time and effort, which makes
them costly. Also, by trying to address everyone’s concern the formal written
report usually contains a lot of information that is of little interest to the
majority of the recipients. Of course, this latter drawback can be mitigated by
creating a good table of contents and index and by carefully organizing the
Project management implementation 569
Informal reports and correspondence are used to keep everyone up to date
on the project. Unlike formal reports, this form of communication generally
addresses a specific issue involving both the sender and the receiver. Because
the parties to the communication usually bring a great deal of background
information to the communication, informal reports tend to do poorly as
stand-alone documents. However, informal reports can be prepared quickly
and inexpensively and they tend to be short.
^ ^ ^
Risk Assessment
Reliability analysis
Safety and reliability are specialties in their own right. The Six Sigma analyst
is expected to have an understanding of certain key concepts in these subject
areas. It is obvious that these two areas overlap the Six Sigma body of knowledge
to a considerable extent. Some concept areas are nearly identical (e.g., traceability)
while others are merely complementary (e.g., reliability presumes conformance
to design criteria, which Six Sigma addresses directly). Modern ideas
concerning safety share a common theoretical base with reliability.
While common usage of the term reliability varies, its technical meaning is
quite clear: reliability is defined as the probability that a product or system
will perform a specified function for a specified time without failure. For
the reliability figure to be meaningful, the operating conditions must be carefully
and completely defined. Although reliability analysis can be applied to
just about any product or system, in practice it is normally applied only
to complex products. Formal reliability analysis is routinely used for both
commercial products, such as automobiles, as well as military products such
as missiles.
*Some of the material in this section first appeared in The Complete Guide to the CRE by Bryan Dodson, # Quality
Copyright 2003 by The McGraw-Hill Companies, Inc. Click Here for Terms of Use.
Reliability terms
MTBFLMean time between failures, . When applied to repairable products,
this is the average time a system will operate until the next failure.
Failure rateLThe number of failures per unit of stress. The stress can be
time (e.g., machine failures per shift), load cycles (e.g., wing fractures
per 100,000 deflections of six inches), impacts (e.g., ceramic cracks per
1,000 shocks of 5 g’s each), or a variety of other stresses. The failure
rate  ? 1=.
MTTF or MTFFLThe mean time to first failure. This is the measure
applied to systems that can’t be repaired during their mission. For
example, the MTBF would be irrelevant to the Voyager spacecraft.
MTTRLMean time to repair. The average elapsed time between a unit failing
and its being repaired and returned to service.
AvailabilityLThe proportion of time a systemis operable. Only relevant for
systems that can be repaired. Availability is given by the equation
Availability ?
MTBF+MTTR ?16:1?
b10 lifeLThe life value at which 10% of the population has failed.
b50 lifeLThe life value at which 50% of the population has failed. Also called
the median life.
Fault tree analysis (FTA)LFault trees are diagrams used to trace symptoms
to their root causes. Fault tree analysis is the term used to describe the
process involved in constructing a fault tree. (See below for additional
DeratingLAssigning a product to an application that is at a stress level less
than the rated stress level for the product. This is analogous to providing
a safety factor.
Censored testLA life test where some units are removed before the end of
the test period, even though they have not failed.
MaintainabilityLA measure of the ability to place a system that has failed
back in service. Figures of merit include availability and mean time to
*b10 life and b50 life are terms commonly applied to the reliability of ball bearings.
The reliability of a given system is dependent on the reliability of its individual
elements combined with how the elements are arranged. For example, a
set of Christmas tree lights might be configured so that the entire set will fail if
a single light goes out. Or it may be configured so that the failure of a single
light will not affect any of the other lights (question: do we define such a set as
having failed if only one light goes out? If all but one go out? Or some number
in between?).
Mathematical models
The mathematics of reliability analysis is a subject unto itself. Most systems
of practical size require the use of high speed computers for reliability evaluation.
However, an introduction to the simpler reliability models is extremely
helpful in understanding the concepts involved in reliability analysis.
One statistical distribution that is very useful in reliability analysis is the
exponential distribution, which is given by Equation 16.12.
R ? exp 
  ; t  0 ?16:2?
In Equation 16.2Ris the system reliability, given as a probability, t is the time the
system is required to operate without failure,  is the mean time to failure for
the system. The exponential distribution applies to systems operating in the constant
failure rate region, which is wheremost systems are designed to operate.
Reliability apportionment
Since reliability analysis is commonly applied to complex systems, it is
logical that most of these systems are composed of smaller subsystems.
Apportionment is the process involved in allocating reliability objectives
among separate elements of a system. The final system must meet the overall
reliability goal. Apportionment is something of a hybrid of project management
and engineering disciplines.
The process of apportionment can be simplified if we assume that the exponential
distribution is a valid model. This is because the exponential distribution
has a property that allows the system failure rate to be computed as the reciprocal
of the sum of the failure rates of the individual subsystems. Table 16.1 shows the
apportionment for a home entertainment center. The complete system is composed
of a tape deck, television, compact disk unit, and a phonograph. Assume
that the overall objective is a reliability of 95% at 500 hours of operation.
Reliability and safety analysis 573
The apportionment could continue even further; for example, we could
apportion the drive reliability to pulley, engagement head, belt, capstan, etc.
The process ends when it has reached a practical limit. The column labeled
‘‘objective’’ gives the minimum acceptable mean time between failures for
each subsystem in hours. MTBFs below this will cause the entire system to fail
its reliability objective. Note that the required MTBFs are huge compared to
the overall objective of 500 hours for the system as a whole. This happens partly
because of the fact that the reliability of the system as a whole is the product of
the subsystem reliabilities which requires the subsystems to have much higher
reliabilities than the complete system.
Reliability apportionment is very helpful in identifying design weaknesses. It
is also an eye opener for management, vendors, customers, and others to see
how the design of an entire system can be dramatically affected by one or two
unreliable elements.
Asystem is in series if all of the individual elements must function for the system
to function. A series system block diagram is shown in Figure 16.1.
Table 16.1. Reliability apportionment for a home entertainment system.
Tape deck 0.990 0.010 0.00002 49,750
Television 0.990 0.010 0.00002 49,750
Compact disk 0.985 0.015 0.00003 33,083
Phonograph 0.985 0.015 0.00003 33,083
0.950 0.050
Drive 0.993 0.007 0.000014 71,178
Electronics 0.997 0.003 0.000006 166,417
0.990 0.010
In Figure 16.1, the system is composed of two subsystems, A and B. Both A
and Bmust function correctly for the system to function correctly. The reliability
of this system is equal to the product of the reliabilities of A and B, in other
RS ? RA  RB ?16:3?
For example, if the reliability of A is 0.99 and the reliability of B is 0.92, then
RS ? 0:99  0:92 ? 0:9108. Note that with this configuration, RS is always
less than the minimum of RA or RB. This implies that the best way to improve
the reliability of the system is to work on the system component that has the
lowest reliability.
A parallel system block diagram is illustrated in Figure 16.2. This system will
function as long as A or B or C haven’t failed. The reliability of this type of configuration
is computed using Equation 16.4.
RS ? 1  ?1  RA??1  RB??1  RC? ?16:4?
For example, if RA ? 0:90, RB ? 0:95, and RC ? 0:93 then RS ? 1  ?0:1 0:05  0:07? ? 1  0:00035 ? 0:99965.
With parallel configurations, the system reliability is always better than the
best subsystem reliability. Thus, when trying to improve the reliability of a parallel
system you should first try to improve the reliability of the best component.
This is precisely opposite of the approach taken to improve the reliability of
series configurations.
Reliability and safety analysis 575
Figure 16.1. A series con?guration.
Seven steps in predicting design reliability
1. De?ne the product and its functional operation. Use functional block
diagrams to describe the systems. De?ne failure and success in unambiguous
2. Use reliability block diagrams to describe the relationships of the various
system elements (e.g., series, parallel, etc.).
3. Develop a reliability model of the system.
4. Collect part and subsystem reliability data. Some of the information may
be available from existing data sources. Special tests may be required to
acquire other information.
5. Adjust data to ?t the special operating conditions of your system. Use
care to assure that your ‘‘adjustments’’ have a scienti?c basis and are
not merely re?ections of personal opinions.
6. Predict reliability using mathematical models. Computer simulation
may also be required.
7. Verify your prediction with ?eld data Modify your models and predictions
Figure 16.2. A parallel system.
System e?ectiveness
The effectiveness of a system is a broader measure of performance than
simple reliability There are three elements involved in system effectiveness:
1. Availability.
2. Reliability.
3. Design capability, i.e., assuming the design functions, does it also achieve
the desired result?
System effectiveness can be measured with Equation 16.5.
PSEf ? PA  PR  PC ?16:5?
In this equation, PSEf is the probability the system will be effective, PA is the
availability as computed with Equation 16.1, PR is the system reliability, and
PC is the probability that the design will achieve its objective.
As seen in the previous sections, reliability modeling can be difficult mathematically.
And in many cases, it is impossible to mathematically model the
situation desired. Monte Carlo simulation is a useful tool under these and
many other circumstances, such as:
. Verifying analytical solutions
. Studying dynamic situations
. Gaining information about event sequences; often expected values and
moments do not provide enough detail
. Determining the important components and variables in a complex system
. Determining the interaction among variables
. Studying the e?ects of changes without the risk, cost, and time constraints
of experimenting on the real system
. Teaching
Random number generators
The heart of any simulation is the generation of random numbers. If a programming
language such as BASIC, C, or FORTRAN is used, random number
generators will have to be created. If simulation languages such as Siman,
Slam, Simscript, or GPSS are used, random number generators are part of the
Random numbers from specific distributions are generated by transforming
random numbers from the unit, uniform distribution. Virtually all pro-
Reliability and safety analysis 577
gramming languages, as well as electronic spreadsheets, include a unit, uniform
random number generator. Technically, these unit, uniform random number
generators are pseudo-random number generators, as the algorithms used to
generate them take away a small portion of the randomness. Nevertheless,
these algorithms are extremely efficient and for all practical purposes the result
is a set of truly random numbers.
A simple way to generate distribution-specific random numbers is to set the
cumulative distribution function equal to a unit, random number and take the
inverse. Consider the exponential distribution
F?x? ? 1  elx ?16:6?
By setting r, a random variable uniformly distributed from zero to one, equal to
F?x? and inverting the function, an exponentially distributed random variable,
x, with a failure rate of l is created.
r ? 1  elx
1  r ? elx
ln ?1  r? ? lx
x ? 
ln ?1  r?
This expression can be further reduced; the term 1  r is also uniformly distributed
from zero to one. The result is
x ? 
ln r
l ?16:8?
Table 16.2 contains some common random number generators.
Simulation modeling
After the desired random number generator(s) have been constructed, the
next step is to mathematically model the situation under study. After completing
the model, it is important to validate and verify the model. A valid
*A unit, uniform random variable can be generated using the ‘‘RND’’ function in the BASIC programming language (Visual
Basic, Quick Basic, GWBasic, and BASICA), and the ‘‘@RAND’’ function in LotusTM 123. MicrosoftTM Excel includes a
random generating tool that allows random numbers to be generated from several distributions.
Reliability and safety analysis 579
Table 16.2 Random number generators.
Uniform f ?x? ?
b  a
; a  x  b x ? a ? ?b  a?r
Exponential f ?x? ? ex; 0 < x < 1 x ? 

ln r
Normal f ?x? ?
 ffiffiffiffiffi 2 p exp 
2  ;
1< x < 1
x1 ? ? ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi 2 ln r1 p cos ?2r2?
x2 ? ? ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi 2 ln r1 p sin ?2r2?
 ?  y
f ?x? ?
x ffiffiffiffiffi 2 p exp 
ln x  
  2  ;
x > 0
x1 ? exp ? ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi 2 ln r1 p cos ?2r2?
x2 ? ? ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi 2 ln r1 p sin?2r2?
 ?  y
f ?x? ?



; x > 0 x ?  ? ln r?1=

f ?x? ?
; x ? 0; 1; 2; . . . ;1 x ?

ln r > t z

ln ri < t 
><>>: Chi-square f ?x? ?
2v=2	?v=2?
x?v=21?ex=2; x > 0 x ?Xv
zi is a standard normal
random deviate.
Beta f ?x? ?
B?p; q?
xp1?1  x?q1;
0  x  1; p > 0; q > 0
x ?
r1=p ? r1=q
Gamma f ?x? ?
x ? 0;  ? 0;  ? 0
1.  is a non-integer shape parameter.
2. Let 1 ?the truncated integer root
of .
3. Let q ? lnQ
4. Let A ?   1 and B ? 1  A.
5. Generate a random number and let
y1 ? r1=A
j .
6. Generate a random number and let
y2 ? r1=B
7. If y1 ? y2  1 go to 9.
8. Let i ? i ? 2 and go to 5.
9. Let z ? y1=?y1 ? y2?.
10. Generate a random number, rn.
11. Let W ? ln rn.
12. x ? ?q ? zW?.
Continued on next page . . .
model is a reasonable representation of the situation being studied A model is
verified by determining that the mathematical and computer model created
represents the intended conceptual model.
Enough iterations should be included in the simulation to provide a steadystate
solution, which is reached when the output of the simulation from one
iteration to the next changes negligibly.When calculating means and variances,
1,000 iterations is usually sufficient. If calculating confidence limits, many
more iterations are required; after all, for 99% confidence limits the sample
size for the number of random deviates exceeding the confidence limit is
1/100th the number of iterations.
Simulation can be used to determine the result of transformations of random
variables. Suppose the mean and variance of the product of two normally
distributed variables are needed for design purposes. The following BASIC
code will produce the desired result. A flow chart for this code is shown in
Figure 16.3.
Binomial p?x? ? n=x ? ?px?1  p?nx; x ? 0; 1; . . . ; n x ?Xn
yi; yi ?
0; ri > p
1; ri  p 
Geometric p?x? ? p?1  p?x1; x ? 1; 2; 3; . . .
ln ?1  r?
ln ?1  p?  x 
ln ?1  r?
ln ?1  p? ? 1 z
Student’s t
f ?x? ?
	??v ? 1?=2

	?v=2? ffiffiffiffiffiffi v p 1 ?
v  ?v?1?=2
1< x < 1
x ?
1=2 zi is a standard
normal random
f ?x? ?
	??v1 ? v2?=2
	?v1=2?	?v2=2?  

?1 ? v1x=v2??v1?v2?=2  ; x > 0:
x ?
v1 P
zi is a standard
normal random
Table 16.29Continued
yTwo uniform random numbers must be generated, with the result being two normally distributed random numbers.
zIncrease the value of x until the inequality is satisfied.
§Statistical Software, such as MINITAB, have these functions built-in.
REM simulation for the product of two normally
distributed variables
FOR i = 1 TO 5000
a = RND
b = RND
REM x is normal with mean=e1 and standard deviation=v1
x = (((–2  LOG(a)) ^ .5)  COS(2  3.141592654#  b))  v1 + e1 REM y
is normal with mean=e2 and standard deviation=v2
Reliability and safety analysis 581
Figure 16.3. Flow chart for the simulation of two normally distributed normal variables.
y = (((–2  LOG(a)) ^ .5)  SIN(2  3.141592654#  b))  v2 + e2
z = x  y
ztot# = ztot# + z
zsq# = zsq# + z ^ 2
PRINT ‘‘ztot zsq’’; ztot#; zsq#
PRINT ‘‘mean=’’; ztot# / 5000
zvar# = (5000  zsq# – ztot# ^ 2) / (5000  4999)
PRINT ‘‘variance=’’; zvar#
Note: ‘‘RND’’ generates uniform random deviates on the interval from zero
to one. The ‘‘LOG’’ function in BASIC represents the natural logarithm.
In the above code, two normal random numbers are generated with the
desired parameters and multiplied. This is repeated 5,000 times, and the mean
and variance of these 5,000 random deviates are calculated. The result is a random
variable with a mean of 12,009 and a variance of 3,307,380.
For the above example, recall that the same result could have been obtained
mathematically. A disadvantage of solving this problem mathematically is that
there is no information regarding the shape of the resulting distribution.
With electronic spreadsheets, simulations no longer require computer code.
The previous example is simulated using Lotus 123TM with the following steps.
1. In cell A1 place the function @RAND
2. In cell A2 place the function @RAND
3. In cell A3 place the formula
(((2@ln(A1))^ .5)@cos(2@piA2))7+100
4. In cell A4 place the formula
5. In cell A5 place the formula
6. Copy the contents of row A 5,000 times
In the above example, each row in the spreadsheet represents an iteration.
The powerful @ functions and graphics tools contained in the spreadsheet
can then be used for analysis. Note that each change in the spreadsheet causes
the random numbers to be changed. It may be helpful to convert the output
from formulas to fixed numbers with the ‘‘Range Value’’ command.
Now consider a system consisting of four identical components which are
exponentially distributed with a failure rate of 0.8. Three of the components
*MicrosoftTM Excel can generate normal random variables directly.
are standby redundant with perfect switching. Information is needed regarding
the shape of the resulting distribution. The following code produces four exponentially
distributed random variables with a failure rate of 0.8, adds them, and
writes the result to the file ‘‘c:\data’’; this process is repeated 5,000 times. A
flow chart for this problem is provided in Figure 16.4.
REM simulation for the sum of four exponentially
distributed variables
OPEN ‘‘c:\data’’ FOR OUTPUT AS #1 LEN = 256
FOR i = 1 TO 5000
REM x1 is exponential with failure rate = 0.8
x1 = -(1 /.8)  LOG(RND)
Reliability and safety analysis 583
Figure 16.4. Simulation for the sum of exponential random variables.
REM x2 is exponential with failure rate = 0.8
x2 = -(1 /.8)  LOG(RND)
REM x3 is exponential with failure rate = 0.8
x3 = -(1 /.8)  LOG(RND)
REM x4 is exponential with failure rate = 0.8
x4 = -(1 /.8)  LOG(RND)
y = x1 + x2 + x3 + x4
PRINT #1, USING ‘‘##########.#####’’; y
By importing the resulting data into an electronic spreadsheet or statistical
program, a wide variety of analyses can be done on the data. A histogram of
the data produced from the above code is shown in Figure 16.5.
As seen from Figure 16.5, the sum of n exponentially distributed random variables
with a failure rate of l is a random variable that follows the gamma distribution
with parameters  ? n and l.
This problem is also easily simulated using an electronic spreadsheet. The
steps required follow:
Figure 16.5. Histogram of the sum of four exponentially distributed random variables.
1. In cells A1 through A4, place the formula
?1; 0:8?@ln(@rand)
2. In cell A5 place the formula
@sum(A1. . .A4)
3. Copy the contents of row A 5,000 times
Again, each row represents an iteration, and the spreadsheet can be used to
obtain the desired simulation output.
Nowconsider a system consisting of three components in series. The components
are Weibully distributed with parameters 
 ? 2,  ? 300; 
 ? 4,
 ? 100; and, 
 ? 0:5,  ? 200. The code below depicts this situation. Figure
16.6 shows a flow chart for this simulation.
Reliability and safety analysis 585
Figure 16.6. Simulation of a series system.
REM simulation three Weibully distributed variables in series
DIM x(99)
OPEN ‘‘c:\data’’ FOR OUTPUT AS #1 LEN = 256
FOR i = 1 TO 5000
REM x(1) is Weibull shape parameter=2 and scale parameter=300
x(1) = 300  (–LOG(RND)) ^ (1 / 2)
REM x(2) is Weibull shape parameter=4 and scale parameter=100
x(2) = 100  (–LOG(RND)) ^ (1 / 4)
REM x(3) is Weibull shape parameter=0.5 and scale parameter=200
x(3) = 200  (–LOG(RND)) ^ (1 / .5)
min = 999999999
FOR j = 1 TO 3
IF x(j) < min THEN min = x(j)
y = min
PRINT #1, USING ‘‘##########.#####’’; y
For a series system, the time to fail is the minimum of the times to fail of the
components in series. A parallel system could be modeled by altering the
above code to take the maximum time to fail of each of the components.
Figure 16.7 is a histogram of the resulting data for the series system.
Figure 16.7. Histogram of a series system.
The large number of early failures are caused by the component with the
high infant mortality rate (
 ? 0:5). The result is a distribution that does
not appear to conform to any known distributions. However, with 5,000
points, a reliability function can be built empirically. The result is shown in
Figure 16.8.
The following steps are used to simulate the above problem using an electronic
1. In cell A1 place the formula
300(^@ln(@rand))^(1 /2)
2. In cell A2 place the formula
100(^@ln(@rand))^(1 /4)
3. In cell A3 place the formula
4. In cell A4 place the formula
@min(A1 . . A4)
5. Copy the contents of row A 5,000 times.
Now consider a system with two Weibully distributed components, A and B.
Component B is standby redundant, and the switching mechanism has a reliability
of 95%. The parameters of component A are 
 ? 3,  ? 85. The parameters
of component B are 
 ? 4:4,  ? 97. The code below models this
system; a flow chart is given in Figure 16.9.
Reliability and safety analysis 587
Figure 16.8. Reliability function for a series system.
REM simulation of a switch for two Weibully distributed variables
OPEN ‘‘c:\data’’ FOR OUTPUT AS #1 LEN = 256
FOR i =1 TO 5000
REM x is Weibull shape parameter=3 and scale parameter=85
x = 85  (-LOG(RND))^(1/3)
REM y is Weibull shape parameter=4.4 and scale parameter=97
y = 97  (-LOG(RND))^(1/4.4)
IF s>=.05 THEN swr=1 ELSE swr=0
IF swr=1 THEN z=x+y ELSE z=x
Figure 16.9. Simulation of a switching system.
PRINT #1, USING ‘‘##########.#####’’; y
Ahistogram of the 5,000 data points written to the data file is shown in Figure
The histogram shows the time to fail for the system following a Weibull distribution.
The reliability function for this system, built from the 5,000 simulation
points, is shown in Figure 16.11.
This situation can be simulated using an electronic spreadsheet. The required
steps follow:
1. In cell A1 place the formula
2. In cell A2 place the formula
97(@ln(@rand))^(1 /4.5)
3. In cell A3 place the formula
4. Copy the contents of row A 5,000 times.
In step 3 above, the reliability of the switch is tested using the unit, uniform
random number generator. If the unit, uniform random number is less than
0.05, the switch fails, and the time to fail for the system is the time to fail for com-
Reliability and safety analysis 589
Figure 16.10. Histogram of a switching system.
ponent A (the value in cell A1). If the switch operates, the system time to fail is
the sum of the values in cells A1 and A2.
In summary, simulation is a powerful analytical tool that can be used to
model virtually any system. For the above examples, 5,000 iterations were
used. The number of iterations used should be based on reaching a steady-state
condition. Depending on the problem more or less iterations may be needed.
Once simulation is mastered, a danger is that it is overused because of the
difficulty involved with mathematical models. Do not be tempted to use
simulation before exploring other options. When manipulating models, mathematical
models lend themselves to optimization whereas simulation models
require trial and error for optimization.
Risk assessment tools
While reliability prediction is a valuable activity, it is even more important to
design reliable systems in the first place. Proposed designs must be evaluated
to detect potential failures prior to building the system. Some failures are more
important than others, and the assessment should highlight those failures most
deserving of attention and scarce resources. Once failures have been identified
and prioritized, a system can be designed that is robust, i.e., it is insensitive to
most conditions that might lead to problems.
Figure 16.11. Reliability function for a switching system.
Design reviews are conducted by specialists, usually working on teams.
Designs are, of course, reviewed on an ongoing basis as part of the routine work
of a great number of people. However, the term as used here refers to the formal
design review process. The purposes of formal design review are threefold:
1. Determine if the product will actually work as desired and meet the
customer’s requirements.
2. Determine if the new design is producible and inspectable.
3. Determine if the new design is maintainable and repairable.
Design review should be conducted at various points in the design and production
process. Review should take place on preliminary design sketches,
after prototypes have been designed, and after prototypes have been built and
tested, as developmental designs are released, etc. Designs subject to review
should include parts, sub-assemblies, and assemblies.
While FMEA (see below) is a bottom-up approach to reliability analysis,
FTA is a top-down approach. FTA provides a graphical representation of the
events that might lead to failure. Some of the symbols used in construction of
fault trees are shown in Table 16.3.
In general, FTA follows these steps:
1. De?ne the top event, sometimes called the primary event. This is the
failure condition under study.
2. Establish the boundaries of the FTA.
3. Examine the system to understand how the various elements relate to
one another and to the top event.
4. Construct the fault tree, starting at the top event and working downward.
5. Analyze the fault tree to identify ways of eliminating events that lead to
6. Prepare a corrective action plan for preventing failures and a contingency
plan for dealing with failures when they occur.
7. Implement the plans.
8. Return to step #1 for the new design.
Figure 16.12 illustrates an FTA for an electric motor.
Safety analysis
Safety and reliability are closely related. A safety problem is created when a
critical failure occurs, which reliability theory addresses explicitly with such
Reliability and safety analysis 591
Table 16.3. Fault-tree symbols.
Source: Handbookof Reliability Engineering and Management, McGraw-Hill, reprinted
with permission of the publisher.
AND gate Output event occurs if all the input events
occur simultaneously
OR gate Output event occurs if any one of the
input events occurs
Inhibit gate Input produces output when conditional
event occurs
Priority AND gate Output event occurs if all input events
occur in the order from left to right
Exclusive OR gate Output event occurs if one, but not both,
of the input events occur
m-out-of-n gate
(voting or sample
Output event occurs if m-out-of-n input
events occur
Event represented by a gate
Basic event with su?cient data
Undeveloped event
Either occurring or not occurring
Conditional event used with inhibit gate
Transfer symbol
tools as FMEA and FTA. The modern evaluation of safety/reliability takes into
account the probabilistic nature of failures. With the traditional approach a
safety factor would be defined using Equation 16.9.
SF ?
average strength
worst expected stress ?16:9?
The problem with this approach is quite simple: it doesn’t account for variation
in either stress or strength. The fact of the matter is that both strength
Reliability and safety analysis 593
Figure 16.12. Fault tree for an electric motor.
Source: Handbookof Reliability Engineering and Management, McGraw-Hill,
reprinted with permission of the publisher.
and stress will vary over time, and unless this variation is dealt with explicitly we
have no idea what the ‘‘safety factor’’ really is. The modern view is that a safety
factor is the difference between an improbably high stress (the maximum
expected stress, or ‘‘reliability boundary’’) and an improbably low strength
(the minimum expected strength). Figure 16.13 illustrates the modern view of
safety factors. The figure shows two distributions, one for stress and one for
Since any strength or stress is theoretically possible, the traditional concept
of a safety factor becomes vague at best and misleading at worst. To deal intelligently
with this situation, we must consider probabilities instead of possibilities.
This is done by computing the probability that a stress/strength
combination will occur such that the stress applied exceeds the strength. It is
possible to do this since, if we have distributions of stress and strength, then
the difference between the two distributions is also a distribution. In particular,
if the distributions of stress and strength are normal, the distribution of the difference
between stress and strength will also be normal. The average and standard
distribution of the difference can be determined using statistical theory,
and are shown in Equations 16.10 and 16.11.
SF ? 2
STRESS ?16:10?
In Equations 16.10 and 16.11 the SF subscript refers to the safety factor.
Figure 16.13. Modern view of safety factors.
Assume that we have normally distributed strength and stress. Then the distribution
of strength minus stress is also normally distributed with the mean
and variance computed from Equations 16.10 and 16.11. Furthermore, the probability
of a failure is the same as the probability that the difference of strength
minus stress is less than zero. That is, a negative difference implies that stress
exceeds strength, thus leading to a critical failure.
Assume that the strength of a steel rod is normally distributed with
 ? 50,000# and  ? 5,000#. The steel rod is to be used as an undertruss on a
conveyor system. The stress observed in the actual application was measured
by strain gages and it was found to have a normal distribution with
 ? 30,000# and  ? 3,000#.What is the expected reliability of this system?
The mean variance and standard deviation of the difference is first computed
using Equations 16.10 and 16.11, giving
STRESS ? 5,0002 ? 3,0002 ? 34,000,000
 ? ffiffiffiffiffiffi