This action might not be possible to undo. Are you sure you want to continue?
Statistical Methods in Engineering and Quality Assurance
Statistical Methods in Engineering and Quality Assurance
PETER W. M. JOHN
A WileyInterscience Publication JOHN WILEY 8r SONS, INC. New York 0 Chichester 0 Brisbanc
0
Toronto
0
Singapore
In rccognition of the importance of preserving what has been written, it is a policy of John Wiley & Sons, Inc. to have books o f cnduring valuc publishcd in thc United States printed on acidfree paper, and wc excrt our best eKorts to that end. Copyright
0
1990 by John Wiley & Sons, Inc.
All rights rcserved. Published simultaneously in Canada.
Reproduction or translation o f any part o f this work beyond that perniitted by Section 107 or 108 of the 1976 United States Copyright Act without the permission of thc copyright owncr is unlawful. Requests for permission or furthcr inforniation should be addressed to the Permissions Departmcnt, John Wilcy XC Sons, lnc.
Library o Congress CalaloginginPublicativnData f
John, Peter William Mcrcdith. Statistical methods in engineering and quality assurance I Peter W. M. John. cm.   (Wiley scries in probability and mathcmatical p: statistics. Applied probability and statistics section) locludes bibliographical references (p. ). ISBN 0471829862 1. Engineering Statistical methods. 2. Quality controlStatistical methods. 3. Quality assurance. I. Title. 11. Serics. 1') 9M TA340.JV 620'.00724 ~ 2 0 003371X CIP
10 Y 8 7 6 5 4 3
Contents
xv
1
Preface
1.
Quality Assurance and Statistics 1.1 1.2 1.3 1.4 1.s 1.6 1.7 1.8 1.9 1.10 1.11 Quality and American Business, 1 Competition, 2 The Reaction, 2 Total Quality, 3 Kaizen , 4 Why Statistics?. 4 The Future, 5 Gathering Information, 6 This Book, 7 Instructional Considerations, 7 Keprisc, 8
2. Descriptive Statistics
9
2. I 2.2 2.3 2.4 2.5
Introduction, 9 The Mean and Standard Deviation, 9 The Median and Quartiles, 11 Ishikawa's Seven Tools, 13 Histograms, 14 2.6 An Example of a Nonsymmetrical Distribution, 16 2.7 I)otplots, I 9 2.8 StemandLeaf Diagrams, 20 2.9 Splitting Stems, 21 2.10 BoxandWhisker Plots, 22 2.11 Pareto Diagrams, 24
V
vi
CONTENTS
2.12 Summary, 25 Exercises, 26
3.
Discrete Variables, or Attributes
29
3.1 3.2 3.3 3.4 3.5 3.6 3.7 3.8 3.0 3.10 3.1 1 3.12 3.13 3.14 3.15 3.16 3.17 3.18 3.19 3.20 3.21
Introduction, 29 Random Digits, 29 A Mathematical Model, 30 Somc Basic Probability Theory, 31 The Coin Tossing Model, 33 The Binomial Distribution, 33 Unequal Probabilities, 35 The Binomial Distribution (General Case), 36 Mathematical Expectation, 37 The Variance, 38 Sampling without Replacement, 39 The Hypergeometric Distribution, 40 Conditional Probabilities, 41 Bayes' Theorem, 42 Bivariate Distributions, 44 Thc Geometric Distribution, 45 The Poisson Distribution, 46 Properties of the Poisson Distribution, 47 The Bortkiewicz Data, 48 Other Applications, 48 Further Properties, 48 Exercises, 49
52
4. Continuous Variables
4.1 4.2 4.3 4.4 4.5 4.6 4.7 4.8 4.9 4.10
Introduction. 52 Probability Density Functions, 52 The Cumulativc Distribution Function, 54 The Uniform Distribution, 55 The Exponential Distribution, 56 The Quartiles of a Continuous Distribution, 58 The Memorylcss Property of the Exponential Distribution, 58 The Reliability Function and the Hazard Rate, 59 The Wcibull Distribution, 60 The Gamma Distribution, 61
CONTENTS
vii
4.11 4.12 4.13 4.14 4.15 4.16
The Expectation of a Continuous Random Variable, 62 The Variance, 64 MomentGenerating Functions, 64 Bivariate Distributions, 66 Covariance, 67 Independent Variables, 68 Exercises, 69
72
5. The Normal Distribution
5.1 5.2 5.3 5.4 5.5 5.6 5.7 5.8 5.9 5.10 5.11 5.12 5.13
Introduction. 72 The Normal Density Function, 72 The Tables of the Normal Distribution, 74 The Moments of the Normal Distribution, 75 The MomentGenerating Function, 76 Skewness and Kurtosis, 76 Sums of Independent Random Variables, 77 Random Samples, 79 Linear Combinations of Normal Variables, 80 The Central Limit Theorem, 80 The Chisquare Distribution, 81 The Lognormal Distribution, 82 The Cumulative Normal Distribution Curve and Normal Scores, 83 Exercises, 86
88
6. Some Standard Statistical Procedures 6.1 6.2 6.3 6.4 6.5 6.6 6.7 6.8 6.9 6.10 6.11 6.12 Introduction, 88 The Three Experiments, 88 The First Experiment, 89 The Second Experiment, 89 The Third Experiment, 90 The Behavior of the Sample Average, 90 Confidence Intervals for the Normal Distribution, 90 Confidence Intervals for the Binomial Distribution, 94 Confidence Intervals for the Exponential Distribution, 96 Estimating the Variance, 97 Student’s t Statistic, 100 Confidcnce Intervals for the Variance of a Normal Population, 101
139 Comparing Binomial Populations.6 8.8 8.5 8. 108 Exercises. 133 Wilcoxon's TwoSample Test. 127 7.4 Introduction. 118 T i e tTest.6 7.12 ChiSquare and 2 X 2 Tables.10 7.1 8. 138 Comparing Exponential Distributions. 124 Tests for the Percent Defective.1 1 . 102 Another Method of Estimating the Standard Deviation. 110 113 Hypothesis Testing 7. Comparative Experiments 129 Introduction.13 7.13 6. 117 The Choice of Alpha.17 7. 141 Exercises.8 7.1 7. 116 Acceptance and Rejection Regions. 106 MaximumLikelihood Estimates.1 1 7. 105 Moment Estimators. 120 The rTest. 126 Exercises.16 6.3 8. 116 OncSided 'rests. CONI'ENTS Chisquare and Exponential Variables. 125 Multinomial Goodness o f Fit.2 8. 119 The Type 11 Error. 113 An Example. 114 The Decisions.2 7.4 8.14 6.9 8. 115 Null and Alternate Hypotheses. 123 Thc Exponential Distribution. 131 Unequal Variances.5 7. 124 ChiSquare and Goodness of Fit. 129 Clomparing Two Normal Means with Known Variance.12 7. 142 8.15 8.10 8.14 7.viii 6. 122 Tests about the Variance. 137 Confidence Intervals for the Variance Ratio. 136 Comparing Variances. 12Y Unknown Variances.9 7.7 8.7 7. 140 8. 104 Unbiased Estimators. 134 The Duckworth Test.3 7. 132 The Paired lTest.15 6.
10 9.3 1 I . 183 .9 9. 158 I Charts.4 0. 172 Demerits. 164 Binomial Charts for a Fixed Samplc Size. 156 Setting the Control Limits for an s Chart.12 9. 175 177 11. 171 cCharts.2 11. 146 Setting thc Control Lines.2 9. 153 Alternative Stopping Rulcs.6 9.CONTENTS ix 9.8 Introduction. 154 Average Run Lengths.1 1 9.7 10.4 10.2 10. 174 Exercises.1 10. 182 What Quality is Accepted? AQL and LTPD.5 10.S 11. 151 Detecting Shifts in the Mean. 158 Process Capability.7 9. 178 Single Sampling Plans. Quality Control Charts 144 9.7 Introduction.1 9. 161 164 10. 180 Some Practical Plans. 154 s Charts. 144 Quality Control Charts. 179 Some Elemcntary Single Sampling Pliins.5 9. 177 Sampling by Attributcs.6 11.1 11. Acceptance Sampling I 11.15 9. 157 Nonnormality.14 9. 167 Interpreting Outlying Samples. 177 The Role of Acceptance Sampling.13 9.6 10. 165 Variable Sample Sizes. 160 Exercises. 149 Another Example.3 10. 148 An Example.4 11 .3 9. 164 Binomial Charts.16 Introduction. 169 The Arcsine Transformation. Control Charts for Attributes 10.8 9. 145 The xBar Chart. 156 Alternative Control Limits. 147 The K Chart.
193 Sequential Sampling by Attributes.12 11.1 13. 191 The Normal Case. 184 Other Sampling Schemes. Acceptance Sampling I1 12.2 13. AOQ and AOOL.14 12.1 12. 186 Chain Sampling Plans. 194 An Example of Sequential Sampling by Attributes.8 12. 204 .2 12.5 13. 201 Acceptance Sampling by Variables. 187 Sampling by Variables.8 11. Further Topics on Controi Charts 13. 189 Sequential Sampling. 192 An Example.3 12.15 Choosing the Sample Size. 204 An Example. 183 The Average Outgoing Quality. 199 Reduced Inspection. 194 Twosided Tests.X CONTENTS 11. 201 Exercises. 187 Skip Lot Sampling. 188 Exercises.6 Introduction. 187 Continuous Sampling. 197 The Standard Procedure. 200 Switching Rules. 203 Decision Rules Based on Runs.9 11.14 11. 198 Severity of Inspection.16 12. 206 \ 203 . 206 Moving Averages.17 12.6 12. 195 An Example of a Twosided Test.4 12.7 12. 197 MILSTD105D.18 The Cost of Inspection.12 12.3 13.10 12.15 12.5 12. 186 Rectifying Inspection Plans. 202 13. 189 Double Sampling.13 11.11 11. 196 Military and Civilian Sampling Systems.11 12. 188 189 12.9 12.4 13. 206 A Table of Values of ARL. 191 The Sequential Probability Ratio Tcst.13 12. 199 Tightened versus Normal Inspection.' Combined Rules.10 11.
245 More Than Two Predictors. 249 243 .10 13. 213 EWMA and Prediction. 233 Other Models. 219 Fitting a Straight Line.19 14. 229 Standardized Residuals.15 14. 243 The Method of Adjusting Variables.4 14.10 14.7 Introduction. 220 An Example.16 14. 221 The Method of Least Squares. 248 A Geological Example. 231 Confidence Intervals for a Predicted Value.7 13. 247 Properties of the Estimates.5 14.4 15. 248 More about the Example.12 14. 225 A Simple Numerical Example. 228 rTests.CONTENTS x i ArithmeticAverage Charts.11 14.8 14. 237 Exercises. 225 The Error Terms.2 15. 244 The Method of Least Squares.7 14. Bivariate Data: Fitting Straight Lines 14. 232 Lines through the Origin. 238 15. 212 Geometric Moving Averages. 228 Sums of Squares.18 14. 217 The Correlation Coefficient.14 14.20 Introduction.3 14. 232 Residual Plots.8 13.6 14. 207 Cumulative Sum Control Charts.2 14. 214 Exercises. 226 The Estimate of the Slope. 218 Correlation and Causality.9 13.1 15.9 14. 231 Influential Points.3 15. 215 217 13.6 15.5 15.13 14. Multiple Regression 15.11 14. 235 Transformations. 227 Estimating the Variance.17 14.1 14. 223 An Algebraic Identity. 209 A Horizontal VMask.
264 The Oneway Layout.7 16. 279 Several Factors. Design of Experiments: Factorial Experiments at Two Levels 17. 286 The Design Matrix.2 17. 264 The Analysis of Variance.10 16. 276 Interaction. 2'92 An Example with Four Factors.2 16.14 Introduction.1 16.3 16. 249 The Second Run. 290 Yates' Algorithm.8 15. 294 Normal Plots. 26U 264 16. 285 An Example of a 2' Experiment.5 17.4 17. The Analysis of Variance 16. 252 Fitting Polynomials.11 16. 293 Estimating the Variance. 280 Exercises.xii 15.9 17. 258 The Quadratic Blending Model. 256 Singular Matrices. 257 Octane Blending. 259 Exercises. 270 Quantitative Factors. 253 A Warning about Extrapolation. 288 Thc Regression Model.9 15.12 16. 276 Interaction with Several Observations in Each Cell.8 16. 267 The Analysis of Covariance. 278 Three Factors. 273 Orthogonality. 284 Experimenting on the Verticcs of a Cube.6 16. 267 The Analysis of Covariance and Regression.9 16. 255 Testing for Goodness of Fit.11 15.15 CONTENTS The Printout.6 17.10 15. 265 Which Treatments Differ'?. 287 Three Factors. 295 .13 16.8 17.10 Introduction.7 17. 274 Randomized Complete Blocks.4 16.3 17.13 15.12 15. 271 The Twoway Layout. 280 284 17.1 17.5 16.14 15.
312 Three Factors in 27 Runs. 296 Three Factors in Four Runs. 305 311 17. 335 Inner and Outer Arrays.12 18. 304 Fractions with 12 Points. 331 Terminology.I1 17. 296 Fractions with Eight Runs. 333 The 'I'hree Stages o f Process Design.18 17. 301 Foldover Designs. 323 HyperCraecoLatin Squares. 335 SignaltoNoise Ratios. 319 A Nonorthogonal Fraction.13 18. Taguchi and Orthogonal Arrays 19. Design of Experiments: Factorial Experiments at Several Levels 18.16 17.17 17.2 19.3 18. 303 Resolution V Fractions.12 17.4 18.3 19.7 18. 337 .8 18.5 19.19 17. 316 Thirteen Factors in 27 Runs. 313 Four Factors in 27 Runs.9 18. 297 Sevcn Factors in Eight Runs.5 18.14 17.11 18. 328 33I 19.13 17. 300 Screening Experiments. 299 The t(16) Lattice.4 19. 318 The L(18) Lattice. 304 Exercises. 333 Parameter Design Experiments.1 18.10 18.I 19. 323 GraecoLatin Squares.6 18. 328 Exercises. 298 The L ( 8 ) Lattice.15 Factors with 'Three Levels. 324 FourLevel Factors and TwoLevel Factors.15 17. 321 Fractions: Four Factors in Nine Points.CONTENTS xiii Fractions. 325 Factors at Five Levels. 333 A Strategy for Parameter Design.6 19. 311 Two Factors.2 18.14 18. 324 Five Factors at Four Levels in 16 Runs.20 18. 319 The L(36) Lattice. 321 Latin Squares.7 Introduction.
340 Tolerance Design. 341 Finale.11 References Tables CONTENTS An Example of a Parameter Design Experiment. The 11. 338 The Confirmation Experiment. 351 tDistribution.xiv 19. 353 Chisquare Distribution. 343 347 35 1 I.8 19.9 19. 356 367 .10 19. The 111. The Index Normal Distribution. 354 F Distribution. 341 Exercises. The IV.
Since returning to academia in 1961. It reflects more than 30 years of teaching statistics to engineers and other scientists. Then came the mainframe. from 1957 to 1961. which began with my experience as the first research statistician in the laboratories at Chevron Research Corporation. a key element of modern quality assurance. Under some accounting procedures. you can imagine how satisfying I find American industry’s current surge of interest in quality assurance. That aspect of the design of experiments has become the basis of offline quality control. both in formal courses at universities and in specialized courses at their plants. we could not perform such tasks as multiple regression. I have devoted much of my research to fractional factorial experiments. statistical calculations were a forbidding chore. Why is it s exciting to have these capabilities available for the PC? o Before the 195Os. after three decades of everaccelerating progress in both hardware and software. with particular emphasis on modern quality assurance. including both chemical engineering and semiconductors.Preface This book is designed to give engineers an introduction to statistics in its engineering applications. The examples in the text cover a wide range of engineering applications. augmented by broad consulting experience. As a practical matter. which meant that the engineer had to work through the computing section with its programmers and keypunch operators. the personal computers that most engineers have on their xv . My own interest in quality assurance evolved from my longstanding specialization in the design of experiments. using them could be intimidating and often complicated. People felt inhibited about using the company’s computers unless they had a relatively big problem and an expert to help them. Given that background. Although the mainframes facilitated wonderful progress. it was also costly. Now. Almost equally satisfying is the burst of technical capabilities for the implementation of quality assurance methods by engineers at plant levela direct result of the wide accessibility and growing power of the personal computer.
1 hope that it will persuade engineers to use statistical thinking as a standard operating procedure. . today’s engineers can make plots that normally would have taken their fathers all day. to a word processor for composition. using Minitab for calculations and Statgraphics for figures. PETER M. Even though engineers now can and should perform many of the applications indcpcndently. It has thc great advantage of being very easy for the students to use. Not only has the PC freed engineers from the mainframes for everyday computations. too. However. they can apply standard statistical procedures easily and quickly. use statistical methods and statistical thinking on an everyday basis. Then. are available in versions for the PC with a hard disk. J can turn to SAS or to GLIM. In just a few seconds. This is particularly advantageous with graphical procedures. Now engineers can. which I find more convenient.xvi PREFACE desks can do more easily and quickly all that the mainframes did in the late 1950s. When I have more complicated problems. JOHN W. I use Minitab on the mainframe for my teaching at the University of Texas. and should. Parts of the last chapters on the design of experiments may look deceptively easy. All of the computations that I have made in writing this book have been carried out on such a PC. As to thc more complicated problems. 1 have been able only to touch upon some of them. it has lately become available in a PC version. I also use Minitab in much of my rcscarch. of course. it is even better to have convenient access to a statistician with whom to discuss problems and seek advice. but better still. they. Moreover. Hcnce. in addition. We are at a stage in our technological development when engineering and statistics can walk together hand in hand. This book is only an introduction to a broad field. as a matter of routine. the engineers themselves can keep their data files on their own disks. both to collect thcir data and to analyze them. and most of them use it on our instructional VAX. 1 intend this book t o help us both to do that. most cnginecrs have little or no occasion t o deal with the mainframe that niay be miles away from their plants. a few words with a statistician before you get started may save you hours of pain later.
Statistical Methods in Engineering and Quality Assurance .
Industry lost interest as soon as the war was over. H. and they knew the conventional wisdom. you had asked engineers or executives what quality assurance was. except for those cornpanics that were engaged in defense work. academic statisticians. It is not too harsh to say that American industry became interested in quality control during World War I1 because the government insisted upon it. 30 years ago. good. as we are engaged in a life and death struggle with foreign nations that hardly existed as competitors 30 years ago.Statistical Methods in Engineering and Quality Assurance Peter W. F Dodge. We will be beaten not by price. accepted some. M. where the government still insisted on its use. bad. worked on the development of acceptance sampling. John Copyright 0 1990 by John Wiley & Sons. but by inferior product quality.1. One hoped that they were able to spot most of the defective parts. mainly at Princeton and Columbia. If you had asked them what quality control was. During World War 11. that group spread out over the country to form the core of the growth of statistics in American universities. After the war. the inspectors inspected the lots. The topic was relegated to a few lectures in departments of industrial engineering. Without dedication to total quality in our manufacturing companies. In some industries. QUALITY AND AMERICAN BUSINESS If. quality has beome a vital part of American industry. They were the work of such pioneers as Walter Shewhart. That is what quality control and quality assurance were. the answer would have been something about Shewhart control charts and government sampling plans. and rejected others. which were then either scrapped or reworked. they probably would not have known the expression and would have said something about being sure that we kept on doing things as we had always done them. the production people churned out the parts. or indifferent. Today. Inc CHAPTERONE Quality Assurance and Statistics 1. Some managers 1 . we will be unable to compete in areas that once we dominated. But they no longer worked on quality control. of whom there were only a few. Two of the main techniques in quality assurancecontrol charts and acceptance s a m p l i n g 4 a t e back to the 1930s at Bell Labs. and Enoch Farreil.
3. Once Americans listened to them. Some did not even bother to control costs. began producing cars. Some Americans tried to attribute the increase in imports from Japan entirely to the alleged fact that Japan’s wages and standard of living were lower than ours. It meant that service after the sale was often subject to cavalier neglect. Deming emphasizes that quality requires the complete dedication of management. Marketing sold the product. The Japanese. but often little real improvement over the previous year‘s product.2. Once the production department shoveled the product out of the plant door. inflation psychology was rampant and extra costs could be passed through to the customer with no great difficultyespecially if that customer was Uncle Sam and the manufacturer had a costplus contract. 1.. who had earned an international reputation for shoddy workmanship bcfore 1940. the user. that were not only cheaper than ours. THE REACTION The reaction came a few years ago. This meant two things. Juran had also been preaching quality control in this country and in Japan for many years. became a prophet in his own country. The number of Japanese cars on our highways and television sets in our homes increased by leaps and bounds. Socalled new models might appear annually with cosmetic changes. but was being completely ignored in his homeland. One of those things was to listen to William Edwards Deming. ctc. He summarizes this requirement in “Fourteen Points for . M. and would have been happier to ship the product as is and let the buyer beware. the division of functions was simple. it became marketing’s problem.2 QUALITY ASSURANCE AND STATISTICS resented the expense and the hassle of this. In some companies. It also meant that there was really very little feedback from the final buyer. COMPETITION Then came the Japanese shock. Then people began to realize that the explanation was that the Japanese had done some things right. For a time. but worked better and lasted longer. an American statistician who had started the Japanese on the road t o quality in 1950. Engineers began to talk of quality assurance rather than quality control. American manufacturers then preferred to focus on cost control rather than quality control. Warranty and f service were pushed down to the level o the local retailer or the local distributor. 1. The term total qualify entered the jargon o f manufacturing. and to the American engineer J. Deming. a radical change began to take place in much of this nation’s industry. called back from the wildcrness. Juran. chips. to the manufacturing department concerning what was wrong with the items and how they could be improved. televisions.
4. I t demands a realization that quality is the concern of everybody. . One of the leading converts to total quality. No step in the process stands alone.” and “Adopt the new philosophy. The word “customer” does not just mean the end recipient who pays for the product. Houghton. Total quality means a dedication to quality by the entire company.” The final point is “Put everybody in the company to work in teams to accomplish the transformation.” of which the first two are “Create constancy of purpose toward improvement of product and service. It emphasizes that quality is a matter of’dcdication of the whole company and all its personnel. N o worker is an island. It emphasizes the old. from the CEO down to the employee who sweeps the corridors. All workers are part of a team working together to achieve quality. All steps are combined in the process. James R. You also have your own customers within the plantthe people at the next stage of the production cycle.” Then he boiled it down even further to two objectives: “Quality is everyone’s job.TOTAL QUALllY 3 Management. and then doing it rightthe first time.” 1. and not just a function o f a few inspectors stationed one step before the customer shipping area. Do it right the first time. much maligned idea that the objective is customer satisfaction: to supply the customer with good product. the C E O of Corning Glass Company. we are in a new economic age.” The first of these two objectives has farreaching social implications. TOTAL QUALITY The first of Deming’s points sets the tone for total quality. summed it up more concisely in 1986: “Quality is knowing what needs to be done. having the tools to do it right.
1. insertions and modifications from middle management. Total quality involves workers in a cooperative effort. and so on. decrease warranty costs by 15%. The key factor is the pervusive use of statistical methody. sometimes with unwise. The division head sets goals or quotas for each of his department heads: increase sales by lo%. It is certainly not a part of the Frederick Taylor model of scientific management that has dominated managerial styles in American manufacturing since it was introduced at the beginning of this century. Its essence is the breakdown of the manufacturing process into pieces that are then bolted together in a rigid hierarchical structure. The Taylor modcl is sometimes called management by control. Why the pervasive use of statistical methods? Where does statistics comc into this picture other than just in teaching people how to draw Shewhart charts? The distinguished British physicist Lord Kelvin put it this way: . with initiativc encouraged and responsibility shared from top to bottom. Simply ordering the workcrs to do this or that is now seen to be counterproductive. Tom Peters.” The Japanese call this kaizen. KAIZEN The fifth of Deming’s points is “Improve constantly and forever every activity.4 QUALITY ASSURANCE AND STATlSTlCS This whole concept of the workers in a plant or a company as a team is revolutionary to many people. WHY STATISTICS? How is kaizen achieved? It is achieved by integrating the research and development effort with the actual production facility and devoting a lot of time and money to “getting the process right” and then continually making it better. who writes on excellence in manufacturing. meaning continuous searching for incremental improvement. There is an expression “overengineering” that is used disparagingly for wasting time fussing at improving a product once it has reached a marketable state. The new thrust in management requires a fundamental change of style. and perhaps unauthorized. I . The CEO gives his division heads a profit objective and down the line it goes.6. These dicta end up as quotas or work standards at the shop level. American manufacturcrs have tended cither to go for the big leap or t o stop caring about improving the process when they think it is good enough t o do the job. points out that some authorities consider this Japanese trait the single most important difference between Japanese and American business.5.
Will American manufacturing. The United States has now established the Malcolm Baldrige award for quality. who. changes must be made in the way things are presently being done. If quality and productivity are to improve from current levels.” But perhaps the most cogent summation is that of W. and a host of other major companies. Quality assurance now has official recognition. your knowledge is of the meager and unsatisfactory kind. We could add Westinghouse. THE FUTURE Some of us who have been involved in the quality revolution have asked each other. once back on track. 2. G. when you cannot express it in numbers. vigorously advocates his program of statistically designed experiments. you know something about it.” Quality assurance pioneer Deming was trained as a statistician. Motorola.7. Key examples may be found in the steel industry. We mentioned Corning Glass earlier. and. Taguchi. Many other. We should like to have good data to serve as a rational basis on which to make these changes.THE FUTURE 5 “When you can measure what you are speaking about and express it in numbers. how should they be analyzed? 4. in our gloomier moments. . with his disciples. It was awarded for the first time in the fall of 1988. named after thc late Secretary of Commerce. but when you cannot measure it. companies are achieving outstanding results from their emphasis on quality. Every year the Japanese present their prestigious Deming Award to the company that makes the greatest impact in quality. It has been won by several of the companies that have become household words in this country. Ford. then lose its appetite for quality assurance and go back to its old noncompetitive ways until another crisis hits‘? There are grounds for increased optimism. There have been wellpublicized success stories in other industries besides scmiconductors. Hunter of the University of Wisconsin. The twin question must then be addressed: what data should be collected. where old dinosaurs are being replaced by small speciality firms that have found niches in the market for highquality items. Now Japan’s principal lcader in quality control is an engineer. once collected. less famous. who said: “1. of which more will be said in the last three chapters. Among the winners was a major semiconductor producer. 3. G . 1. Statistics is the sciencc that addresses this twin question. whether or not this is perhaps a passing fad.
And that is before taking into account sales lost through customer dissatisfaction. Two statistics can be cited in support of this principle. he has tended to increase that earlier figure of 85%. It should be a long while before they forget it. When another engineer uses a catalyst bought from another manufacturer. Engineers conduct experiments all the time. In some companies the cost of poor qualitythe cost of the twin derrions SCRAP and REWORK runs somewhere between 20 to 30% of sales. and morewho had won Preferred Quality Awards for the quality of the parts that they supplied to the automotivc manufacturer. she is doing an experiment. thc data should be obtained in an organizcd way b that their yield of information shall be as rich as possible. The second. O n the day that this paragraph was drafted. Juran has said for some time that at least 85% of the failures in any organization are the fault of systems controlled by management. more telling statistic came from James Houghton. we must learn more about them. 'They have to be gathered and analyzed. down to the grass roots of American manufacturing. Customers demand and expect quality from their manufacturers. and.6 QUALITY ASSURANCE AND STATISTICS Concern about quality is becoming pervasive. Quality pays. whcther we like it or not. in wcights o f pigs from the same litter. They may not realize it. Fewer than 15% are workerrelated. 1. The statistician o can help to extract all the information that is to be found in the data. We are working. When an engineer nudges thc temperature at one stage of a plant by a degree or so. Manufacturers demand quality from thcir suppliers and train them to produce the quality parts that they need. and management is learning that. We cannot just sit in our offices. Better still. 'There are data out there. Quality is working its way down to the small subsubcontractor. developing the mathematical theory. For them it was the reality of variability in weather. Brazil.8. Shewhart's great contribution was to educate engineers about the existence of Variability. Agricultural scientists already knew about natural variability. if . and then turn to the computers to solve the differential equations. They ncccssarily learncd to conduct experiments in its presence. American nianagement has been learning that statistic the hard way. the Wall Street Journal carried an advertisement from Ford giving a list of their suppliers in the United Statcs and other parts of the worldMexico. Liechtenstein. he is doing an experiment and geting some data. GATHERING INFORMATION If we are to improve our manufacturing proccsscs. in soil fertility. Lately. in the presence of variability.
The semiconductor industry is full of horror stories about devices that have been designed miles away from the future fabricators and “tossed over the transom.9. 1. INSTRUCTIONAL CONSIDERATIONS There is enough material in this hook for a twosemester course. fitting models to the data by the method of least squares. How can the process be improved‘? Chapters fourteen and fifteen arc devoted to rcgression. will help to choose a designed experiment to yield good data. is mainly devoted to the socalled Taguchi method. The last chapter.c. Now let the manufacturing engineer make their design nianufacturable! But how‘? The manufacturing engineer has to learn more about the process and to find good. operating conditions. i. That procedure has become the basis of modern experimental design. The instructor who has only one semester will necessarily make some choices. nineteen. preferably optimum (whatever that really means). 1. Chapter sixteen is an introduction to the analysis of variance. . where it is widely used. Taguchi in Japan. The Taguchi method combines the ideas developed by Fisher. It is now used increasingly in the west as well. They consider how to conduct experiments to determine thc best conditions for the operation of the proccss. Although based on the work of Fisher and Frank Yates.10. Those topics deal with passive quality assurance. The last six chapters point the reader toward active quality assurance. Chapters two through eight can be taken as a course in probability and statistics designed for engineers. That means carrying out experiments at the manufacturing level. Cuthbert Daniel. these chapters focus on the developments of the past 40 years in industrial applications by George Box.INSTRUCflONAL CONSlDERATIONS 7 given the opportunity. Chapters sixteen through nitietcen are concerned with offline quality control. The charts tell whether the process is moving along under statistical control. Chapters seventeen and eighteen are about factorial experiments.. Box. The sampling procedures tell whether the incoming or outgoing batches of product meet the specifications.” The design people have made one batch of the prototype device under laboratory conditions. Chapters nine through thirteen deal with the traditional topics of quality controlcharts and acceptance sampling. THIS BOOK There are 19 chapters in this book. and others with some new insights and emphases by Taguchi to form a systematic approach to conducting online experiments. and others. This is thc procedure for designed cxperirnents developed by G . a procedure that Sir Ronald Fisher introduced 60 years ago to design and analyze agricultural experiments.
or at least to omit the sections on sequential sampling in the latter chapter. judicious cuts could be made in the assignment of introductory segments. In the last part of the book. and communicating the results to the user so that valid decisions can be made. The emphasis in the later sections would depend on the focus of the course. 1.11. Others may elect to omit the advanced topics on control charts in chapter thirteen. Those techniques for designed experiments are fundamental to the improvement of manufacturing processes.8 QUALITY ASSURANCE A N D STATISTICS For example. he could omit some of the more advanced and specialized topics in chapters three through eight. However. REPRISE Why do engineers need to learn statistics? We have mentioned some reasons earlier. . some instructors may choose to leave out all of the material on acceptance sampling in chapters eleven and twelve. one could postpone the last half of chapter eighteen and some of the specialized topics in the two chapters on regression. any more than a general can make good decisions on the battletield without good intelligence. any modern course in statistics for engineers should include the material in chapters seventeen and nineteen. That is what statistics is all about: gathering data in the most efficient way. We have to gather good data and interpret it properly. If the students already know some statistics. preferably by properly designed experiments. For example. analyzing that data correctly. Less thorough coverage could be accorded the chapter on analysis of variance. You cannot make good decisions about quality without good information.
1. Two of the most commonly used statistics for describing a sample are the mean and the standard deviation. There are two ways of doing this. Inc CHAPTERTWO Descriptive Statistics One picture is worth more than ten thousand words. which is written as X.1. . which at first glance convey little.. .1 shows the daily yields (percent conversion) in 80 runs in a chemical plant. In the next sections.or xbar. As they stand. is defined as 9 . we show how the points can be presented graphically by a histogram and also in several other ways. the mean. we make this set of 80 numbers more intelligible in two ways.x. x 2 .2. we discuss how one might summarue and prcsent a large sample of data. but nowadays computers have made it easier. We can calculate from the sample certain important statistics like the mean and the standard deviation that. THE MEAN AND STANDARD DEVIATION The word statistic is used to describe a number that is calculated from the data.Statistical Methods in Engineering and Quality Assurance Peter W. 2. INTRODUCTION In this chapter. calling for hours of tedious tabulation. John Copyright 0 1990 by John Wiley & Sons. Then. It is the simple avcrage of the data. 2. First. . in somc scnsc. . Table 2. we derive some numbers (statistics) that summarize the data points.. We use this data set to illustrate the ideas throughout this chapter. An obvious candidate is the average of the 80 observations. these are only 80 numbers. . If there are n observations denoted by x . characterize the data. M. One is by words and figures. The other is by using graphical methods. 'The name mean is short for the arithmetic mean. This used to be a chore.
2 70.3 72. The average of the squares of the deviations is the variance of X.5 75.9 71.2 78. .0 69.1 72.q2 The divisor is n .5 69. we do riot know p .4 74.5 77.2 68.3 69.0 69. the sample mean provides a good estimate of p .9 66.4 66.0 73. ..6 63. w’.3 DESCRIPTIVE STATISTICS 73. Then the variance is estimated by the sample variance ~ c (x. so we redefine dev x .1 72. Our set of data is a sample of 80 observations from that unknown population..0 77. For the set of data in table 2.1 76.1 73.0 80.5 64.9 74. 71.3 69.9 67.1 74. The variance and standard deviation of x are measures of the spread of the data.4 75.1 73.2 68.) Usually. xbar = 71.5 73.2 where the summation is over the values i = 1. denoted by the greek letter p. we could also have written X = G.7 60.5 72.5 69.8 70.7 76.1.7 73.5 72.5 75. (We say more about the criteria for “good” estimates in later chapters.6 74. x .p is the deviation of the ith observation from the population mean.3 68. X.4 61. For any observation x .1 72.4 66. .1 73.7 72. 2 .5 68. observations.8 76. the deviation from the sample mean. Its average is the population mean. represents the center of a distribution of mass along the xaxis. .4 70.. .0 70.. the variance is estimated from the sample by (The hat denotes that the expression is an estimate from the data.1 rather than n because xbar has been used instead of p.8 71. We try to use uppercase letters. n.X.8 69.4 68.9 70.8 73.9 71.9 70.0 71. to denote the names of variables and lowercase..) The mean represents the center of the data in the same way that the centroid.6 69.7 69.2 70.6 74.8 77.7 69. i.1. . as x . = xi .4 68. . The population mean is often unknown and has to be estimated from the data. or When is known. The set of all possible daily yields of the plant forms a statistical population.7 68. the quantity dev x . It is denoted by V ( X ) .0 68. or center of gravity. but it is very hard to be absolutely consistcnt in this notation.7 68.8 76.5 65. not been taken under unusual or biased circumstances.5 70. If the sample ha.5 77.9 76. to denote values that thc variables actually take. .6 75.3 70.438.e.7 68.10 Table 2 1 1 Yields of 80 Runs in a Chemical Plant .4 73.4 68.
Q . when N = 80.25. We define Q . any number 4 5 x" 4 8 will satisfy the definition. 2 3 THE MEDIAN AND THE QUARTILES . Suppose that the ) data are written down in ordcr and that x ~ . In the example of table 2. If n = 2t + 1. x"= (x(?. lies to the left of the median. s = 3. I.75 and so Q 3 is defined by .. with only five observations. The median is the sccond quartile. is the ith largest observation. then Qi = x(U) and ( 2 3 = y(_?. we use linear interpolation.))+ x(. Just as the median divides the data into two halves. on the average. It is usually f estimated by s. the median is defined as If n = 21.1. 8 . For this set of data. For example. and 9. and take the middle observationthe third one.') . 3u = 60. we arrange them in ascending order. 1. I . is the square root o the variance. although we use a different method when we comc to set up Shewhart charts.. 8 .79. u = 20. and 2. . 4. For example. The median is mother rneasurc of the center of the data. The first quartile. If u is an integer. Approximately oncquarter of the points are below it and approximately threequarters are above it. If we only have four observations.. 2. The sample median is a number with the property that there are as many points above it as below it. the sample median is clearly defined.35. the points arc arranged in ascending order.THE MEDIAN A N D THE QUARnLES 11 We see later that dividing by n would give a biased estimatean estimate that is. and so we define Similarly.and Q 3 in the following way.)/2=71. 8. If n is odd. 4. too low. 4. a. 9. If u is not an integer. As before. and 9. A less ambiguous definition follows.1. The standard deviation. the quartiles divide it into quarters. Let u = ( N f 1)/4.
If you are in the habit of storing your data in computer files.1. the quartiles.3. the data and the results of calculations are contained on an imaginary work sheet. the standard deviation. etc.925 Q1 STDEV 3.I & High school counselors may say that you havc to be “in the first quartile” of your high school class if you are to go on to college. Description of the Yield Lhta N 80 MEAN 71.175 SEMEAN 0. Some other packages intcrpolate differently when ZI is not an integer. The machine prints out the mean. the trimmed mean (TRMEAN) and the standard error of the mean (SEMEAN).788 03 74. They mean that you must bc in the top 25% of the class. The points that divide the data into ten Table 2. maximum. > .350 MAX 80.00 ‘TRMEAN 71 .3. With N = 80. Onc sets the data in thc first column and types the command DESCRIBE C1. In the Minitab package. as is shown in table 2. quartiles. the maximum. or MINI CI.546 68. These statistics are easily obtained on personal computers. but they are not universal.12 DESCRIPTIVE STATISTICS Those are the definitions that the Minitab package uses. They place the quartile midway between the observations immediately above and below it. Statistics such as thc median. median.423 . Others interchange &. and the minimum are called order statistics. We have defined a quartile as a particular value that separates onequarter of the data from the remainder.438 MJN b0.700 MEDIAN 71. and minumum values and two other statistics that are introduced later. so that (2. Some authors refer to all the observations in the bottom quarter as constituting the first quartile.1. and Q 3 . they define and There is another possibility for confusion about quartiles. Individual statistics can be obtained by typing MAXI C1. it takes only one or two typewritten commands to read a data set into the work shcct.
but their visual impact is simple and directly useful. Each student is a dot on . you count the number of marks in each row.11. tally sheets (or check sheets). to present the breakdown of the grades as a histogram. B. They involve no deep mathematics and are rarely mentioned in mathematical statistics courses. The Pareto diagram is another type of bar chart. 2. and then you take the exams one by one. On a piece of paper. His seven tools are the following: 1. the fifth is a diagonal mark through those four. you make a vertical mark in the B row. graphs. 4. You might also choose to stratify the data by making separate charts for men and women. These are simple graphical tools. making the set of fivc look like a gate. using linear interpolation if needed. How many of the components that were rejected failed for which of several defects? How many votes were cast for each of several candidates? How many students in a large class got A. Pareto diagrams. 5. or for graduates and undergraduates. causeandeffect diagrams. The tally chart is only one step removed from the histogram. 2. or F?The idea i s simple. it is discussed in section 2. Scatter plots demonstrate the relationships between two variables. or bar graph. ISHIKAWA’S SEVEN TOOLS lshikawa (1976) proposed a kit called “The Scven Tools” for getting information from data. The tally chart has long been used for counting how many objects in a group fall into each of several classes. and there are the totals for the various grades. After you have gone through all the papers. stratification. 3. The first four marks are vertical. you make rows for A. If the second student failed. B. and so on. after tallying. one can plot the entrance examination scores in English and mathematics with English along the vertical axis and mathematics along the horizontal. . C. 6. . It would be quite appropriate.ISHIKAWA’S SEVEN TOOLS 13 groups are called deciles. and 7. graphs. . For a batch of incoming students at a university. If thc first student got B. . scatter plots. whose use is enhanced by modern computers.4. histograms. D. The counting is usually made easier by counting in groups of five. it is exceeded by 5% of the data. and we can also use percentiles. which is the topic of the next section. Control charts fall under the last category. The ninetyfifth percentile may be defined as the observation in place 95(N + 1)/ 100. you make a mark in the F row.
4. It contains one observation.0 + 1). Into each of these diagonal lines we draw lines for subareas and then lines for subsubareas.0 centered on the even integers.1. that we are concerned about the number of defective items being produced.7. and so on.14 DESCRIPTIVEsTAmsrirs PE0PI.O occurs in both intervals. usually of equal width.1. and maintenance. There is no agreement on . The resulting diagram looks like the backbone of a fish.0 (=60.1.4.8. Would that suggest a pattern? A causeandeffect diagram looks like the diagram that students in English grammar classes used to draw to analyze sentences. Suppose. A rectangle is then erected on each interval. the graph. It was obtained by typing the command HISTOGRAM C1. e.g. one could use stratification. An example is shown in figure 2. The second interval runs from 61. showing that scores in the two subjects are highly correlated. As we have phrased it.1 shows the histogram for the data of table 2.5. for example. The area of the rectangle is proportional to t h e number of observations in the interval.1) to 61. 61. The first interval runs from 59. The prograni has arbitrarily elected to use intervals of length 2. men.0 and also contains one observation. Figure 2. The range of values taken by X is divided into scveral intervals.5. or do they fall into the shotgun pattern that suggests independence? Again.0.0 (=60. Then we draw diagonal lines running to it to denote major areas that lead to that defect. machines. 60. the value 61 . HISTOGRAMS One way in which a sample can be represented graphically is by using a histogram. 2. materials. Minitab prints the histogram sideways. Fishhone diagram. using one symbol for engineers and another for arts students. Is there a pattern? Do the points tend to lie on a line. We draw a horizontal line on a piece of paper to denote the defect.1. and so these causeandeffect diagrams arc commonly called fishbone diagrams.0 to 63.E PA I NT Figure 2..
Statgraphics puts it in the lower interval.0. 74.1.0. 68. Most software packages can give you intervals of your own choosing by a simple modification in the histogram command. For example.5.0 < x 6 63.1. In this case. The number of intervals is also arbitrary. 72. It has a slightly diffcrcnt shape. respectively) are close to one another.0 that are treated differently in the two packages. It is not a question of one writer being corrcct and the other wrong. so that the intervals are 59. Minitab has 12 observations in the interval centered on 68 and 19 in the interval centered on 70. . it is a symmetric distribution.1 are usually almost symmetric.0. this too happens with samples of data from symmetric distri butions. or normal. which shows another histogram tor the same data set obtained with Statgraphics. Minitab puts it in the upper interval. 69. . The normal distribution is discussed in later chapters. Other programs put them in the lower interval. the modal interval is centered on 70. 70.0.35. There i s an obscrvation at the boundary. Similarly. there are two obscrvations at x = 73. 62. 66. 80. NUMBER OF OBSERVATIONS 1 * I * 2 ** 4 **** 12 ************ 19 ************* 12 ************ *4*********** I4 9 ********* 5 ***** I * Figure 2. distribution. The interval with the most observations is called the modal interval. Thib difficulty is illustrated by figure 2. The distribution of the yields appears to follow the bellshaped curve that is associated with the Gaussian. We have already noted that the sample mcan and the sample median (71. ****** * handling this. Some authors suggest counting them as being half a point in each interval. 64.5. Statgraphics has 13 in the former interval and 18 in the latter.2. Histogram for table 2.HISlOGRAMS 15 MIDDLE OF INTERVAL 60. 61. .1. but it gives the user freedom to choose 10.1. 78. The difference between the two is a matter of choice. 76. Sevcral other things should be noted about these histograms. Statgraphics suggests 16 intervals for this data set.44 and 71. The histograms of sarnplcs o f normal data like that in table 2. . Minitab puts points on the boundary in the higher interval.0 < x C 61.
8 Yields 72. It is obvious that the two highest priced houses are markedly more expensive than the others.2.4 67. AN EXAMPLE OF A NONSYMMETRICAL DETRIBUTION Some years ago a realtor listcd in the Austin AmericanStatesman newspaper the asking prices of all the houses that he had on his list. figure 2. shows those two houses appearing as outliers. The prices are listed in descending order.6. They are given in thousands of dollars in table 2. The histogram. there is a clear explanation of such outliersthere are always a few expensive houses on the market for the wealthy buyers. I . Suppose that we can regard this list as typical of a realtor serving clients in the middle and upper income brackets in those days.1. we might have doubted that . In this case. 2.2 76.5. If thcse had been measurements of the yield of a chemical plant. Frcquency histogram.16 20 I I I I I I I I I DESCRIPTIVE STATISTICS I l l ! I l l / I l l 1 59 63.6 81 FIgure 2.6.6.
5oc) 74. The histogram also shows that the distribution is not symmetric.950 67.950 119.500 62.CwK) 156.500 134.950 69.950 85. On the other hand.000 54.I .ooo they were valid and gone back t o the records to check whether mistakes had been made in recording the observations.200 76. 140.000 79.500 62.000 7Y.500 59.950 53.000 149. A case can be made that. 60.000 144.050 124.950 94.900 L29.000 65.500 74.500 122. 240. the mean will vary considerably. We saw earlier.W) 60.6.500 96.950 195.500 129. 80.900 73. _  Table 2.0(K) 55.950 92. 160. NUMBEK OF OBSERVATIONS 3 *** 18 15 7 8 4 ****************** * * * * * * :k * * * * * * ******* ******** :j~ 2 1 0 0 2 1 1 **** ** * * * ** Figure 2.900 71. Ilistogram of housc priccs.1.900 __ .500 49.050 275.950 51.500 94.950 70. for an ordinary person coming to town.OOO 154. 220.AN EXAMPLE OF A NONSYMMETRTCAL DISTRIBUTION 17 .000 93.950 74. This is because the median is much less affected by outlying observations. 260.500 66.500 125 104.OOo R9.900 52.950 56.000 61.900 39.ooo 225. Each provides an estimate of a central value of the distribution. 100. in the case of a sample from a symmetric distribution.500 88.500 51.950 48. that the sample mean and median are closc to one another. the median is a better estimate of how much one should expect to pay for a house than the mean. according to whether or not the head football coach was included in the sample. .000 110.500 77.000 139.500 114. the coach has less effect on the median. 120.6.500 95. one can argue about what is meant by a central value. 180. MIDDLE OF INTERVAL 40.000 69. 280. For a nonsymnietric distribution. House Prices (in Thousands of Dollars) 208. It is skewed with a long tail to the bottom.950 178.900 87. 200.450 57.950 71.1. If one takes a small sample of faculty incomes at an American university.
a decrease of 6%.925 .998. the mean drops to $91.25%). or to provide a morc robust estimate of p . The median decreases to $78. It can be argued that the median itself is essentially a 50% trimmed mean because all the obscrvations are trimmed away save one! Sports fans will note that in some athletic contests. outlying values directly. The amount of trimming is arhitrary.2 4.1 3.U 5.6 4.982 and i = $79. 71. The median is said to be a morc robust statistic. it is more scnsitivc to outliers.8 4. observations are close together. When the two most expensive houses are removed. her highest and lowest MIDDLE OF INTERVAL 3.2. a loss of only 1.6. In this example. respectively. This is why. where the . some packages trim as much as the upper and lower 20% o f thc data. some of which may bc erroneous observations.6 3.0 5. This is obtained by dropping the largest 5 % of the obscrvations and the lowest 5% and finding the mean of the remaining 90%. The DESCRIBE subroutine that was used in section 2. thc four highest and the four lowest observations were dropped. Thc reason is that the calculation of the median never involved the . even when estimating the population mcan for a normal distribution.430. among othcr statistics. .925. which is symmetric. Although the sample mean can be shown t o be ti more precise estimate mathematically when all the observations arc valid. some scientists go further and use trimmed means. the trimmed mean is 71. ) ) /a2 small change in the middle of the data.)+ ~ ( ~ .+ ~ ( ~ ~ to ) / 2 (xf3[.18 DFSCRIPI'IVE STATISTICS The mean and the median of the housc prices are .0 NUMBER OF OBSERVATIONS I * 1 10 12 12 7 8 4. the rating of each competitor is a trimmed mean. some statisticians prefer the median over the sample mean. Indeed. such :is Olympic gymnastics. Each competitor is evaluated by a panel of several judges.2 5.546 as opposed to thc mean of the whole set. The median m o w s only from (x(?[.U = $97.4 6 2 ********** * * * * * * * #: * * * * ************ ******* ******** ****** ** * * I Figure 2. the 5% trimmcd mean of the yield data.3 gave. IIistograin of iiarural logarithms of house prices.
important that the engineer decide on thc level of trimming before looking at the data and n o t aftcr seeing the observations and forming an opinion. the curve is still skewed.1 on the same horizontal scale. I 74.00 __ t 70.m . her rating is the average of the remaining scores.2 shows dotplots of the first and third rows of table 2.I  76 00 . Wc have no reason to expect that the housing prices should follow the normal bcllshapcd curve. Figure 2. Figure 2.DOTPLOTS 19 scorcs are discarded (trimmed). It has been suggested that this is a prccaution against outliers caused by unduly partisan judges. DOTPL. we see immediately that the averages of the two rows arc about the same. I selected the third row on purpose to illustrate this point: there may actually have been an improvement in the variance during that period. It is. of course. 2. but i t sccms to have been temporary.2.. The engineer draws a horizontal scale and marks the observations above it with dots. 72. It is very easy to make and is useful when there are fewer than 30 points. .1. we can try a transformation.6. This second histogram is morc like a bell curve than the first. but the third row has less variation than the first row.OTS The dotplot is a simple graphical device that is rathcr like a quick histogram. It is a simple diagram that can bc made on any piece of paper without any fuss.7. If it is important to have a variable that has a bellshaped distribution.W .1.2 shows the histogram of In(price) with the two highest houses omitted. + . Figurc 2..7..00 .r~) Figure 2. Dotplots arc also useful when comparing two samples.  .Kou 3 7fi. Dotplot o f rows 1 and 3 . 1 68.7. One possibility is to replace the price by its natural logarithm. Even when we omit the two highest houses.7.1.00 .1 shows a dotplot of the last 10 yiclds in table 2.
STEMANDLEAF DIAGRAMS A variation of the histogram was introduced as a tool in exploratory data analysis by J.8. 51. This can be considered a point in favor of the stem and leaf because it shows more information without adding more clutter to the picture. 49. or rectangles. The histogram only has diamonds. The other digits form thc stems. It is called the stemandleaf diagram.1. 54.20 DESCRIRIVE STATlSTICS 2. there has to be a stem every $10. Stcniandleaf display o f house p r i a b . 53.8. 48 and 49 becomc stem 4 with leaves 8 and 9.0OOO 1 2 REPRESENTS 12. 51. respectively. O n the other hand. W. We now illustrate the procedure for the house price data in table 2. separates 225.1. 52. 48. The last digit of each trimrncd observation is called a leaf. 275. In the cxample. thc digits are written as leaves. 280 . Tukey (1977).6.1. In the stemandleaf diagram. There are two main differences between the histogram and the stemandleaf diagram. 195 is leaf 5 on stem 19. The Minitab program. . stars. the stcms have to come away from the trunk at unit intervals. In ascending order. . 1 3 12 21 (11) 30 26 20 19 16 11 3 9 4 89 5 I 12345679 6 0 12256799 7 01 134446799 8 5789 9 234456 10 4 11 049 9 7 5 5 4 4 14 49 1s 46 16 17 8 18 12 24500 1.000. 39 is leaf 9 on stem 3. .8. and 280 at the end as high values. The diagram is shown in figure 2. with the command STEM C1.7 49 19 5 Figure 2. . thc data arc now 39. HI 225. This means STEMANDLEAF DISPLAY LEAF DIGIT UNIT = 1 . We begin by rounding the price of each house down to the number of thousands by trimming off the last three digits. 275.
SPLI’ITING STEMS 21 that there tend to be more intervals in the stemandleaf diagram than in the histogram. (11) indicates that the stem has eleven leaves.9. which may. 7s contains 76 and STEMANDLEAF DISPLAY LEAF DIGIT UNIT = 1. 7* contains observations 70 and 71 .I . Stemandlcaf display for table 2. the diagram may not have room for them. be a good idea. Figure 2. By the time we have reached the end of the sixth stem. with a maximum at about $71 . This is a matter of sales tcchnique$W. there are 11 observations to go.OOO..950 is less alarming to a potential customer than $101. 8* 7* Figure 2. If there are too many leaves on a stem. There is only one house in the $lOo. the stem and leaf draws attention to the fact that the curve may have two maxima. LO 60 2 3 5 9 30 (14) 36 20 11 2 1 6* 6T 6F 6s 6. 1 3 45 6667 888888888889999999999 00000000011111 2222222333333333 444445555 666667777 8 0 7T 7F 7s 7. which makes it hard to determine the general shape of the distribution of the observations. The first column of the figure gives running totals of the number o f observations from the ends. the stemandleaf diagram may have too many intervals.1. 7T contains 72 and 73. .1. we have counted 21 observations from the low end of the data.1. Each yield is rounded down to an integer by trimming the last digit. or may not. The seventh stem contains the median.ooO range. In this example. This makes it easier to compute the median.1 illustrates this for the chemical data in table 2. At the end of stem 12 (the beginning of stem 13). Once we have passed the median.OOOO 1 2 REPRESENTS 12. Note that stems 16 and 18 have no leaves. 2 9 SPLITTING STEMS .9.OOO to $1 10.0o0.1. 7F contains 74 and 75. The seventh stem has been split into five shorter stems. If there are not many data points. the histogram suggests a smooth curve. Tukey has a method for splitting stems into five smaller stems. we start to count from the high end of the data.
BOXANDWHISKER PLOTS Another of Tukey’s graphical representations is the boxandwhisker plot. Thc box is the rectangle in the middle of the diagram.2.925 . with 8 and 9. and. Q. + * 64 I hn + c I 12 Yicld in pcrccnt 76 I 10 1 Figure 2.1. Notice Tukey’s choice of symbols for the last digit of the integer: * with 0 and 1.925.05 and 74. The Hspread is 74. T with 2 and 3. which is shown in figure 2.) are called the upper and lower fences. Q. lies bcyond the fence. I . .7.175 7.22 DESCRIPTIVE STATISTICS 77.6. the quartiles.7.0. and (3. 7.5( Q3 . the median is 71.68. I SO + I00 I I so ?(HI * 250 ** I’riccs in thousmids of dollars Figure 2.1. which is 80.5(Q3 .10. The two vertical ends of the box are the quartiles. In this example.2.7.925 = 5. This plot emphasizes the spread of the data.1.10.25. .10. Boxandwhisker plot of chemical yiclds in table 2. The whiskers protrude in either direction until they reach the next observcd value before the fence. which Tukey calls hinges. the righthand whisker goes as far as the largest observation. the three outliers are all on the high side. respectively. Thc lefthand whisker goes as far a s the sccond lowest observation. or hinges. 60. F with 4 and 5. 2. and the high and low values. It shows the median. Boxandwhiskcr plot for houses data in tabic 2.875 = 82. or Hspread.175.Q .1. = 68.175 .1.Q. The median is denoted by +. = 74. it is classified as a low value and marked with an asterisk.35.. finally. ) and Q. 60. which is considered to be unusually low by a criterion described in the next paragraph. is called thc intcrquartilc range. and so the fences arc 68. The asterisk at the far left denotes the lowest value. (1? Q . + 1. contains 78 and 79.10.8. S with 6 and 7. The horizontal lines coming out from either side of the box are called whiskers. . Recall that for this data set. The boxandwhisker plot o f the chemical data is shown in figurc 2. One observation.875 = 61.05.10. In thc box plot of the houses data. 61.
0 09. Multiple boxandwhisker plots.240. =$62.$90.10.840 . = $123.925.4. Q3 80 76 72 1.000 + $90.900. stops at $39. % w I 5 a 68 . I I 60  0 1 / 1 1 2 l i i ! I 4 I i Row I 6 l l : ! 8 10 Figure 2. and the fences are at $62.M0.400 and $123. the left whisker.000. therefore. The hinges are Q.3. $195. The lower fence is negative and we have no observations below it.240 = $213.10.BOXANDWHISKER PLOTS 23 I ** 8 I t I + I I 63. . leaving the three highest values marked by asterisks. Multiple boxandwhisker plots f o r thrce samples.000.0 72.0 7s. We recall that the median is $79.o Figure 2.240 = $27.0 + I I CI 66. The right whisker stops at the fourth highest observation.
PARETO DIAGRAMS One of the main purposes of a quality assurance program is to reduce the cost of poor quality. In most processes.1 gives a set of data from a group of 97 nonconforming items from a production run.1.4 goes a step further and shows. thcre are several types of defect that could cause an item to be rejected as nonconforming. Table 2. I.11. An obvious step is to reduce the number of items produced that do not conform to specifications.3 shows such a multiple plot for rows 1. Analysis of Defects Type of Defect liisulating varnish Loose leads Solder joint A Solder joint B Rcsistor 1 Resistor 2 Capacitor Number Reported 54 39 20 9 Percent 40 20 15 7 5 4 Cumulative Percent 40 68 83 5 2 7 95 90 1 100 99 .1 1. as manifested by distance between the hinges.1.10. Figure 2. Juran. is much less than the variance for the other two samples. M. Which should we work on first? J. one of the pioneers of quality assurance. boxandwhisker plots for each of the eight rows considered as separate samples. We are not going to be able to eliminate every type of defect at once. and 8 of the data in table 2.” and introduced Pareto analysis. argued that only a few types of defect account for most of the nonconforming items. One should be careful not to read too much into the appearance that the variance for sample 3. multiple boxandwhisker plots may raise more questions than they can answer.10. With small samples. We inspect some of the items produced and set aside those that are nonconforming or defective. in a figure. Figure 2. He called this phenomenon “the rule of the vital few and the trivial many. which he named in honor of the Italian economist and sociologist Alfred0 Pareto.1 1.24 DESCRIPTIVE STATISTICS It is sometimes convenient to make box and whisker plots of several samples on the same scale. 3. 2. Remember that there are only ten observations in each of these samples. considered as three separate samples. Note that there Table 2. The essence of the analysis is to rccord the number of each type of defect found in the nonconforming items that were rejected and then to make a chart rather like a histogram that will emphasize clearly the important types.
and thirteen. The corresponding chart is figure 2. such as percent conversion or housing prices.8 81.11. SUMMARY In this chapter. followed by xy plots for bivariate data and residual plots in the two chapters on regression.SUMMARY 25 136 108. dotplots. In later chapters. and the solder on joint A between thcm account for over 80% of the defects. Pareto chart. The sets consisted of observations on a single variable. With a few easy commands on a personal computer. loose leads. 2. The table shows that insulating varnish. it took an hour or more to calculate the mean and variance and to . ten. each of these plots gives the viewer a simple visual summary of the data that provides a picture of “how the data looks. allowing for false starts and interruptions along the way. and boxandwhisker plots.4 27. Control charts appear in chapters nine.2 0 CL1 CL2 CL3 re CL4 CL5 CL6 CL7 2. we have discussed some methods of presenting data sets graphically.1. stemandleaf plots. are more than 97 defects recorded because some items had more than one.11. we mention other types of plots.12. The methods used in the main part of the chapter were histograms.” Not many years ago.6 L 2 0 54.1.
2.2 13.6 13.9 17.7 14.0 14.6 14.6 15.7 16.0 15.3 16. 15.8 16.8 13. median.0 14.3 13.6 15.8 13. and standard dcviation if you rounded off each observation to the ncarcst integer.5 14.7 17.9 15.5 13.2 15.8 15.4 16. Verify that you would gct thc same results in exercise 2. for convenience. Always have a good look at your data set before you try to do anything complicated with it! EXERCISES 2.6 17.4 15.5 13.8 16.5 15. and cven longer to make a histogram. median.3 18.8 16.2 14. 2. 2.4 14. The reported thicknesses.2 .5 16.9 12.4 19.3 14. During World War 11.1 15. thin slices of mica were used as dielectric material i n condensers for radios. there is n o C X C U S ~ not to make thcm a routine first step in any analysis.6 17. 2.6 13.4 14.4 13.6 13.0 14.3 13.0 from each observation before making your calculations (and added it back to obtain the mcan and median).2 16.8 13.7 15. you subtracted 10. Verify that in exercise 2.1 16.2 16. Perhaps that was an excuse for a busy engineer not taking the timc to do it.6 13.0 15. are 1831 1839 1828 1800 1831 1850 1813 1817 1816 1832 1781 1832 1874 1578 1876 1821 1847 1823 1848 .4 13.6 15.9 15.0 14.1.1 if.8 16.7 16.6 15.2 17.6 15.1 15.3 16.0 13.0 12.2 10.26 DESCRIPTIVE STATISTICS locate the maximum and minimum and potential outliers in a large set of data. Make a histogram of the data and calculate the mcan.4 14.4 15.5 11.9 15.1 14.3 15.5 12.8 12. and standard deviation.0 13.1 13.2 16.2 14.7 15.3.0 18.5 15.8 14. An engineer takes a sample of 19 wafers and measures the oxide thickness.5 11.5 14.0 12.1 12.3 18.7 17.1 15.7 18.4 14.8 17.4.5 16.6 16.9 15.2 14. Now that these simple graphics take only a few minutes at the terminal that is conveniently situated on one’s desk.4 14. The following data set is made up of measurements of the thickness of 100 strips of mica in onethousandths o f an inch.5 15.1 15. in angstroms.1. there would be little changc in the values of the mean.5 14.
the trimmcd mean. Does this set of data approximate a bellshaped curve? The observations are the rcciprocals of the observations in exercisc 2. the median is unchanged and the trimmed mean and quartiles increase slightly. 2.4 2. standard deviation.9. Makc a histogram of the following set of 50 observations and use it to find the median and the quartiles. D o that and recalculate the mean.8. and the quartiles of the data. median. Make a stemandleaf diagram for the 18 observations in exercise 2.4. Make a dotplot and a boxandwhisker plot.6. (after 1578 has been excluded) and use it to find the median and the quartiles. Docs the data set seem to have the traditional bell curve (the normal distribution)? 180 215 179 193 247 222 189 262 202 211 183 219 193 188 187 182 184 171 201 185 160 207 208 249 231 20Y 241 185 243 189 285 258 264 223 175 218 214 183 211 257 228 231 187 219 204 155 199 202 189 205 2. 56 47 56 52 40 45 53 38 50 47 46 52 53 53 55 55 54 63 58 54 5U 48 48 40 43 48 41 54 41 53 35 39 38 45 57 46 47 55 47 39 44 43 53 46 49 65 50 SO 53 40 2. and the quartiles for the remaining 18 wafers.7. On the other hand. 2. trimmed mean. . m' k e a d new dotplot and a new boxandwhisker plot.EXERClSES 27 Obtain the mean.7.7 multiplied by lo4 and rounded. thc engineer decides. 2. that 1578 is an invalid observation and that he is justified in throwing it out. standard deviation. Find the hinges and fences. In both plots. Explain why the maximum observation is flagged as a possible outlier. the observation 1578 is obviously an outlier. Make a boxandwhisker plot of the data in exercise 2.10. there arc appreciable changes in both the mean and the standard deviation. Make a boxandwhisker plot o f the data in exercise 2. in exercise 2. Make a histogram of the following sct o f observations and use it to find the median and the mean.why is it n o t starred as an outlier in the boxandwhisker plot? Note that when we drop the outlier. but the niinimum observation is not.9.5. The observation 1781 is the only observation below 1800. mcdian. Suppose that after investigation.
Sixty of the components are put on test and their lifetimes arc given below. with an average of 20 defects per machine A2 F 67 B5 G 15 c 54 H 13 D 12 I 10 E 8 K 14 Make a Pareto diagram that provides a graphical illustration o f the importance of focusing on machines C and F. An electric component is spccified to have an average life of looc) hours. A plant has ten machines working a single shift. together with the standard deviation and the quartiles. In a week.28 DESCRIPTIVE STATlSllCS 2. Make a histogram of the data.11. 15 213 434 722 1108 1576 2217 3159 72 241 452 723 1171 1601 2252 3215 74 270 464 877 1316 1631 2316 3832 99 285 504 959 1328 1657 2684 4295 100 302 526 1024 1410 1893 2858 127 361 652 1051 1451 1975 2862 I95 366 666 1075 1463 2059 2873 195 430 704 1087 1529 2076 3088 2. Find thc mean and mcdian lifetimes of the components tested. This sounds encouraging. the following numbers of defects were noted for the ten machines. The average exceeds the specification by 27%. .12. and 25% of the components lasted more than twice as long as specified. but what fraction of the data fails to meet the specification? This suggests that it is not enough just to look at the sample average.
29 . An inspector takes a sample of rods from the production line and measures the lengths. the last digit but one. Hence. M. The inspector can record the actual length.2. I. . In the jargon of quality control. and then 1 took the third digit after the prefix in each number. Y is a random variable.1 shows 320 digits taken from the Austin phone book. i. The reason that I chose the third digit rather than the last was that I was not sure whether the phone company held back some numbers ending in zero for test purposes. I took the righthand column from several pages (for convenience). as opposed to values over a continuous range. John Copyright 0 1990 by John Wiley & Sons.e. acceptable or not acceptable. A simpler alternative is sampling by attributes. . Because the rods will not have exactly the same length. Inc CHAPTERTHREE Discrete Variables. in which case it is said to be a continuous random variable. X. 3. of each rod. Y. . perhaps. 2 . It takes only integer values 0. Thus. either too long or too short. INTRODUCTION Suppose that we have a process for manufacturing metal rods and that we are concerned about the uniformity of the rod lengths. It may. the inspector is sampling by variables. Y will not take the same value every time. A rod may be usable if its length lies in the interval 109 mm 5 Y I I 1 mm.. We now develop some theoretical background for discretc variables. . X is said to be a discrete random variable. The inspection can now proceed in two ways.Statistical Methods in Engineering and Quality Assurance Peter W. but unusable if its length is 1 outside that interval. in the sample. take values anywhere in a certain range. The inspector may record the number of defective rods. RANDOM DIGITS Table 3. Continuous variables are discussed in the next chapter. the inspector may classify each rod as either good or bad. or Attributes 3.2. 1. The customary ternis are defective or nondefective.
.1. in the sense that each of the ten values is equally likely to occur. X takes the values 0. . . as a consequence of this probability law.3. Table 3. Let X denote the value of a digit.1. or law.eros ~ 1620 6773 4114 2269 7510 4305 1516 8796 3076 676 1 4255 7063 9485 1573 1589 4879 2476 6699 5270 3992 6298 6370 8966 5308 6136 4241 9104 2140 8518 6807 6089 0325 474 1 0722 2024 3073 7673 3733 6653 8068 8350 1449 5518 7781 5272 1663 3902 3472 7957 0229 2440 1042 2953 8593 1085 9706 0480 5054 0546 9131 5659 0158 0863 2118 1057 8404 0037 8534 8008 2191 1938 8227 1301 5198 4590 6081 5253 5343 8764 1013 It is reasonable to think that the digits ought to be evenly distributed. 9 has probability 0.30 DISCKETE VARIABLES. more generally. A plot of them appears as Table 3. or P(O).1 The probability distribution of the random variable X in this example is callcd the multinomial distribution. In this example. the probability law is simple.1 shows the actual frequencies of occurrence of thc digits. for thc probability that X takes the value i. That is clearly not so.3. onetenth must be 1. . It is a generalization of the binomial distribution that is discussed later. We write P ( X = O). Frequencies of Occurrenms of Digits 0 40 1 36 2 29 3 33 4 29 5 34 6 32 7 30 8 31 9 26 . . OR A'ITRIBUTES Table 3. The probability model for the process is given by P ( 0 ) = P(1) = P ( 2 ) = . Some go further and call it a uniform discrete distribution bccausc all the probabilities are equal. . for the probability that X takes the value zero. or P(i). 320 Digits Taken from the Phone Hook Nun1ber of 7. We do not claim that. it is. A MATHEMATICAL MODEL Suppose that we regard this data as a sample of the output from a process for producing digits. and. and so on.2. I. called a random voriuble.1. 3. 1. exactly onetenth of the digits in our sample must be 0. . This implies that cach value 0. . therefore. P ( X = i).3. = P ( 9 ) = 0. . Wc may also say that X has a particular probability distribution. 9 according to a probability model. 2. . another tenth must be 2.
On the other hand.1. In section 7. I I I I I I I I I I I 10 I I 0 0 1 2 3 4 5 7 8 9 Figure 3. I I I I I I B ti u. SOME BASIC PROBABILITY THEORY We can also look at this from the point o f view of an experiment. It occurs more often in the first row than in any of the others. Out o f thc total of 320 digits in the sample. Is this unreasonable? The answer to that question can be shown to be no. each digit should have occurrred about 32 times. The deviations of the frequencies from 32 that we see in this particular set of data are not unreasonable. In only one o f the rows are there fewer than four zeros. figure 3.3. and yet 0 occurred 40 times and 9 appeared only 26 times. the set of possible results of the experiment i s .15.3. We should also expect that out o f the 40 digits in each row. we present a statistical test for the hypothesis that this set of data did indeed come from a process with equal probabilities. I I .4. Is that unrcasonable? 3.1. It has ten possible results. tossing a die can be regarded as an experiment with six outcomes. or outcomes. I I 8 1 I I I I I I I I 30 I 1 I I I I I I L I I I I I I I I 2 .ITY THEORY 31 40 I I I . the number 0 should occur four times. In the general case. Similarly.SOME BASIC PRORABI1. 2 20 + I I I . . The experiment consists of picking a digit. Frequencies of digits. we expect that each digit should appear about onetenth of the time.
Suppose that E. 7. occurs. 30% have an output defect. We are given that P ( E . ) = 4/52. be the cvent that a part has the output defect. The intersection..P ( E ) . it occurs only if both El and E. not necessarily in equal portions. or product. f E 2 ) l (3. which consists of the points 0. is a set of points in S. Somc might get no jam at all. It follows that f ( E c ) = 1 . 1 . it occurs if either E.. we can consider an ordinary deck of playing cards. more precisely. a probability measure. be thc event of drawing a spade.. For example. E. 20% have an input defect. An cvent is said to occur if the result. An example in which S has an infinitc number of points appears in section 3.1) Exumpfe 3. It contains six points. U E . of the experiment is one of the points in the set.2. It is the set of all points that are in either E. is the set (1.e.6 = 1 . or. of E l and E.7. 5 .4. These probabilities satisfy two requirements. or E. OR ATTRIBUTES called the sample space and is denoted by S.16. ) = 0. ) + P ( E 2 ). 3) with probability 0. 3. As another example. 1. ) = 13/52 and P ( E . For a discrete variable.f ( E . n E. In computing the probability of El U E. It is the set o points that are in both E . occur. The ace of spades is thc only card in E. In this case. is assigned to each point of S. is written as E. Of the parts offered by a supplier. is the event that X is odd. The following formula expresses this: P ( E . be drawing an ace. and f E. One can picture a 1pound jar of probability jam being ladled out. An event. In this example. ) = 16/ 52. We can now define unions and intersections of events. What is the probability that a part will have no defects? Let E l be the cvent that a part has the input defect. we must be careful not to count twice points 1 and 3. 9) with probability 0.. not (13 + 4)/52. Its probability is 0. or E. or outcome. X = 1 or 3 or 5 or 7 or 9. The latter answer is wrong because the ace of spades has been counted twice.1.. or both. which occur in both sets. among the various points.P ( E ) . either an input defect or an output defect. 2. but P ( E . The union of E . E l ilE. They are nonnegative and their sum is unity. and let E.. The probability that the event occurs is the sum of the probabilities of the points in the set. is written E l U E. A probability. Thc event E' is the complcmcnt of E . E' is the event X > 3 . the event El might be the event X 5 3. It consists of 52 cards. U E .32 DISCRETE VARIABLES.4. 2. A simple part in an electronic system can have two kinds of dcfcct. The numbcr of points in S may be either finite or infinite. i. ) = P ( X 5 3) = P ( 0 ) + P(1) + P(2) + P(3) = 0. and 3.4. P ( E . in the digit cxamplc.20. P ( E . Thus. and E. and E. and 10% have both defects. r l E2. Let E. ) = P ( E . Y(E l f l E 2 ) = 1/52. S is a set of discrete points. of which 4 are aces and 13 arc spades. El U E2 is the set (0. . it is the set o f points of S that are not included in E and corresponds to the event that E does nof occur. those points have zero probability.
8.6. 6. 2. Let X take the value 1 if the digit is even. X = 1 if the coin shows a hcad. 2. Y(E. 6.00 0. In both cxamples. 320 Digits Taken in Groups of 4 (Number of Even Digits) Number of ‘rwos 3 2 1 I 2 2 2 0 2 2 3 2 3 1 3 2 2 4 1 2 1 1 2 2 3 1 3 3 0 4 1 2 1 2 4 1 1 2 1 2 1 4 2 2 2 2 0 2 3 3 2 2 2 1 2 3 3 1 0 4 2 1 2 2 0 2 2 4 4 3 1 2 3 2 1 3 1 3 1 1 6 8 7 11 .0 .2.5. 7. the behavior of X follows the probability law P ( 0 ) =.5 . X . THE BINOMIAL DISTRIBUTION The random variable. The probability that a part will have at least one defect is.5. 5. This time the digits have been classified as odd (1. The mathematical model for drawing a digit at random from that set of possible values is the same as the model for tossing a fair coin.5. and the value 0 if the digit is odd.5. the same as the probability that the digit will be odd. P(1) = 0. ) = 0. 0 3. n E . and let X denote the number of heads that appear. from equation 3. and X = 0 if the coin shows a tail. 7.4.10. This is equivalent t o observing the number of Table 3. toss a fair coin once.1. This i the simple coin tossing model. The probability that a digit from the set 0. 4.1 is thc number of even digits in each set of four. ) = 0. 4. The first entry in the table is 3. s 3.THE BINOMIAL DISTRIBUTION 33 P ( E .60. 3. that is observed in table 3.40 = 0. 9 will be even is onehalf. 5.1. 9) or even (0.5. It follows that the probability that a part will be free of defects is 1. ) = 0 . which contains three even numbers.1 is based on the data in Table 3. 3.30. 3 0 . sincc the first set of four digits in Table 3.1.u E . 1 0 = 0 .1 is 1620.1 is the number of even digits in a group of 4. 2 0 + 0 . 4 0 . 1. The rundom variable recorded in Table 3. THE COIN TOSSING MODEL Table 3. and P ( E . Alternatively. 8).2.
2. TITr Figure 3. TTTH. The numbers of times that these five values occurred arc shown in table 3.6. THTT. IIIITH. but not with equal probabilities. 1.1.. TI'HT.6. HHHT. 4.. S.( f >< $)< ) = . The probability. They are equally likely. 3. X takcs the five values 0. Frcqucncies o f values of X . HTHH. 5 1 21 20 2 32 30 15 3 4 5 20 7 hcads that occur when a coin is tossed four times. We multiply the probabilities of tails on the individual tosses. has 16 points. HTTH. HHTT.1.6. THTH.34 X= Observed frequency We expect 0 5 DISCRETE VARIABLES. each has probability 1/16. = P(4) The sample space. that X takes the value 0 is the same as the chance of tossing four tails in succession. . HTHT.1 and in figure 3. P(O). THHT. OR ATTRIBUTES Table 3 6 1 Frequencies of Occurrence of Values of X . I T H H . The points are HHHH. HTTT THHH. P(0) = ( t.
4/ 16. THTT. Onc might ask the following question. respectively. HTHT.’’ in which each trial has exactly two outcomes. rcplacc all the nonzero digits by ones. and 90% that are nondefective. What are the probabilities that X = 0. 3 7 UNEQUAL PROBABILITIES . The binomial distribution was introduced in the previous section from the coin tossing model. exactly two heads occur at six points: HHIT. A fair coin is tossed five times. thc statistician thinks of a process. called success or failure. HTHH. Again.2.. cl Hcnce. Now we have a sample of data from a process that produces 10% zeros and 90% oncs. and so P(2) = 6 / 16. HTTH. If the data were given in the form of a sequence o zeros and ones. The random variable X is said to have a binomial distribution. 4/16. 0.1. which takes thc . corresponding to the values taken by X with probabilities 1/16. X = 3. 1. Finally. X = 0 occurs only at the point 1TITT. THTH. HHTH. Thus.UNEQUAI. ITI’H. Exactly one head in four tosses occurs at each of the following four points: H’I?T. how would you decide whether the process is f operating under control at a 10% defective rate? In the general case.6. 1 / 16. or a “sequence of trials. THHT. 4. This time. P ( 0 ) = 1 /32 = P ( S ) . and X = S? There are 2’ = 32 points in the sample space. This topic is developed further in the next two sections.5. Example 3. 3. THHH. Thc probability of a head at each toss is 0. 2. PRORARILITIES 35 X = 0 for the point T I T T . we use thc data in table 3. Associated with each trial is a random variable X. while X = 3 at MHHT.1. X = 3 requires three heads and two tails. F(1)= P ( 3 ) = 4 / l h . We can think of it as represcnting the output from a production process that produces 10% defective parts. we should expect X = 2 to occur (6/16)(80) = 30 times. That is a special case of a more general model. lTHH. An alternative version of the sample space would be the five points. In a set of 80 observations. TTHT. represented by zeros. Y ( 3 ) = 10/32. 6/ 16. three places out of five for the heads can be found in 5!/(3!2!) = 10 ways.
This would lead one to conclude that p is considerably less than 0. In the coin tossing model with “heads” as the succcssful outcome.90. If we were to classify a defective as a failure. since the probability of one head and nine tails in ten throws is greater than the probability of ten tails.90.1 an example of a binomial variable with n = 4 and p = 0. That argument takes us into thc area of hypothesis testing and decision making. The probability of failure at each trial is 1 .5 and to bet on tails for the tenth toss.e. It is not reasonable to bet in favor of heads on the tenth throw.8. If we toss a fair coin nine times and observe tails each time. We saw in table 3. Indeed.5. and so P ( 1 ) = 4/16. p = q = 0. The idea that the trials are independent with constant probability of succcss is important to the argument. the example in the first paragraph of this section would have p = 0. they would have p = 0. which is often denoted by q . we multiplied p 2 q z by 6. and so P(1) = 4pq3 . They apply this model with defectives classified 21ssuccesses. each of which has the letter H once and the letter T three times. 3.. the probability of a head on the tenth throw is still 0. To compute P ( 2 ) . We have described what is known to statisticians as a sequence of Bernoulli trials. the number of ways of choosing . As a practical matter. the chance of a head on the tenth is greater than a half. Thc system is postulated to have no memory. and with the probability of any part bcing defective denoted by p.5. quality control engineers are more interested in focusing upon the number of defectives produced by a process than on the number of nondcfectivcs. which is the topic of chapter seven. and let X denote the total number of successes in all n trials.20 in the example and q = 0. and q = 0.5.5 is false. a case could be made for the argument that a sequence of nine tails in succession is strong evidence that the original assumption of p = 0. It would not be reasonable to argue that. By generalizing this for p # 0. X is said to have a binomial distribution with parameters n and p. named after the Swiss mathematician James Bernoulli (16541705). for example. that X = 1 occurs for four sequences. OR ATTRIBU’rES value one if the trial is a success and the value zero if the trial is ii failure. The probability of success is thc same for each trial and is denoted by p.10. We argued. the probability of a sequence with one head and threc tails is pq3.5. i. Each sequence has probability 1/16.36 DISCRETE VARIABLES.p . Thus.50. THE BINOMIAL DISTRIBUTION (GENERAL CASE) Suppose that we have a scries of n independent Bernoulli trials with probability p of success at each trial. which is the number of sequences with H twice and T twice.
it is defined by the center of gravity of the probability distribution.MATHEMATICAL EXPECTATION 37 two places out of four for the successes. we should not be surprised to see a few more or a few less. the number of sequences available is n! q n . In practicc. The Poisson distribution is discussed in sections 3. 6. For a discrete variable. it is easier. Consider throwing a single die.9. let X . Let X be the number that shows on the top face.x ) ! and the probability that X = x is P(x) = n !p"q"" x ! ( n. then the average number of p i n t s pcr throw would he approximately 3.5. and adequate. 2. the values of I I may be rather large. each with probability 1/6.weighted according to the probability law. 5 . 3. this is the number of combinations of two objects out of four. X takes the values 1.1) where the summation is made over all values x that X takes. 3. we should expect to have 32 zeros. denote the number o f succcsscs on the ith trial: (3. Mathematically. It is denoted by E ( X ) .t f 2 + : =3.x ) = x ! ( n .x ) ! In quality control work. to use the Poisson distribution as an approximation to the binomial.173. We do state that if we were to throw a fair die for hours on end. Then E(X)= x +: + 2 + . If we wish to have x successes in n trials. and it is hoped that the values of p are small. or expected value. The mathematical expectation. We do not believe that we shall ever see the die showing three and onehalf points. Under those circumstances.9. In a series of Bernoulli trials.19. In the language of permutations and combinations.9 2 ) . the formula for E ( X ) is E(X)= c x&) 9 (3.5. 4 . MATHEMATICAL EXPECTATION We said earlier that out of 320 random digits. of a random variable X is the longterm average of X.
2) For a single toss of a coin.. respectively. E ( X ) = E ( X .p 2 = p( 1 .9.Y. E(X)= p whence 9 V ( X >= p .E(X)I’} = E ( X 2 ). Definirion.38 DlSCRErE VARLABLES. (3.5) THE VARIANCE We arc now able to give a more formal definition of the variance of a random variable.E ( X ) I L }.1) E { ( X .. . then E ( Y ) = a. . and if Y = a .p ) = p q . X . .. If X has been measured in yards and if Y is the corresponding measurement in feet. .* .m. q=p . 3. ) + E ( X . The following properties of expectations are easily verified. If X.2 E ( X ) E ( X ) + [ E ( X > ] * = E ( X ’ ) . . If a is some constant.10. + a2m2I. .X. . (3 . 4) This is merely a matter of change of scale. E ( X 2 )= 1 2 ’ p and +o’. . OK AITRIBUTES ri X . .10.* = C a. .3) E ( u X )= u E ( X ) . This definition is compatible with the definition that was given in chapter two. and so E(Y)=3E(X). X is the sum of independent variables X.. (3.m. we note that 2 (r3= E { [ X .9. . (3. = np .10. and so For a binomial distribution. . The variance of a random variable X is V(X) In practice. are random variables with expectations m .[ E(X)]’ .. then Y = 3 X . ) 3. . + u2X2+ * *.. .m. (3.
. p . then Applying equation (3. we should have p = 2 / 5 . .3) and if X . we should have. we can develop properties of the variance that parallel equations (3. If the first item chosen were defective. If a and b are constants. or batch. However. *. A. the denominator is N . X . for example. ’. SAMPLING WITHOUT REPLACEMENT In acceptance sampling by attributes. which is independent of the parts that have been sampled so far. of which three are defective. and if Y = a . for the second item.X. we have sampling with repluc*ement. . stretching a piccc of yarn until it breaks in ordcr to determine its strength.4) and (3. sampling with replacement is impossible. rn2. X . In computing p for thc second item. .95). * . . + a. The alternative is sampling wirhout rrplucernent.10. Suppose that a batch consists of only six items.9.1 1. This was the basis for deriving the binomial distribution. . the probability of drawing a defective item remains constant because the composition of the batch from which the inspector is choosing is unchanged. + * . of defectives in the sample. However. as.10. of which only two are defective. = npq * II (3.4). for the second itcm. where N is the size of the . then V ( X ) = V ( X . When the first item in the sample is chosen. there would now be five items left in the batch. and the next item is drawn from the remainder of the batch.5) 3. V(uX + b ) = U’V(X) . . (3. . p = 315. )+ V ( X . If the first item were nondefective. 1. p = 3/6. the sample can be drawn in two ways. Clearly. The inspector sets each item aside aftcr it has been tested. an inspector draws ii sample of n items from a lot. We have argued until now that the probability of drawing a defective part at any stage of the sampling is a constant. of N items and observes the number. respectively. If the inspector returns each item to the batch aftcr examination and then draws the next item from all N members of the batch as before. With this scheme. . are independent random variables with expcctations m . it is quite possible for the same itcm to be counted more than once before the inspector has reached the quota.SAMPLING WITHOUT REPLACEMENT 39 For independent variables. we sec that if X is a binomial variable with trials and probability p .10. ) + . if the testing of the items is destructive.
Of these 20 ways. The correct model would not be the binomial model. Returning to the batch of six items. 3 X 3 = 9 give us the sample that we seek. The sample of n items contains x out of d defectives and n .333363. n . we can choose a sample of three out of six in 20 ways.40 DISCRETE VARIABLES. Let x be the number of defective items in the sample. In the general case. n .001/44. THE HYPERGEOMETRIC DISTRIBUTION Example 3.OOO arc women. of which three are defective. both of which are close enough to 1 / 3 for practical purposes. What is the probability that it will contain two defectives and one nondcfective? We can choose two defectives out of three in three ways. and the number of defectives are not very small. Altogether.d . suppose that a large university has 45. x ) ways. We wish to find the probability law. so that P(x) = C ( d . x)C(N. When the batch size. That model is called the hypergeometric law. and that we draw a sample of n items. The probability that a student chosen at random will be female is 113. the binomial model is still a good enough approximation to the real world to justify its use. suppose that a random sample of three items is drawn. hence. . 3. or 15.x ) ways.000 are men and 15. On the other hand. and cl so the desired probability is 9/20. for P(x). but a model based on the change in batch size and content.d are nondefective.d in C(N . n ) (3.999 = 0. It can be shown that the sum of these probabilities for the various values of X is indeed unity and that E ( X ) = n d / N .1) The algebraic expression in the numerator resembles an expression that appears in classical mathematics in a term in the hypergeometric wries.X ) C".1.x of N .3333185. the name hypergeometric for this distribution. the number of combinations of three out of six. We can choose x defectives out of d in C ( d . we can choose one nondefective out of three in three ways. When a second student is chosen. OR A'lTRIBUTES original sample.12. n ) ways.000 students. and a sample of n items out of N in C ( N .x nondefectives out of N .999/ 44.999 = 0. which is the number of combinations of two items out of three. even if we sample without replacement. of whom 30. the numerator depends on the result of the first drawing. suppose that we have N items of which d are defective and N . N. or distribution. Similarly. That is a simple example of the hypergeometric distribution. the probability of selecting a woman will be either 14.d nondefectives.12. or the hypergeometric distribution. n .d .12.
and P ( A l B ) = 215. We see from the data that P ( A ) = 0.048. given that event A has occurred. In 8% of the components. the ratio has no meaning if P ( A ) = 0. In 40% of the components. The coniponents each contain two resistors. = Substituting the values given above. This is an example o f thc basic law of conditional probability: P ( A B ) = P(A)Y(HIA)= P ( B ) P ( A [ B ) . in this case.1.12 The two events are not independent.1) P ( B ) = (. P(B1A') = 3 / S . i.13. Exurnple 3. both resistors are defective. An inspector examines a component and finds that the first resistor is defective.08. The notation P ( B I A ) means the conditional probability of event H . If the first item is defective.13. we calculate (3.0.2) Mathematically. Hence. either it is preceded by A or else it is preceded by A'.08 = 0.12.40) =0.) + = j Let A B denote the joint event that both items are defectivc (we drop the symbol n). and B is the event that the second item is defective. (3. Similarly.CONDITIONAL PROBARILlTIES 41 3. then P ( A I B ) = P ( A ) and P ( A B ) = P ( A ) P ( B ) . CONDITIONAL PROBABILlTIES Suppose that in the previous problem. A supplier is manufacturing electronic components. the first resistor is defcctive. the event that A did not occur.12)(0.13. and P ( A H ) = 0.13. We write this as P(BIA) = 2 / 5 . .. Y ( B ) P ( A ) P ( B / A )f P ( A ' ) f ( B I A ' ) . we say that the events A and B are (statistically) independent. Event B can occur in two distinct ways. thc probability that the second item is also defective is 2 / S . A denotes the event that the first item drawn is defective. We see that P ( A B ) = 1 / S = P(A)P(HIA). P(A) 0. Let A' denote the complement of A .)( i) (i)(:.67 . the second resistor is defective. P(BIA) = I _ _ P(AH) . They would have been independent 0 only if we had P ( A B ) = (0. A" is the event that the first item is nondefcctive. If P(B1A) = P ( B ) . P(A1B') = 315. In 12% of the components.In that case. P ( B ) = 0.40. we define P ( B ) A ) as the ratio P ( A B ) I P ( A ) . What is the probability that the second resistor will be defective also? Let A be the event that the first resistor is defcctive and H thc event that the second is defective.e.
On the other hand.14. if the first cxperiment has a sct o f outcomes A . What is the probability that A has occurred also? We wish to calculate the conditional probability P ( A 1 B ) .42 3. P(Al23) is thc postcrior probability of A .2) In equation (3. You purchase most of them from a supplier in Dallas who is on record as producing only 5% defectives. or has not. . and so you havc to buy the remaining 20% from another supplier in Houston who makes 10% defectives. It is the probability of the occurrencc of A before we know whether or not R has occurred. From equation (3..1..13. It is defective.13. but you cannot read the scrial number. after his death. .14. What is the probability that it came from the supplier in Dallas? Let A denote thc event. Let R be the event that the component is defective. Suppose that there are two experiments.1).14. Your company uses a large number of a certain electronic component. and that probability would remain unchangcd. and Bayes' formula. (3. the knowledge that B has. and applying equation (3.2). The theorem seems at first to be just a routine exercise in manipulating formulas. If you did not . You pick a componcnt at random. Unfortunately.1 leads us in the direction of Bayes' theorem. Thomas Bayes.14. The first has outcomes A and A'. It was proved by the Rev. Its importance has been realized in the last half century. Example 3. and published in 1763. Wc notice that I3 has occurred.13. The second has outcomes B and B'. or outcome. he can only supply 80% of your need. occurred would provide no information about the occurrence of A . If A and B are independent events. It is our revised value of the probability of the occurrence of A in the light of the additional information that H has occurred. OR ATFRIUUTES Example 3. It is called by various authors Bayes' theorem. Bayes' rule.1). A L ? the theorem becomes .1) More generally. It enables us to make use of prior knowledge in our probabilistic calculations. (3. that the component came from Dallas. P(A) is the prior probability of A .14. BAYES' THEOREM DISCRETE VARIABLES.
999)(0.2. applying Rayes' theorem. If we assume that there are no false negatives. By equation (3. this means that the test may indicate.001)( 1 .000) = 0. Y(B1A)= 1. Applying Bayes' theorem. 0.001)( I . P(A1D) would fall to about 2%.OOo) = 0. P ( f 3 )= (0. Let B be the event that he tests positive. Example 3.67. (0.80)(0.005) + (0. They may give a small percentage of false positives. chosen at random from the population. If P ( A ) were O.10)= 0.001.1).00s995 . In medicine.80. This can lead to unfortunate consequences.13. Suppose that one young male in a thousand has a certain disease that has n o obvious unique symptoms.WS% (onehalf of onc percent) of false positives. P ( B ) = (0.(HK)l (one in ten thousand) and the rate o f false positives were the same. 0. By equation (3. Somc diagnostic tests are not perfect. P(B1A') = 0.l).O00. in some cases. that a patient has a disease when that is not so. the prior probability is P ( A ) = 0.20)(0. Another example occurs in medical statistics.BAYES' THEOREM 43 know whether the component was defective.14.13. in this example. It is important to noticc the assumption thiit the young inan was choscn at . and that the routinc tcst has probability U. P(AIH) = (0. A young man.005. takes the test and is rated positive. the knowledge that the component is defective reduces the odds that it came 0 from Dallas and increases the chance that it came from Houston.06.05) + (0.80)(0.05) Y ( A I B ) = = 0.1668. then. What is the probability thitt he actually has the disease'? Let A be the event that the patient has the disease.06 Because of the comparatively good quality record of the Dallas supplier. This is the overall rate of defectives for the supply of components entcring your plant.OOSY~S 0 Only onesixth of the men who test positive actually have the disease! P ( A l 8 ) i s strongly influenced by the rate o false positives and the f rareness of the disease. you would argue that P ( A ) = 0.
The table has four cells. pertains to length and X. X . Table 3.” We usually prefer to say that each has two levels. In that case.14. we use data from a similar example to illustrate some of the basic ideas of bivariatc distri hutions.1 Strength Length OK 30 I20 90 Defective 10 Total 1 00 OK Defective Total 20 30 50 150 . A sample of 150 rods gives the results that are shown in tablc 3. Each test classifies the rod as either OK or defective. It would not be so appropriate if he was chosen because he was already in a doctor’s office or a clinic as a consequence of feeling ill. P ( A ) might be larger. we saw an example in which a component could be classified on the one hand as having come from Dallas or Houston. In this section. These data show the output of a plant that makes rods and is not very good at it. IS.In our example. The probability that XI takes the value i and X. They take the values 0. 1: zero if the test result is OK. OR ATTRIBUTES random from the whole population. Thc ijth cell corresponds to X . whether or not he had the disease.ES.. 3.I . This is an example of a bivariatc distribution. The reader with experience of test procedures can think of similar examples in industrial work. The totals have been written at thc rightmost column of the table and along the bottom row“in the margins. it contains tz.15. pertains to strength.” The corresponding probabilities lire called the marginal probabilities.. BIVARIATE DISTRIBUTIONS An object can sometinies be classified in two ways. and one if the rod is defective. Each rod is subjected to two tests: length and strength. Every man in the population had the same chancc of being chosen. X. = j . observations. This might be an appropriate model for routine medical exams for all the potential inductees to the armed forces under conscription or of a large company that routinely gives medical examinations to all potential employees as a condition of employment.15. Now there are two variables: location and quality. The data have been modilicd to make the arithmetic simpler. There are N observations altogether. The data can be arranged in a 2 X 2 tablc: two rows and two columns.= i .44 DJSCRETE VARIABT. Each variable can take two “values. In section 3. Denote the two variables by XI and X. and on the other as being defective or nondcfectivc.
I for every i and every j . p. = "I. The conditional probabilities are defined in terms of the cell probabilities and the marginal probabilities: P ( X . X has a binomial distribution with parameters p and N. (3.. .2) It is sometimes appropriate to ask whether the two variables are independent. pI. = n . The mathematical argument is that we inspect N items and find X defectives.. = y (3. n. = yO/150. we should have expected to have had (2/3)(4/ 5)(150) = 80 parts passing both tests.15.. items produced by a manufacturing process. In this example. p. we have poo ..60 # ( $ ) (. PI PI. (3. in section 8. = jlx.12. = 301150. or defective.3) In the example. inspect them.. p.33 is the probability that a rod is defective in length./ is the total for the jth column. and I p.. and B be the event that it passes the strength test. pol = 10/150. The events are independent if P ( A B ) = Y ( A ) Y ( B ) . The corresponding marginal probabilities are .. p .. and pl. We discuss this type of problem in more detail in chapters six and seven. = i ) = .15.THE GEOMETRIC DlSTRlBUTlON 45 takes the value j is pl. = 0. is the total for the ith row of the table. We rephrase that and say that the variables are independent if. = p. and only if. and n 11. An obvious procedure is to take a batch of items.. 3. We return to this point when we discuss the chisquare test for 2 X 2 tables later. p.0. Are the two types o f defects independent or do they tend to go hand in hand? Let A be the event that a bar passes the length test.1) In the example. This suggests that there might be a tendency for parts that fail one test to fail the other also. / N .) . The marginal totals are denoted by ni. and then argue that the percentage of defectives in the batch is a good estimate o f the actual process percentage.16. its . With independence.15. THE GEOMETRIC DISTRIBUTION Suppose that we wish to obtain an estimate of the percentage of nonconforming. = 20/150..
17.I)p'q"' (3. L. THE POISSON DISTRIBUTION The last discrete distribution that we shall discuss is the Poisson distribution. and so P(x) = qx'p . 3 . Mathematically. Another approach to the problem would be to inspect items one at a time until we found ii defective. This aspect o f the Poisson distribution is one of the topics of the chapters on acceptance sampling (chapters 1I and In 1898. 3.' by 1 /X. named after the famous French mathematician D. . Let X be the number of items inspected (including the defective). It is known to some mathematicians as the Poisson limit law. It is featured in the c charts in chapter ten. . . Its sum is (3. and let X be the number of trials until the first succcss.161) ii  * * is geometric series with first P = 14 L P 1 We can show that E ( X ) = l / p and V ( X )= q / p ? The negative binomial distribution is an extension of the geometric distribution. Poisson's mathematical contribution was to obtain an approximation to the binomial distribution when p is small. The negative binomial variable is merely the sum of r independent geometric variables. The Poisson approximation is commonly used by engineers in acceptance sampling. This procedure fixes the number of items inspected and. . OR ATTRIBUTES expected value is E ( X ) = N p . He was investigating the death by accident of soldiers in the Prussian cavalry. thus. He observed ten corps of cavalry over a period of 20 years and 12). Poisson (17811840). and so we estimate p by X I N . r .2) with E ( X ) = r/p and V ( X ) = rq/p'. and is infinite. 2. The infinite series P(1) + P ( 2 ) + P ( 3 ) + term p and common ratio q .1 failures followed by the lone success. Bortkiewicz found a use for Poisson's distribution in his research. The sample space. We may then estimate 1.1. S. P ( x ) = C(x . the cost of obtaining the estimate. consists of all the positive integers 1. We continue until we have had r succcsscs. . In that case.46 IIISCRETE VARIABLES. This means that we have to have a sequence of X . we derive the geometric distribution from a sequence of Bernoulli trials with the probability of success being p . The probability distribution of X i s called the geometric distribution.16.
18. we can estimate p by taking several samples and letting 3.p2 = p (3. Each corps contained about the same number. P 2 P3 2! 3! (3. t z . p . Summing equation (3. This says that X has a discrete diwibution with X = 0. 2.17. . we have =Pp(0+ 1+p = p . If p is small. Sincc E ( X ) = p..+ . PROPERTIES OF THE POISSON DlSTRlBUTION We now confirm that equation (3. This is the Poisson distribution. o f dying from such an accident. E [ X ( X . Wc do not know either n or p . p . we can replace y by unity and write V ( X )= tzp = E ( X ) . and V ( X ) = rip4. 3 .+ . each of whom had a sniall probability.[ E ( X ) ] ' = p2 + p . + . .17.I ) ] = p2 It follows that E ( X 2 )= pz + p . X.PROPERTIES OF THE POISSON DIS'I'RIBUTION 47 recorded for each of the 200 corpsyears the number of men who died from being kicked by a horse.. What probability law did his data follow? One may make this kind of loose argument. .1) from x = 0 to infinity. of soldiers.18.3) . The number of deaths. but we can write np = p and use Poisson's approximation to the binomial. has roughly a binomial distribution. (3.18. lt has a single parameter. . and so p ( 0 ) e cl.2) and so V ( X ) = E ( X * ) . and We recall that O! =: 1.18.1) Similarly. with E ( X ) = np. 1.1) defines a probability distribution.
The p ( 0 ) = e 6' = 0.7 to be zero.1. this model seemed to be used only for rare events. and had no place in industry. or the number of chromosome interchanges in cells. or the numbcr of misprints per page in a book. We estimate p by x = 0. Poisson Distribution of Deaths X of Observations 109 Frequency fx 0 65 9 Fitted Value 108.1 0. or the number of ships arriving per day in a seaport.19. FURTHER PROPERTIES It can be shown that if X. The last column of fitted values is explained in the following.6 0 1 2 3 4 200 65 22 3 1 44 4 122 3. OTHER APPLICATIONS For quite a while.19. Similarly. 3. we substitute 0.21.1 mm2 in area or counting knots in sheets of plywood. Industrial examples include the number of meteorites that collide with a satellite during an orbit. To make the fit to the data. and so on.61 for p. the expected frequency of x = 1 is (200)(0. it began t o be a popular and useful model for situations in which we count the number of occurrences of some event that wc regard as happening at random with a certain average frequency. Eventually. such as the number of deaths of centenarians in any month in a big city.54335) = 108.3. Examples are the number of cancerous cells per unit area of a section of tissue in a biopsy. Out of 200 observations.61..48 DISCRETE VARIABLES. THE BORTKIEWICZ DATA The data are given in table 3.X 2 arc independent Poisson variables with .20.1. and the number of failures of insulation in a length of electric cable. the number of breaks in a spool of thread.61e" 6 1 ) = 66.54335. we should expect 200(0. ' 3. The Poisson model gives a very good fit to this set of data. Other examples in the literature have inspectors taking sheets of pulp and counting the number of dirt specks greater than 0.3 20.7 66.19. There were 200 observations with a total of 122 deaths.2 4. OR AlTRlBUTES Table 3.
There is light as long as at least one of the lamps works. to taste.. (1. X. and Y is their sum.. Let Y denote the sum of the numbers on the two die. (d) the t w o numbers are the same.4. there are 36 possible results (1.EXERCISES 49 parameters p. . When two die are thrown. Ten judges are each given two brands of coffee. . (g) both (c) and (d) occur simultaneously. EXERCISES 3. in a sample of six devices. (f) both (d) and (e) occur simultaneously. what is the probability that exactly five judges will prefer A and exactly five will prcfcr H? 3. A device consists of six lamps in parallel. The following results are reported by thc registrar of a college. each with parameter p l . 80% failed at least onc o f thc two subjects. + X. . . their sum. 3. l ) .1. Twothirds of the devices that we make are defective.95.3. The probability that any given lamp works is 0. A and 13... (e) Y >8. Last year.). . then Y hits a Poisson distribution with parameter k C (I*. each with probability 1/36.has a Poisson distribution with parameter p. (c) Y is odd. It follows that if X. + &. What percentage of the students passed English but failed mathematics? 3 6 Suppose that the thicknesses of slices o f mica made by a machine are . This is the additive property of the Poisson distribution.. What are the probabilities of the following events? (a) Y = 7. What is inconsistcnt in that report? The amended report from the registrar said that 50% o f the freshmen failed at least one of the two subjects. . ( 6 . and k .2. 40% of the freshmen failed mathematics and 30% failed English. (b) Y = 11. (a) exactly two are defective? (1)) At least two are defective? 3. If there is no real difference between the two brands. What is the probability that. . . 6 ) .2).5. X . are independent Poisson variables. X. What is the probability that at least one of the lamps works? 3. A certain electronic device is rather difficult to make.
he puts it back. if it is OK. chooses two of them at random without replaccment and tests them. An inspector tests one device.9. A piece of tcst equipmcnt detects 99%) of the defective parts that it sees. A second inspector comes along later and he. What is the probability that he chooses a defective device? Note that this is not the same as sampling with replacemcnt. A and 13. 3. 3. Unfortunately. Construct a table like table 3. Eight of them are thicker than specified. At the end of the day.. what is the probability that it came from machine A? 3.0015 in. An inspector . N makes 300 items a day with 5% defective. Our process produces 4% defective items. the outputs of the three machines are collected in one bin. bccause the first inspector only replaces the device that he tests if it is good. R has failed also'! Suppose that the results came from examining 1000 devices. tests one device. A batch of ten items is known t o contain exactly one defective.12. OK ATIKIBUTES symmetrically distributed iibout the median thickness. Suppose that iin itcm is declared to bc defective aftcr testing. Machine A makes 200 items a day with 4% defective. Items arc tcsted one at a timc until the defective is found.11. Component A is found to have failed in 6% of the devices. A quality control engineer measures ten slices. Another batch o f eight devices also contains three defcctivcs. What is the probability that as many as seven slices out of ten will be thicker than the amount specified if the slicing machine is working properly? 3 7 A batch of eight devices contains three defectives. 3. Are the evcnts (a) A has failed and (b) R has failed independent? What is the conditional probability that if A has failed.50 DISCRETE VARIABLES. What is the probability that it actually is defective? Repeat the exercise assuming that the tester catches only 9S% of the defectives and misclassifies 8% of the good items. it also erroneously declares 1 % o f the good parts to be defective. too. and that the manufacturer specifies a median thickness of 0. What i s the probability that it is defective'? If it is defective. and C' makes 400 ii day with 2% defective. component B in 8%. An itcm is takcn from thc bin and testcd. Failures in the device are accompanied by failures in one or both components. What is the probability that both are defective? 3.10. A plant has three machines.8. If it is defective. Both components have failed in 4% of the deviccs. A device has two critical components.1 t o prescrit the data.15. he throws it out. Let X be the .
17. 5 . thcrc is a probability p = 0. 3. What is the probability that he will unnecessarily llag thc machine for overhaul if the process is on specification? 3. Calculate E ( X ) .EXERCISES 51 number of items tcstcd. each with probability 1/6. He will reject the lot if there is more than one defective item in the sample.50 that the valuc will bc above the specified value.16. which is the ccntcr line o n his chart. 3. Calculate E ( Y ) and V ( Y ) . X takes the values 1. If the process is on specitication. 3. 3. Hc decides that he will flag the machine for overhaul over the weekend if thcrc arc either seven (or more) out of ten points abovc the line that week or seven (or more) out of ten below t h c line.14. and 6. A plan for examining lots of incoming items calls for . 3. what is thc probability that a lot will pass inspcction? You may assunic that the lot size is so large that it is appropriate to use the binomial distribution rather than the hypergeometric distribution. A quality control engineer takes two samples each day from a production linc and computes the average of somc characteristic. .15.in inspector to take a sample o f six items from the lot. 4.50% of a value below the line. and that the probability will be 0. Calculate E ( X ) and V ( X ) for the geometric distribution. If the vendor is producing items with 10% defectives. Let Y dcnotc the sum of thc numbers on two die as in exercise 3. Calculate E ( X ) and V ( X ) . Let X denote t h c numbcr that is shown when a single die is thrown. 3. Generalize your result to thc case whcre therc is one defective in a batch of n items.13.1.
M. for example.r) to +a or the half line from 0 to +m. A might. PROBABILITY DENSITY FUNCTIONS The thickness of the layer of probability jam in our analogy is formally called the probability density function. S. One can also have hybrid variables that take values over both intcrvals and discrctc points. We turn now to continuous variables ' . as one might spread jam unevenly over a piccc o f bread. we discuss the exponential and gamma distributions. Events are subintervals or collections of subintervals of S. or distributions. 4. take any value in the interval 4 < x < t ( . distribution. The most important of the continuous probability laws. or Gaussian. is defined in terms of intervals. Inc CHAPTERFOUR Continuous Variables 4. In theory. The density at t h e value X = x is denoted by f ( x ) .2.2. we considered discrcte variables whose sample spaccs are sets of discrete points. The sample space. John Copyright 0 1990 by John Wiley & Sons. I . Now we picture the probability being spread over the interval in a layer whose thickness may vary. Thc picture of the contents of a 1lb jar of probability jam being allotted in lumps changes. The reader who is pressed for time may prefer to postpone some of thc sections at the cnd of the chapter and then refer back to thcm as ncedcd when reading the later chapters. we could have variables for which S consists o f a number o f disjoint intervals.Statistical Methods in Engineering and Quality Assurance Peter W. sometimes continuous and somctimcs discrete. The probability that X takes a value in t h c intcrval a < x < b is dcfincd by the integral P(a < x < b ) = 52 b f ( x ) dx . is the normal.1) . whose sample spaces are intervals. A continuous random variable. We shall usually consider random variablcs for which S is either the whole line from . To considcr thcm would take us beyond the scope of this book. INTRODUCTION In the previous chapter. That is thc topic of the next chapter. In this chapter. (4.
in equation (4. and p and u in the normal distribution. (ii) (4. we take the limit as b approaches a .2. P(x > 1.2.0) = Note that it would have made no difference had we asked for P ( x 21. They are called the parameters of the distribution. One advantage is that formulas such as that in the condition of equation (4. The portions where f ( x ) = 0 make zero contribution to the value of the integral. 15 2 dr =  3 4. A random variable takes values in the interval 0 5 x 5 +2.0. If. We have f ( x ) = k x . where k is a constant to be calculated.2. The integral that evaluates the probability is the same in both cases. Find f ( x ) and calculate the probability that x > 1.2. the following conditions are imposed on the density function: for all x in S.2.3) can be written as the integral from a. is zero. From equation (4. In practice. to +=. This implies that the probability that X takes any single value. such as 8 in the uniform and exponential distributions that follow in later sections of this chapter. . n The probability density functions usually contain constants. as opposed to the set of values in an interval.1). and onc of the main problems in statistics is to obtain estimates of the parameters from a set of data. Example 4.3) where the integral is taken over the whole sample space.1. their values are unknown.3) where k = 112. It is customary to denote them by Greek letters.0). the integral shrinks to zero. It is convenient to adopt the convention that f ( x ) is zero for values of X that are not in S. The probability density function is proportional to x.PROBABILITY DENSITY FUNCTIONS 53 Since the probability of an event has to be nonnegative and the probability that X takes a value somewhere in the wholc space S is unity.
0..F(a) 8 l h e distribution function. x and equation (4.3. F ( x ) . (Thcrc may be cases when the graph of y is a collection of pieces of curves or a step function. THE CUMULATIVE DISTRIBUTION FUNCTION In general. The graph o f y = f ( x ) is an isosceles triangle with a vertex at the point (I. I). i. Alternatively. elsewhere Find F ( x ) and evaluatc P(O.3. if b > u .e. we may define F(x) by F(x) = P ( X I ) . F( +s) 1.3. 15x52. For the righthand portion. (iii) F(xI)= 0. = (4.3.3.5) Example 4. .4) (4.5).5 5 x 5 1. The cumulative distribution function of X (we usually omit the word cumulative) is defined by (4.2. or areas under the curve. F(x) = [x dx = x22 '  We may take the lower limit as zero because f ( x ) = 0 for x c.) The defintion of equation (4.54 CONTlNUOUS VARIABLES 4.2) P(a . (ii) F(x) is a nondecreasitig function. the probability density function defines a graph y = f ( x ) that lies on or above the xaxis. For the lefthand portion of the triangle.3. F(b) 2 F(a).1)may be modified to read (4.0.3) (4.1) defines probabilities in ternis of integrals.1) It is the area under the curve between the lowest value in S and x.1. satisfies the foliowing conditions: (i) F(x) z 0 for all x. A random variable has density function F(x) = x =2x =0 05x51.= X < b ) = F(b) .2.3.
4. has the distribution that we used before in example 4. 1  =o . I ) . = 0. X = X . . xrl. (4. and X .5) = 1.THE UNIFORM DISTRIBUTION F(x)= = I .+ X. has More generally. There arc two common generalizations of the uniform distribution on (0.1.3) O2 > 0.4.3.(2 . f(x) = 1 o<x <1 .. O<x<l.1) This density function obviously satisfies the conditions of section 4..5 5 X 5 1. For this distribution.5) = F( 1. each of which has a uniform distribution on (0.4. the uniform distribution on the interval 8.' 55 x dx + (2 . 1). e*<x<t). c x < B2 with c f(x) = 1 02 1 .F(0. are two independent random variables. =0 elsewhere . .4.5 + 0. elsewhere.5 . It is immaterial whether we use the open interval or the closed interval in the definition. (4. 1). The distribution function is F(x)=O =x 1 X<0.6. U Some writers prefer to develop the theory by starting with the distribution function and then introducing the density function as its derivative.O) has P(x)= s 1 O<x<B.4. T h e uniform distribution on the interval (0.5) .~ 2 R= . =0 elsewhere .2) We will see later that if X .X ) d . THE UNIFORM DISTRIRUTION The simplest continuous distribution is the uniform distribution on the interval (0.x)' 2 P(0. (4. their sum.xj2 (2 .2.
5. Indeed. because. The density function is defined as (4.56 CONTINUOUS VARIABLES A special case occurs when the interval is symmetric about zero.5. we have f ( x ) > 0. It is a good model for the lifetimes of electronic components that do not receive the unusual stresses of sudden surges in current or voltage and have a constant failure rate.2.1) The probability density function for an exponential distribution with 8 = 100 is shown in figure 4.1. (x 1 0 3 ) 10 8 4 2 0 0 100 200 300 400 500 X 600 Figure 4.1.2) is clearly satisfied. it is sometimes called the constant failure rate (CFR) model. The condition of equation (4. THE EXPONENTIAL DISTRIBUTION The exponential distribution is commonly used in reliability work. when x 2 0. 4.6 and $ = +6. = . . 6. i s . . Probability density function for an exponential distribution.9.5.
How do you estimate the average lifetime of the wings'? One possibility is to wait until a11 the wings have failed and then calculate the average.5.393.2.1). has a physical meaning.1) and equation (4.5.3) is satisfied by showing that F( +m) = 1 .3% will last longer than 1500 hours.THE EXPONENTIAL DISTRIBUTION 57 We confirm that equation (4.5.5. Integrating f ( x ) gives us The limit. 39. 22.1) or equation (4.2. of F(x) is indeed unity as required.5. Even if one ' . It is the population avcrage lifetime of the Components. so that F(6OOO) = 1 . Example 4.. One can rewrite equation (4.' = 0.04078. The new parameter a is the death rate. If the lifetimes OO have an exponential distribution. By 6ooo hours.3) is that we have written a instead of 1/8.3% of thc components will fail before 500 hours. If that report is correct.3) The only difference between equation (4. That is not a very practical solution because by the time that the last wing has failed. It will be shown later that the parameter 6.3) to define the distribution is a matter of personal preference. Only 5% of the wings will last for the requircd length of time! n There is a difficulty with this example. We will stay with equation (4.e' = 0.0. measured in components per unit time.e.1) as follows: f(x) = X>0.'.5. 95% of the wings will have failed.223.e xiinon F(500) = 1 .1. what percentage of the wings will last for the desired 6000 hours? We have CI = 2000. there are no planes Icft.5. (4. 0 Specifications call for 95% of the center sections of the Exumpfe 4.F( 1 5 0 ) = e. i. What percentage of them will fail before SO0 hours? What percentage will last for more than 1500 hours? We have F ( x ) = 1 .5.5. what should 8 be? It was reported in the press that the average lifetime appeared to be only 2000 hours. at which the components are failing. Whether one uses equation (4. the rate. 1 . The lifetimes of electrical components made to a certain set of specifications have an exponential distribution with an average lifetime (0) o f 1000 hours. wings of a new type of fighter plane to last for 6 O hours. as x tends to infinity.
4. the last one might take quite some time. whence x" = Oln(2) = 0. Find thc median and the quartiles of the exponential distribution. is a scale factor. Similar calculations show that ( . THE MEMORYIABS PROPERTY OF THE EXPONENTIAL DISTRIBUTION An important property of this probability model is that the system has no memory.69328. Perhaps the reporter was confusing the mcdian with the mean. 2/6 = ln(0. I = 0. When 8 is doubled. The quartiles.7.75.58 CONTINUOUS VARIABLES were to take a sample of the planes and wait until they all failed.2 (cont. 0 4.6. Using this value for 6. Example 4. half the wings had failed? This statement implies that thc median is 2000.5) = In(2). We return to this example in the next section. Example 4.6932 = 2885 hours. The mcdian of X is the value for which F(x) = 0. we can calculate that about 12% o f the wings would last for 6000 hours.5.38638. so that we can estimate 8 as 2000/0. = 1. THE QUARTILES OF A CONTINUOUS DISTRIBUTION The quartiles are simply defined in t e r m of the distribution function.5. respectively.28778 and &.6. (2.). If a lamp hiis lasted until now. the probability that it will last for another 100 hours is the same as the probability of lasting another 100 hours . The median is obtained by solving the equation or Taking natural logarithms of both sides. are the values for which F(x) = 0.1.25 and 0. and Q 7 . the 0 median and the quartiles are doubled too. We see here that the parameter 6. What if the reporter had said that after 2 W hours.
This can be established by using some of the probability concepts that were introduced in the previous chapter. Then. defined by R ( x ) = 1 . is the probability that an item will survive to time x .2) K ( 8 ) = 0. It is defined in the following way: h ( x ) dx is the incrcmcntal probability that thc component will fail in the time period x to x + d x . R ( x ) . Events A and B can occur if. P ( X > ( a .8. by equation (3. givcn that it has lasted until time x .1) R ( x ) = e"'" .8. therefore.2).368. + blX> a*) for all values o f a . (4. Going through all the stages of the calculation and taking the limit is similar to computing a derivative using the step process. The computation of h(x) uses the conditional probabilities.12. 4. Let A be thc cvent that the component lives at least until time a. The hazard rate h ( x ) is the instantaneous failure rate. We end up with the formula (4.8% o f the items will survive past time H.8. Let X denote the lifetime of a component. and B be the event that it lives at least until time u + h. Formally. THE RELIABILITY FUNCTION AND THE HAZARD RATE 'Two other probability functions are conimonly used by reliability engineers. we have A : X > a . and s o only 36. B occurs. or the probability that it would have lasted for 100 hours from the time that it was first switched on. Thus. + hlX u l ) = P ( X > (a.THE KELlABIW'lY FUNCTION AND THE HAZARD RATE 59 was last week.3) . For the exponential distribution (constant failure rate). so that AB is the same as B. and only if. F(x) . (4A. arid u 2 . and I): X > a + h. We have to show that P(RI A ) = P ( X > b). It is. The reliability function.
2% of the devices have failed. h(t) = prp. R ( t ) = exp(t’) . PI If we set f = a.9. Suppose that the lifetime of a piece of equipment has a Weibull distribution with a = 15 years and p = 2. If we write tlu instead of just t . Readers who are interested in reliability will also want to read about the lognormal distribution in section 5. Example 4.x ) . we have R ( u ) = e .exp(f’)).9.368.1) and F(f) = 1 . () p . and replace x by t’. a is called the characteristic life.’ This is the basic Weibull distribution. the element of probability becomes and the probability density function of t is (4. These itenis are warranted . we obtain a more general density function: (4. Like the parameter 8 in the exponential distribution.’ = 0. If wc take the simplest form of the exponential distribution. it is the lifetime by which 63.60 CONTINUOUS VARIABLES The exponential distribution has a constant hazard rate: 4.9. We now modify the hazard rate to make it timedependent.2) and R(f) = exp[ ( 93. f ( x ) = e“ = exp( .h ( t ) = .9.12. THE WEIBULL DISTRIBUTION We have just seen that the exponential distribution has a constant failure rate and a constant hazard rate.1.
THE GAMMA DISTRIBUTION 61 to last for 3 years. 4. the gamma distribution reduces to the exponential distribution with dl = p. is a special case of a gamma distribution. we have the reduction formula When m is an integer. The name gamma comes from the gamma function and is defined by ~(m)=[x"1e'& m>0. What percentage is still alive at the end of the warranty period? Only 4% of the items d o not survive the warranty period.I)! as T(rn).2 ) ( m . The sum of the lifetimes and also the average lifetime have gamma distributions. The gamma distribution is defined for a continuous random variable that takes nonnegative values.10. which is very important in later chapters.10.1) Integrating by parts. we keep applying the reduction formula until we arrive at where (rn .10. it is conventional to define ( m . the distribution is skewed with a maximum .3 ) . (4.l ) ! denotes the factorial (nt . The density function is (4. When a > 1. When m is not an integer.4) with a > 0 and p > 0. Whcn m is not an integer. 0 THE GAMMA DISTRIBUTION Suppose that we put on test n items that have exponential lifetimes. * * ( 3 ) ( 2 ) ( 1 ) That is the usual definition of the factorial. The chisquare distribution. we cannot evaluate the gamma function by the methods of elementary calculus.l ) ( m . When a = 1.
11.1 ) and a long tail on the right as x gocs to infinity.0 and p = 5. (4.1.01 0.1. E(X)= c xP(x) .11. THE EXPECTATION OF A CONTINUOUS RANDOM VARIABLE In section 3. Probability dcnsity function for a gamma variable. 4.9.0 is shown in figure 4.10.10. ru For a continuous variable.005 0 0 20 40 60 X 80 too 120 Figure 4.62 0.03 CONTlNUOlJS VARJABLES 0. (mode) at x = P(a . the corresponding definition is E ( X )= xf(x) dx . The probability density function for a gamma variable with a = 8.9.1) . we defined the expectation of a discrete random variable by the formula of equation (3.1).
11. Example 4.2j Exampfe 4.4) The moments of a distribution are defined in two ways.l)! = ap (4.12. It is sometimes called the mean of the distribution.6) The second ccntral moment is the variance.11. Let X have a gamma distribution with parameters a and .11.ii. or the rth central moment. can be thought of as the longterm average or the center of gravity of the probability distribution. The rth moment about the origin is E ( X ' ) . ! ? The numerator has the form of the gamma function with a instead of a .THE EXPECTATION 0 A CONTINUOUS RANDOM VARIABLE 1 63 This.1. too. = H.3) More generally. . It is the topic o f thc next section.1 1. .S) The rth moment about the mean.1.2. SO that E ( X )=  p a ' 'a! p y a . is defined by (4. We write (4. the expectation of a function g ( X ) is defined by +w (4. (4. Find the expectation of a variable with the exponential distribution.11. Find E ( X ) . and is often denoted by p.
It is often casicr to use the following formula: V ( X )= E ( X 2 ).1) It corresponds to the moment of inertia of a distribution of mass.12.12. we calculate Then V ( X ) = /3”(a + 1) . Following the method of example 4.2.e2 = e2 .1 1.2.4) 4.1.12.11. is dcfined by V ( X )= ysm ( x . Find the variance of an exponential variable. (4. Find the variance of a random variable that has a gamma distribution with paramctcrs a and p. (4.3) Example 4. v.1 that E ( X ) = 0.13.2) Example 4. = (4. (4. Then v(x)= 2e2 .p ) y ( x )dx . That is. denoted by V ( X ) or by a’.a2p2 a@.64 4. MOMENTGENERATING FUNCTIONS A useful technique for finding the moments of a distribution is to take derivatives of the momcntgcncrating function.12. THE VARIANCE CONTINUOUS VARIABLES The variance of a continuous random variable. The positive square root of the variance. .[ E ( X ) J 2 = m.12. is called the standard deviation.12.12. Wc saw in example 4. 2 . as the name suggests. a function that gencratcs the moments. .p .
+ .=r+l.aplace transform. Suppose that X has a uniform distribution on the interval (0. Exumple 4. Engineers will notice the similarity between the momentgenerating function and the L. It follows that we can obtain the rth moment about thc origin by taking the rth derivative of M ( t ) with respect to f and evaluating it at I = 0. but it illustrates the basic idea. 0). .13.+ . .+ .2.1 10t * . and wc can see immediately that the rth moment is The momentgenerating function is .f. is defined as We recall that e" can be expanded as the infinite series When we evaluate the integral of equation (4.j of a random variable.13.0t (01)' . For the exponential distribution =. + . (el)' 2 3! (r + l)! and the coefficient of f'/r! is 6' m.MOMENTGENERATING FUNCTIONS 65 The momentgenerating function (or m.1) and expand it as a series in powers oft.g. . We have f ( x ) = 116. This is a simple example. Find the rth momcnt about the origin. the coefficient of f' in the series is mi/r!. Example 4. .13. X .1.
For example. Summations are replaced by integrals.f. the m.14.P O ] P In later chapters.= O'r! .1)!/3"] [x " . * * and so Example 4. we can say that the variable has a gamma distribution with parameters a = 3 and /3 = 2.1) 'The density function satisfies these two requirements: (4.15.3.13.l)!P"]' = [ ( u. I . The probability density is spread over an area rather than along a line.g. The probability that both (1 < X < h and c < Y < d is givcn by the double integral (4. if we are investigating a random variable and we find that its m.2) . This is useful when we wish l o talk about the sums. We have a bivariate density function 4 ( x . We will see that the sum of 11 inclcpcndcnt exponential variables is a gamma variable with cr = ~t and p = 0. . y ) .g. 4. Suppose that we denote two continuous random variables by X and Y.or the averages.f. For the gamma distribution. 1 'e'" dx x rrl cxp[ 41. is ( I .14.2 f ) . BIVARIATE DISTRIBUTIONS We discussed bivariate discrete distributions in section 3.66 This is the geometric series 1 CONTINUOUS VARIABLES + (Of) + (Of)? + (Of)' m . of random variables.' e =(I pt)". is M ( r ) = [ ( a . The same ideas carry forward to continuous variables.3 . we use the momentgenerating function in an inverse way.14.
15. given that X = x .4) The conditional density function for Y. 4. It must be zero when both variables arc at co and one when they are at +m. Suppose that the joint density function f o r X and Y is and zero elsewhere.15.14.5) II X Y ~ 4 ( x .3) In this function. the covariance measures how the variables vary together. evaluated at the value x: (4.15. y ) = It has to be a nondecreasing function of both x and y.2) As the name implies. is obtained by dividing the joint density by the marginal density of X. it is no longer a variable. It is defined by Cov(X. Y : (4.14. In the simplest caw. x is a given value. h( y l x ) is a function of y alone. Example 4.14.S ( X ) ] [ . . ‘The conditional density is not defined when / ( x ) = 0. (4. if Y > E ( Y ) whenever X > E ( X ) . Y) = E { [ X .1) can be rewritten as Cov(X. we obtained the marginal probability of one variable by summing over all the values taken by the other. In the discrctc case.COVARIANCE 67 The distribution function is defined by P ( X < x . the product [X . COVARIANCE The covariance o f two random variables.15.1) The definition holds whether the variables are discrete or continuous. and if Y < E( Y) whenever X < E ( X ) . For purposes of calculation. Here we obtain the marginal density of X by integrating out the other variable.E( Y ) ] }.E ( X ) ] [Y E( Y ) ] is never negative. (4.E ( X ) E ( Y ) . Y (4.1 5. Y < y) = F(x.1. y ) dx dy . Y) = E ( X Y ) . is a bivariate generalization of the concept of variance. X and Y. and so the covariance is positive. equation (4.
) 12. This fact does not niake X and Y independent. O 4. Then 6’ + (x y ) dy = x + 2’ 1 and It follows that and whcnce 1 7 7 Cov(X.68 CONTINUOUS VARIABLES The marginal density of X is f ( x )= which is similar to thc marginal density of Y. = p . p . This gives The conditional density of Y when x = 1 / 2 is (2y + 1)/2. we need h( y l x ) = f( y) for all values of x. INDEPENDENT VARIABLES In section 3. . which is the same as the marginal distribution of Y.5). For independence.( . we said that two discrete random variables are independent if p.14. we have just seen that this is not true for x = 1/3. is obtained by substituting x = 1 / 3 in equation (4.15.) ( .16. for each i and each j . given that X = 113.. Y ) = 5 . . the joint probability function is the product o f the two marginals.12 = 144 I The conditional density of Y.e.. i.
1) holds. in exercise 4.p. However. each with probability 1/3..\ . If. Example 4. ppot. 0 EXERCISES 4.16. if X and Y arc independent. The marginal probabilities for Y are p. . E ( X ) = 0 and . a buyer wishes to have a 95% probability that a .~. When the variables are not normal. Suppose that X takes thc values 1. and that Y = X'. ( + 1 .. and +1. The two variables are indcpendent if.1. + l ) .EXERCISES 69 The corresponding statement for continuous random variables is this: Let f ( x ) and g( y) be the marginal density functions of X and Y respectively. This is illustrated in example 4. 0.16. The sample space consists of the three points (. The lifetime of a certain component has an exponential distribution. = J . = . what percentage of the components will last for more than 400 hours'? 4. even though the variables are not independent. If 0 = 1500 hours.15."= p0. + 1) with equal probabilitics. 0). When the condition of equation (4.1.1. Y) = 0. The variables are not independent because poo f ~ ~ . we have Thus.2.. (0. and It follows from equation (4. The converse is true when both the variables are normally distributcd.16. their covariance is zero.2) that Cov(X.1.1. and only if.= ( f )( f ) = f J I and 2 p. . zero covariance may not imply independence.
What fraction lasts for 10. 4. what should the buyer specify the average life to be? 4.9. 63. Calculate the constant. What percentage is still operating at the end of the warranty period'? moments about the origin: 4. and m. A continuous variable has the density function f(x) .6. Usc it to calculate the mean. = kx2 1< x < 3 . 710 of the lamps havc failed.000 hours? 4. Sliow that the momentgcnerating function for a Hernoulli variable is 4 +per.5.70 CONTINUOUS VAHIABLES component will last for SOOU hours.000 hours. What would the engineer cstimate the value of 0 to be? 4.10. A device is dcsigned to have a mean lifetime of 20. After lo00 hours of use. Lifetimes of field windings for a gcnerator havc a Weibull distribution with cy = 12 ycars and p = 2.2% of them last for SO00 hours. variance.) 4.8. What is the median lifetime? What percentage fails in the first year (8760 hours)? What percentage lasts for 5 years? What should be thc mean lifetime if wc wish 90% of the devices to last for 40..7. Suppose that in exercise 4.11. and find the mean and variance. Suppose that the lifetime is cxponentially distributed. An engineer puts 100 lamps on test. Compute m3 and rn4 for the exponential distribution.01)0 hours'? 4. Derive the following identities for obtaining central momcnts from 4. (See exercisc 4.5.3. The life of ii certain device has a Weibull distribution with p = 2. What fraction lasts for 10. k.4.ooc) hours'? 4. They arc warranted for a life of only 2 years. but that we did know that 25% of the devices lasted until 9(HH hours. .8. we did not know p.
have joint density Calculate the constant.12.15. Find the marginal densities of X and Y. A variable has density function f ( x ) = k sin x 0 <x < T Calculate the constant. Two variables.14. 4.I ) ! (02 + n .x m . It is stated in the text that the sum of two independcnt random variables that have uniform distributions on the interval (0. V ( X ) . Confirm that. 4. and calculate E ( X ) . n ) = J J . r ) " . 4. n .' ( l . . 4. and Chv(X. Derive the momentgenerating function for ii Poisson variable.l)!(n . (0.EXERCISES 71 4.17.' d r = (rn .. +0 ) : X. k . Y ) .13.3.. A random variable.1. The Beta function is defined by I R ( . has a uniform distribution on the interval =o elsewhere. X and Y. k. 1) has the distribution of example 4.l ) ! A variable has a beta distribution with parameters a and /3 if its density function is Calculate the mean and variance of X. 4. and find E ( X ) and V ( X ) .16.
2. arguments about priority between supporters of the two men. provcd in advanced calculus texts that (5. in his early ninctecnthcentury work with astronomical data. The curve is symmctric about the value z = 0.1) Its graph is the familiar bell curve shown in figure 5. which is also the maximum value. of course. or probability law.2. INTRODUCTION The normal distribution. THE NORMAL DENSITY FUNCTION The simplest form of the normal density function is (5. A t about the same time. the median. it is true that many random phenomena have probability distributions that are well approximated by the normal. Laplace published the same distribution in his book on probability theory. It is perhaps too extreme a simplification t o say that it is called the normal distribution because it is the distribution that naturally occurring random variables normally hdvc. However. who. It is.2. Abraham DcMoivre derived the normal distribution as early as 1718 as the limit of the binomial distribution as n becomes large. This distribution is sometimes called the Gaussian distribution after K.1. Actually. so there are. used it as a law for the propagation of errors.Statistical Methods in Engineering and Quality Assurance Peter W. however. Gauss. M. John Copyright 0 1990 by John Wiley & Sons. and the mode of the distribution arc all at z = 0. is the cornerstone of statistical practice. 5.2.1. F. Inc CHAPTER FIVE The Normal Distribution 5. The mean.2) 72 . The normal density function cannot be integrated by the ordinary methods of calculus.
p and v*.2. P .1 0 5 3 1 X 1 3 5 Figure 5. Probability dcnsity function for a normal distribution.and setting x . We write X .2.3 x 0. V ( X )= a Y 2 is said to have . except that it is centered at x = p and that the scale is changed.3) The graph of 4 ( x ) looks exactly like figure 5. we have dP = f(z) dz = 4 ( x ) dx . E ( X ) = p . The general form of the normal density is obtained by introducing the two parameters. * ) to denote that X is normall distributed with a mean. or expectation. V ( X ) = v .u Then.1. where dz = dxfa. The density function for X is (5.THE NORMAL DENSITY FUNCTION 4 73 0. and variance.1. writing d P for the element of probability.2. We show later that the arameter v is the scale factor.N ( p . It is the standard deviation of X.
We see that 99% of the population lies within the 2.Oc147 o.9554.   5.%a) .5). within a distance t 2 u of the mean.S). (b) P(X < 81.3.1. and 2.9750.0.5% of the time it will be in the upper tail.0250 0. Probabilities for Special Values of z 2 _ I P(z 5 z c +z) P(Z 2) I .96.P ( X < /A .3. we see that the values of F ( z ) change from 0.96. From the symmetry of the curve.9(#)0 0..3. we see that F(1.00 0.96)=0. Let X N(89.1. given E ( X ) and V ( X ) .9750 to 0.95: /A .36.6 sigma limits. ii stirndard normal variable will lie in the interval (. (5.1. 9 65 x ~ I p + 1.1 . I ) .96)0. 1). tind values of F ( x ) . one observation in 800 on the high side.&I5 I .954 0.2) If we enter the table at z = 2. The change is small enough for us to take as a rule of thumb that 9S% o f the members of a normal population lies within the two sigma limits. We have seen that if X N ( p.96) = 0.1.0500 0. 2.96c:Z~1.575 2. and one in 800 on the low side.0 < X < 98.  Table 5.1. and (c) P(86.74 THE NORMAL DISTRIBUTION a standard normal distribution or to be a standard normal variable.96 2. Z < .9974 0.96U) = 0.95. Z > +1. Several other USC~IJI values of z are listed in table 5.3.60 3.1.0250 and 0.1584 0. By writing X = p t zu.96). then (X .9906 0.0 rather than 1. THE TABLES OF THE NORMAL DISTRIBUTION Table I (see Appendix) lists values of F ( z ) . We write Z . Example 5.025 and P(1.96~ . (5.00). + 1.025 .e. we can. Find (a) F(x > 93.5% of the time in the lower tail.N ( 0 .1) Y(X >p i.9900 0. For 95% of the time.0050  o.96.1 .96) = 0. v ’ ) .0).96.p ) / v N ( 0 .3.00 1.96.3.950(3 0. we also have P(Z < 1.025. so that P ( Z > 1.0228 and 0.6832 0.0228 0. Only one observation in 400 lies outside the bounds set by /A C 3u.00 2. i. Entering the table at z = 1.ot113 . From this table. this translates to the following two statements that are each true with probability 0.
0. When X = 93.75.5.89. It follows that the third central moment of X is also zero.5. t cc Zf(2) dz and E ( X ) = p.7734 = 0.THE MOMENTS OF THE NORMAL DISTRIBUTION 75 (a) The mean of X is 89. . Z = t 1. (c) When X = 86.5j6.F( 125)= 1 .'(0.25) = 1 .8943 = 0. and the third and fourth central moments of the normal distribution.0)/6. (b) When X = 81.5) .5) = P ( Z > 0. when X = 98. THE MOMENTS OF THE NORMAL DISTRIBUTION In this section.2266. A similar argument shows that all the odd moments of Z are zero. We then argue that since X = crZ + p . We need to calculate F( 1.5) .5) = F( 1.1 + F(0. E ( X ) = a E ( 2 ) + p and the rth central moment of X is vr times the corresponding moment of Z.0. Z = (93.6247 5. we have s o that and zf(z) ctz = Hence.75) = 1 .5.0.6915 = 0.0 = .5 . the variance.0. Z = 0.I. and so P ( X > 93. From symmetry.5.9332  + 0.25) P(Z > f 1. Our procedure is to calculate the moments for the standard normal distribution. we calculate thc expectation.75) = 1 F(0. JT.0.1.1056. P(Z < . 2 = 7.5) 1 = 0.4. Since the standard normal density function is symmetric about the line z = 0.0 = 0.0.25.1. the standard deviation is 6.
'0 z2f(z)dz = j in ou z[zf(z)] dz 1 V ( 2 )= 2T Application of L'HGpital's rule shows that the first term on the right side is zero.5. 5. V(Z) = By integrating by parts.6. It can be shown that the momentgenerating function for the normal distribution is M ( t ) = exp( p1+ is $) . j t OI. and the w.1) We actually want the generating function for the central moments. Thus. meaning that t h e curve f ( x ) is symmetric . A similar argument shows that the fourth moment of Z is m. The skewness is measured by For a symmetric density function.76 THE NORMAL DIS'IRIBUTION Since E ( Z ) = 0 . r = 0 gives us Scalefree versions of thc third and fourth momcnts are used a s measures of the shape of density functions.5. = 3. THE MOMENTGENERATING FUNCTION The results of the previous section can also be obtained from the momentgenerating function. V(Z)==l and V ( X ) = w2 . The second term is unity. fourth moment of X is 3 ' 5. which M * ( t ) = M ( t ) exp(p t ) = exp  Differentiating M * ( r ) as many times as necessary and putting the desired central moments. (5. SKEWNFSS AND KURTOSIS (3 .
7. which is less than the kurtosis of the more widely spread normal distribution. In section 5. we begin our investigation of statistical procedures. The even moments are V(X)=j o2 and m. Denote the momentgenerating function of U by M(1). The result of this section is expressed as a theorem. Let U = X. They subtract 3 from the kurtosis and call the differencc thc excess.. The difference is manifested in the kurtosis.6. or average.8. of a random samplc of observations. rcspectively.SUMS OF INDEPENDENT RANDOM VARlABLES 77 about the line x = E ( X ) . ( Z ) = 3. Suppose that X has a uniform distribution over the interval (0. We now consider how the sum of two independent random variables behaves.( x ) and f 2 ( x ) . it is possible for two densities to have the same mean and variance and still differ markedly in shape. Exumple 5. The normal distribution has tails going off to infinity. The density is 1/26. The kurtosis of the normal distribution is m . SUMS OF INDEPENDENT RANDOM VARIABLES In the next chapter. we introduce the important central limit theorem. The graph of the uniform distribution is a rectangle. In the next three sections. It is symmetric about the line X = E ( X ) = 0. and momentgenerating functions M . The kurtosis of a density function is measured by This a measure of the pcakedness of the density curve. we discuss the behavior of sums of random variables.this density has the same mean and the same variance as the standard normal distribution. Theorem 5. Some use the normal distribution as a reference distribution.+ X. The kurtosis o f the uniform distribution is 915 = 1.1.and X . Although the standard deviation is a mcasure of the spread of a density curve. The odd moments are zero.O).7.=4 0 5 ' If 8 = fi. be two independent random variables with density functions f. respectively. Let X. and so the skewness is zero. They clearly do not have the same shape. which is their sum divided by 12. One of the most important statistics we use is the mean. 0 5.1 . ( t ) and M2(r). the odd central moments are all zero.10.
7. . ) E ( X 2+ E ( X .7. Furthermore.3) are that the expectation of the sum of indcpendent random variirbles is the sum of their expectations.(O)M. E ( X . and the variance of the sum is the sum of the variances. . we diffcrentialc M ( t ) with respect to t and equate t to zero. are constants.7.(O) f 2 M .7. ) . ( O )+ Ml(o)M. (5.3) and V ( U ) = E(U') .5) W is said to be a linear combination of the random variables.[ E ( U ) 1 2= V ( X . . .:(U*)= M.I*. t. W'.78 THE NORMAL DISlHIBU'rION More generally. X I . ( O ) M . ) ) (5. ) c wfaT .(o) = E(X?) + 2 E ( X . 'To find E ( U ) . 'Thc generalizations of equations (5. Then (5. then E( W ) =. The details of the generalization are left to the reader.7. we can show that if U is thc sum of several indcpendent random variables.4) V( W ) = W " ( X . . if w I . and 2 w . its momentgenerating function is the product of the individual momentgenerating functions. w .2) and (5. )= 2:  (5. and W= 2: W .2) Similarly. )t V ( X .7.
RANDOM SAMPLES A random sample of n observations from a population is a set of n independent observations. it is a “fair” sample that is typical of the population being sampled. Jndcpendence is essential.7. = + 1 and w2 = 1.wc have w .5) become (5. in which the pollsters reached wrong conclusions because their samples were not random.X. fraternities. Then the formulas of equations (5. because that will bias the sample against students who take their lunch in residence halls.. .8. we see that the expectation of the difference is the difference between the individual expectations.4) and (5. hecause that will bias the sample against the arts students. and its precision (thc reciprocal of the variance) improves as we take more observations. There are numerous examples. .4) These equations show that the sample mcan provides a good estimate of E ( X ) inasmuch as. we do not just go into the student union at lunch time.2) If we substitute 1 l n for w.5). X = E ( X ) .8. Every member of the population must have the same chance of being included in the sample. We say more about this in the next chapter. the variance of the difference is the s u m of the individual variances. but had biases built in to them. and sororities.4)and (5. but because w: = wf = C I.. . .8.7. each of which has the same probability distribution. If W = X . 5.7.8. on the average. We can treat them as realizations of n independcnt random variables.1) and V( W ) = c w2V(X) .RANDOM SAMPLES 79 An important special case involves the difference between two random variables.8. we obtain E ( i )= E(X) and (5. Suppose that we do have a random sample of n observations: xl. x. . Basically. especially in political polling. If we want ii random sample of the student population. .3) (5.7. (5. Nor do wc just set up a booth outside the engineering building or the mathematics department. in the formulas of the previous section. Substituting in equations (5. It is not easy to obtain a random sample.
80 THE NORMAL DISTRIBUTION 5. a larger sample is needed.9.t) . the higherorder terms tend to zero. 5. the distribution of the samplc mean approaches normality for large enough n. How large is large enough depends on the shape of the parent distribution. In particular.1) If the X . We have shown that if W is a linear combination of independent normal variables. M ( t ) . We lean heavily on this in our applications. (5. As n increases. One approach i s to look at the logarithm of M(r) as a power series and to note that the coefficient of 1‘ contains a multiplicr. What if the variables do not have a normal distribution‘? The central limit theorem tells us that for all practical purposes. the momentgenerating function of rvX is M ( w t ) . and we are left with which is the logarithm of a normal momentgenerating function.(w.2) which has the same form as equation (5. are all normal variables.1) and add the exponents to get (5. It consists essentially of showing that as rt increases. LINEAR COMBINATIONS OF NORMAL VARIABLES If X is a random variable with momentgenerating function M ( r ) .7 that the momentgenerating function of W is M ( t ) = M . no matter what the parent distribution of the population may be. t ) M 2 ( w z t )* * ~ M. For a uniform distribution.1). The proof of the central limit theorem takes us further into mathematical theory than we want to go in this book.9. . The importance of the central limit theorem is that we can quite reasonably act as if the mean of a random sample of observations is normally distributed.9.. converges to the form of equation (5. THE CENTRAL LIMIT ‘THEOREM We saw in the last section that the mean of a sample of independent normal variables itself has a normal distribution. the sample mean has a normal distribution. It follows from the results of section 5.1). we take equation (53.10.9. ( w . a s given in equation (5. (1 /rz)’.2). then W too is a normal variable.5. For a skewed distribution such as the exponential distribution. the distribution of the mean is well approximated by the normal when n is as few as six.9.
10.11. ( t ) .11. f.2) We recall from example 4.3 that the momentgenerating function for a gamma variable with parameters a! and p is (1 . Hence. (5.13.1) This has the same form as the integral for a normal distribution with zero expectation and variance (1 . we shall have no occasion in this book to use this formula o r to quote it again! . thc value of the integral is ’ M(r) = ( I  2 r y. we see that W has a gamma distribution with parameters n / 2 and 2.. We can now write the density function of W as (5. . The momentgenerating function of Z 2 is M(r) = E[. Thc chisquare distribution is a special case of the gamma distribution that was introduced in section 4.2 1 y = . (5.“ .xp(t’r)] (5. Table IV (see Appendix) gives tabulated values of the chisquare distribution for various values of degrees of frqedom. be independent standard normal variables and let W = C Z l .11.4) Fortunately.21)’” is missing.2)‘ except that the factor 1 /u= ( I . . . The distribution of W becomes important when we investigate the behavior of the variance of it sample from a normal distribution i n the next chapter. It follows that the momentgenerating function of W is MIV(f) (1 .1 1.3) with M . l h e random variable W is said to have a chisquare distribution with n degrees of freedom. The expression “degrces of freedom” is usually abbreviated as d. THE CHISQUARE DISTRIBUTION Let Z . . Z. Z. ..3) Comparing equation (5. and we can use this property to derive the density function.’THE CHISQUARE DISTRIBUTION 81 511.11.P I ) .
There is a 20% probability that the engineer's 0 ration of lamps will last for a year. It is 1 (1 2t)" ' which we recognize as the 1n.03) . The probability density function for Y is . we see that the probability of a value less than 25. 5.f.1. and let Y = C x. The momentgenerating function of Y is The momentgenerating function of 2Y/B is obtained by replacing t by 2r/6. Example 5.g.f. THE LOGNORMAL DISTRIBUlION The lognormal distribution is another useful probability model for the lifetimes o f electronic devices. be exponential variables with the same parameter.12. Then..0 is 0. . Let Y denote the lifetime of a device. 2Y/6 has a chisquare distribution with 2n d.f. The model assumes that In( Y ) is normally distributed with mean p and variance a2. .80. of chisquare with 2n d. An engineer has a supply of ten lamps whose lifetimes are cxponentially distributed with an expected life of 700 hours. What is the probability that the supply of ten lamps will last for a year (8760 hours)? P( Y > 8760) = P( 700 > 25.13.f.. .x . a new one is plugged in.2 that the momentgenerating function of an exponential variable is M ( t ) = 2 1 16t' Let Y be the sum of n independent exponential variables with the same 6. As soon as one lamp fails. We sum up the result of the previous paragraph as follows: Let x l r. 2Y Entering the chisquare table with 20d. We showed in example 4. 8.82 THE NORMAL DISTRIBUTION An important consequence of the mathematical result that we have just obtained is that we can establish a connection between the exponential and chisquare distributions.22.
its normal score is 1. hence. ) Example 5.2. This is the familiar ogive curve.12. Table 5. Normal probability paper changes the scale on the vertical axis and converts the ogive into a straight line. The median lifetime is ZOOe6 00 = 40. +m) along the horizontal axis. We can turn this around and argue that if we plot the cumulative distribution curve for some distribution on normal probability paper and the plot is not a straight line. the 0 avcragc lifetime is 1 0 e6 . Similarly.1.744) 4%. 0 and a=0. The first observation. in order of size. It is called the eighteenthorder statistic for the sample. . The average value of the eighteenthorder statistic from a random sample of 20 observations from a standard normal population is 1.00 0.59.13.. The median of the lognormal distribution is The expectation is E(Y)=exp(p + . It is shown as figure 5.Ooo hours? In(100)=4.THE CUMULATTVE NORMAL DISTRIBUTION CURVE AND NORMAL SCORES 83 where p is the expected In(life). only two observations exceed it.13.605 .6. the fourth observation. 4 60.2. F(I .200 hours.800. From the normal tables. only 4% of the devices fail = before 10. The plot of the normal scores ( Y ) against the observations (X) should be a straight line. THE CUMULATIVE NORMAL DISTRIBUTION CURVE AND NORMAL SCORES The graph of the cumulative normal distribution is shown in figure 5. 0.13. It plots the cumulative probability along the vertical axis and A ' ( . .13.744. or lazy S.1.1 lists a random sample of 20 observations from a standard normal distribution and their normal scores.13.000 hours. The lifetime of an electronic device (in hundreds of hours) has a lognormal distribution with p = 6 . and cr2 is the variance of the In(1ife). is the thirdorder statistic. What percentage of the devices fails before 10.800 = 1. then the distribution is not normal.15.605 and z= 4.13. We see it again in chapters seven and eleven.300 hours. 5. . A similar result can be obtained by using normal scores. is the eighteenth member of the sample. That is the normal score of the first observation in the sample.
99 0.3.13 0. It is clearly not a straight line! Table 5.06 0.31 0. The plot of normal scores against the observations is shown in figure 5.59 0.4 0.31 1.46 0.0.13 0.13.13.13.30 0.45 0.0) 0.74 0.45 0.1.40 0.36 0.w 1. One would not expect the order statistics for an exponential distribution to behave likc order statistics from a normal population.8 0.87 0.06 0.52 0. Cumulative normal distribution function.84 1 THE NORMAL DISTWIBUI’ION 0.15 0. t 7 = 1. table 5.13 0.20 0. Twenty Normal Observations ( p = 0.40 1.6 0.92 0.2 0 5 3 1 X 1 3 5 Figure 5.11 1. O n the other hand.1.74 1.13.19 1.19 0.32 0.59 .97 0.07 1.96 4.16 0.21 0.59 0.33 1.87 0.18 Normal Scores for che Sample 1.52 0.2 shows a sample of 20 observations from an exponential population with 8 = 100 and their normal scores.80 1. and they do not.
70 0. Normal scorc plol f o r an exponential sample .13.40 c11 Figure 5.40 0.00 Observations a .0  * * * * ** * 1.10 1.70 normal sample. Normal score plot for * * * * ** ** ** ** ** * 0 60 1 20 180 Ohsewations 240 300 c33 Figure 5.3.2 0.13. 1.2 ' * * * * ** 2 * 2.2.THE CUMULATIVE NORMAL DISTRIBUTION CURVE AND NORMAL SCORES 85 Scores * * * * 1. 0.
To what nominal weight should the filling machine be set so that only 2% of the cans of grease fail inspection? 5. The containers in exercise 5. Let Y denote the amount. (b) Y < 74. What percentage of the rods mects specifications? 5. and (e) 68.40 0. Suppose that the lifetimes are normally distributed with p = 1200 and cr = 400.06 1.0 kg and CT = 2.74 1.To what value should the filling machine be set if the company does not wish to havc more than 2% of the containers rejected as below specified weight? 5. The diameters of the rods are normally distributed with p = 1. (c) YC69. A machine makes metal rods.13 0.5.1. both of which have circular cross sections. An engineer picks a rod and a socket at random. The lengths are normally distributed with p = 19.45 0.3. the diameters of the sockets are normal with p = 1 .31 0.87 0.87 0.74 0.13 1. The specifications call for the length of a rod to he in the interval 19.86 24 248 130 76 THE NOKMAL DISTRIBUTION Table 5. 5.4.3.8 cm and u = 5 mm. The inspector weighs T thc cans of greasc and expects them to weigh at least 59Og.5 kg. . Thc quantities of greasc that the tilling machine puts in the containers are normally distributed with the mean at the designated value and ( ~ = 2 5 g .2.59 0.5. Find the probabilities of the following events: (a) Y > 75. A device is designed to have an average lifetime of 1200 hours. What is the probability that the rod will go into the socket'? 5. The amount of product made pcr hour by a chemical plant is normally distributed with p = 72.40 0.31 0. (d) 75.0.59 EXERCISES 5.1.8< Y<77. An oil company has a contract to sell grease in 500g containers.0cm and cr = 2 mm.13.5.92 0.92 1.5 4x < 20. Their empty weights are normally distributed with p = 90 g and C = 8 g.6. Twenty Exponential Observations (6= 100) 2 56 86 86 17 1 60 53 69 146 80 10 314 49 03 90 Normal Scores 0.45 0. Ten of those devices are put on test and their lifetimes are recorded.06 1.2.19 1.4 are metal cans.2.5 mm.1<Y<69. A manufacturer makes rods and sockets.0S cm and CT = 1.19 0.
the average of n exponential variables tends to behave like a normal variable.12.11. 0. for the gamma distribution. Calculate the third moment about the mean: m R= npq( q . g .10. 5. g. The lifetimes of locomotive controls (in thousands of miles) have a lognormal distribution with p = 5.90? 5. the shape of the chisquared curve becomes closer to the normal curve. g. for the exponential it distribution. What percentage of the devices operates longer than 10.4.9.6 assuming that the lifetimes have exponential distributions with 0 = 1200. Derive the momentgenerating function for the normal distribution. and (c) 1320 < Y < What are the first and third quartiles of the average lifetime? How many devices would you have to test if you wanted P(Y > 1100) = 1400.13. g . It also confirms that we are justified in using the central limit theorem t o claim that. Find the probabilities that (a) Y > 1300.5. 5. The life of a device measured in hours has a lognormal distribution with p = 8.p ) . (b) Y < 1150.OOO hours? What are the median lifetime and the average lifetime? 5. .6. .8.9): 5. Show that as parameter a increases. tends to zero and value of the kurtosis converges to 3. The momentgenerating function of a binomial variable is M ( f ) = ( q + pe')" . g .7.EXERCISES 87 Let Y denote the average lifetime of the devices that were tested.. Calculate the skewness. as n increases. What is the median lifetime? What percentage of the controls lasts more than 500. 5.OOO miles? 5. and the kurtosis.700 and u = 0. equation (5. Repeat exercise 5. . H n :You should use the following formulas (see exercise 4. .65 and u = 0.1). This verifies that as the number o f degrees of freedom increases. Calculate the skewness. and the kurtosis.
THE THREE EXPERIMENTS 6. and five can he regarded as a short course in probability theory.2.1. Inc CHAPTERSIX Some Standard Statistical Procedures 6. 88 . In each experiment. or distribution. I n this chapter. . In each expcrimcnt.2. John Copyright 0 1990 by John Wiley & Sons. p and u in the normal.. an inspection is made.1. Experiment One We have a process for producing electronic components. .Statistical Methods in Engineering and Quality Assurance Peter W. exponential. 6 in the exponential. Alternatively. we turn our irttcntion to applications by considering three basic experiments. We consider both point and intcrval estimates for the parameters. . and make our first venture into testing hypotheses.. we select an appropriate niathematical model: the experiments are chosen to illustrate the binomial. . designed to be the foundation €or the statistical applications that are the main purpose of this book. x. INTRODUCTION Chapters three. 6. and normal distributions. Each component is tested and classificd as either dcfective or nondefective. We wish to estimate from the data the pcrcentagc of defectives being produced. These models involve unknown parameters that have to be estimated from the data: p in the binomial. x . we take a sample of n observations. Before thcy leave the plant. x 2 . M. or statistical investigations. . A sample is taken from each batch of components. we can write them in a column and denote thc column by the vector x. The data are a set of n numbers that can bc regarded as the realizations of n independent random variables. four. They can be written as x . A geometer would regard that vector iis a point i n an ndimensional sample space. each of which comes froni the same population. .
Unfortunately. on the iiveritge.is that of a Bernoulli trial. f An appropriate model for X. 6. THE FIRST EXPERIMENT Suppose that the batch that was being inspected had SO00 components and that the sample size was n = 100. 6. the estimate that wc have chosen is corrcct. They are put on test and their lifetimes are recorded. We wish to find the average lifetime of the lamps being produced on the line that day. (6. bc the lifetime of the ith bulb. with probability o succcss p that is the same for all n trials.3. An estimator. (It ciin be argued that since we are not talking about sampling with replacement. It takes the value one if the component is defective and zero if the component is nondefective.1) a) . day out.2.THE SECOND EXPERIMENT 89 6. An appropriate model for x.2.4. that has the property that it gives. THE SECOND EXPERIMENT L. Then X has a binomial distribution with parameters n = 100 and p.1) We recall from chapter three that E ( X ) = np. This does not say that the estimate will be exactly equal to p for every sample. On the average.3.2.3. It does say that if we follow this procedure. the binomial model is adequate. A reasonable estimate is p= X i. Experiment Three We have a process for producing resistors with a nominal resistance of 90 ohms. but. with the batch size SO times the sample size. we will get the correct value.) We have to estimate p from the data. a hypergeometric model would be correct. Experiment Two A sample of n light bulbs is taken from a production line. Let X he the total number o f defectives in the sample. or recipe for obtaining an estimate. denotes the state of the ith component tcsted. The data vector is a sequence of 100 zeros or ones. which is unknown. the resistance of the resistors varies from item to item. Wc wish to estimate the average resistance of resistors being produced by the process that day. day in. 6. x. is the exponential model with density function (6. so that E( = n p / n = p. thc correct value is said to be an unbiased estimator.4.et x. on the average.
say. we . X. Let us assume that the resistances are normally distributed.5. p. and so is a standard normal variable.to be estimated from the data.of E ( X ) is a point estimate in the sense that it is a single point. X itself has a normal distribution. THE THIRD EXPERIMENT Let XI denote the resistance of the ith item. N( p. = with E(X) = E ( X ) and V(X) V ( X ) / n . If we were reasonably sure that the true value of the parameter would lie within 0. we have estimated a parameter. or p .7. by the sample average: the average number of defectives per component. the average resistance of a large batch of resistors.4. The simple statement that our estimate is. Now we have two parameters. Since E ( X ) = p . CONFIDENCE INTERVALS FOR THE NORMAL DISTRIBUTION The estimate. In the last chapter.5. the average lifetime of the light bulbs.1) We discuss estimating the variance from the data later in this chapter. 77. p and v2. 6.6. X.1 of our estimate.9 gives no indication o f how precise that estimate might be. In the other two experiments. (6. we discussed the probability distribution of X at some length. 6.2) 6. A reasonable estimate is (6. THE BEHAVIOR OF THE SAMPLE AVERAGE In each of the three experiments.90 SOME STANDARD STATISTICAL PROCEDURES We want to estimate the parameter 8 from the data. u’). We recall that in the normal case. it is reasonable to estimate it by  @=X. or value. we can use the central limit theorem and act as if X were normally distributed with expectation p and variance p q / n in the first experiment and with expectation 8 and variance 0 2 / n in the second experiment. 8.
965 5 +1.965 c= ~1~ + 1. If.575 cm.0.5% of the cases.7.5%. They were introduced by J.96. we develop a class of interval estimates known as confidence intervals. The average length of the four rods in the sample is 29. The two estimates.95: Hence. We start by obtaining a 95% confidence interval for the mean of a normal distribution. we were to take samples of sizc n from this normal population and compute the interval given by equation (6.1 and 77. In 2. The 95% confidence interval is given by equation (6.9 Ifi 10.1). In this section. Lct Z be a standard normal variable. S. day after day.1) Equation (6. 2p 1. Pearson.96.CONFIDENCE INI'ERVALS FOR THE NORMAL DISTRIBUTION 91 might feel that we had a good estimator.7. Assuming that the lengths are normally distributed with standard deviation 0. u l ~ Then p . Wc have alv'iz = 0. p would be outside the interval on the low side. if we thought that the true value was quite likely to be anywhere in the interval X rf: 10. 95% of the intervals would indeed contain p. produces rods about 30.S): .. are examples of interval estimates. Neyman and E. obtain Yo%.95. in the long run.250.9 rf: 0. We have 95% confidence in the interval because it was dcduced from a statement that is correct with probability 0. and multiplying by 1. 95%. 77. in the remaining 2. we should be less comfortable. and 99% confidence intervals for the process mean. Example 6 7 2 A sample of four rods is taken from a process that .500cm.7. p would be outside the interval on the high side.500/2 = 0.7.. The following statement is true with probability 0.x and (6.0cm in length. 1. On the other hand.1)is a 95% confidence interval for p.0.
250 n I 15.288cm. (6 * 7. 0 The intcrval defined by equation (6.us . We can start with this statement: With probability 0.1. 0.250 cm? We wish to have 1.96(0.1): 29. (6. 0 Shorter intervals can also bc ohtained by taking larger samples.1. a shortcr interval can be obtiiincd with a decrease in Confidence. Then ~. The cost of increased confidence is a longer interval. It sets both upper and lower bounds to the value of E ( X ) .7.250) 5 p 29.931 5 p 5 30.2) For t h e YO% intcrval. i. 0 500 0.95.7.96(0.~ px 5 I .7.986. Example 6.96 in equation (6.. we substitute 1. we substitute 2. On the other hand. In example 6.4) The length of the %% interval is 0.7.2.96(+)500 50.92 SOME STANDARD STATISTICAL PROCEDURES 29.575: 28.e.065 + 1.085 5 p 5 29.1) is a twosided interval. i t 0.3) For 99%1 confidence.645 for 1.822 cm. onesided intervals are more appropriate.575 5 30.250 0 so that v'iz21.98cm.7.219. A sample of at least 16 observations is needed. In somc problems.575 . if we arc not contcnt with an interval in which we have only 9S% confidcnce and call for 99% confidence. the length of the intcrvd increascs to 1.50C)cm.250) . The 90% interval is shorter.92.96()=3. how large a sample should the engineer take to have a 95% confidence interval no longer than 0. (6.7.164 5 p 5 29.36 .
The upper confidence interval is given by I.58.82.C. so be it. The data arc 2. 2.7. Thc value of z that corresponds to a prchibility of 5% in one tail is z 2 1. 2. 2.7.98. A purchaser of acid might specify that the vendor deliver acid whose strength is at least 60%. 2. whence (6. Intervals like those of equation (6.74.68.6) are used if the engineer wants assurance that p is not too high. Example 6 7 3 Ten measurements are made of the percentage of .57. we take some action.59. We say more about this when we discuss hypothesis testing in the next chapter.84. 2.5) Alternatively. We arc not concerned about taking action if the water is purer than required. 2.7.83 c1 .5) would be used if the engineer wants to be confident that p exceeds some lower bound and is not concerned that the value of p might be too high. use twosided intervals. The choice between using a onesided or a twosided interval depends on the application. we can start with this statement: With probability 0. 2. DISTRIRUTION 93 and (6. in a crude oil shipment. butenes.CONFIDENCE INTERVALS FOR THE NORMAI. For this data.95. Find an upper 95% confidence limit for the amount of butenes present..745 and (7 =0.73.. If it docs.. One example might be water quality. 2.15. we have X=2. 2. an impurity. X could be the percentage of some undesirable contaminant.6) Intervals like those of equation (6. if the vendor wants t o send him 65% strength at the same price. A good working rule i s to use twosided intervals unless the situation clearly calls for ii onesided interval.92. p < 2. We wish to be sure that the amount of contaminant present does not exceed a specified level. Sometimes it is obvious that the situation calls for a onesided interval.645.7. In case o f doubt.
the required sample size increases. 9 6 W s p5 n n +1 .8.04) 0.8. (6.8.1. 4 = 0. (6.96(0. 2 ) ( 0 .96d0.8. The ends of the interval involve py. 8 ) / 1 0 0 ~ p 0.the number of successes in trials. As p moves farther away from 0.2) Example 6. They showed that for large enough values of n .2).80. Suppose that the sample of 100 electronic components in the experiment in section 6. If p = 0. 9 6 ~ ( 0 . In these formulas.4) .I . and n q > 5 . and 4 = 1 . 6 0 + 1.np .8.1) and. Following equation (6.8.x m* (6.8.12 5 p 5 0.x/n for y. SO.3) If the sample size had been 1000 with 600 defectives. (6.20.1) becomes X . Laplace generalized De Moivre’s results to values of p other than p = 0.28 .96~0.5 nearly 100 years later.20+ 1. We may assume that X is normally distributed with E ( X ) = np and V ( X ) = npy.24/ loo0 I 5 0 . The best solution to this difficulty is to substitute our point estimates for the unknown quantities: = x / n for p. De Moivre showed that for a fair coin. jj = 20/ 100 = 0.8. Find a 95% confidence interval for the fraction of defectives in the whole batch.94 SOME STANDARD STATISTICAL PROCEDURES 6. the confidence interval would have been 0.57 5 p 5 0.1 . CONFIDENCE INTERVALS FOR THE BINOMIAL DISTRIBUTION As early as 1718. following the development in section 6. the distribution of X.24/10(x). the binomial distribution approached the normal distribution in the limit as ri became large. How large is “large enough”? A useful rule of thumb is to choose ti large enough for nzp>5. the normal approximation will be adequate for a sample of only ten observations.63 .1 . This was the first version of the central limit theorem. this leads to the interval There is an obvious problem. we substitute prior estimates of p and q .5. and we d o not know p. 9 6 m . This means that we can obtain a 95% confidcnce intcrval for p by writing z = .5.7.60. Then equation (6. is well approximated by a normal distribution. p 0.3 has 20 defectives and 80 nondefectives.20 .
~ + 3 .1). 0 The other approach. This procedure is clearly conservativc.25.133 or p =0. where p is only 0. ~ + In this example. is to note that the maximum value that pq can take is 0. and to substitute this value for p y in equation (6.4 or 0.8.20. 8 4 ) ~ n.' = 0 .1). One is to take the upper and lower inequalities as quadratics and solve them for p .96fiqZ or Squaring both sides. not on I.7. For cxample 6. It is reasonable for quite a broad range of values.20.3).8.1). + whence p =0. and the dcgrec of improvcmcnt in the cstimatc docs not really justify thc effort. we substitute 100 for n and 0. When p = 0. = 0. which is useful when p is closer to 0. so that the equation becomes 103. Evcn with p = 0. That is a more lengthy procedure.8.289. pq = 0.458.24. The interval has approximately the same length as the interval that we obtained earlier in equation (6. It is centered on 0. but it is moved to thc right.21.5. When the terms arc rearranged. The end points of the interval can be written as p = 2 1. this is a point against the method. this equation bccomes the following quadratic in p: (n + 3 .4 3 .3 or 0. 8 4 ~ 4 = 0 .8.CONFIDENCE INTERVALS FOR THE BINOMIAL DISTRIBUTION 95 There arc two other ways of handling the difficulty with the ends of the interval in equation (6. we have pq = 0.so that this procedure gives a confidence interval that is only 9% longer than that from equation (6. this estimate would give the interval . ~ 4)(2ni. when p = 0.1.8.20 f o r b .21 and pq = 0.6.5.84p2 .
8. no pennant winner in the National League. and only four in the American League. and few of the bottom teams have failed to win at least 30%.8.SOME STANDARD STATISTICAL PROCEDURES Table 6. replace 1. we recall that V ( X ) = 0'.1). 6. The central limit theorem.8. Applying the central limit theorem to the exponential distribution.0 and py by i. which is longer than the other two intervals.57 59.96 by 2. which was .09 61 3 4 43. this leads to a quick method of obtaining approximate confidence The intervals. Finally. Another example is found in baseball. The pcrcentagcs of the votes obtained by the winning candidates in the presidential races from 19601984 arc shown in table 6.91 50.56' 60. Since 1910. 0. which is the model in the third experiment.8. introduced in chapter five. The formula for the 95% .30.1.99* * The opposition vote was split hctwecn two candidates. says that for most densities. In equation (6.5. section 6.105p20. since it is rare for one candidate in a twoperson race to receive more than 70% of the votes.5) One hopes that this approximation will not be relevant to industrial situations in which p represents the percent defcctives! It is a useful approximation for political polling.17 50.9. CONFIDENCE INTERVALS FOR THE EXPONENTIAL DISTRIBUTION The binomial distribution is not the only distribution whose mean a p proaches normality for large n.1. equation then simplifies to (6. Presidential Election Results Year 1960 1964 1968 1972 1976 1980 1984 Candidate Kennedy Johnson Nixon Nixon Carter Keagan Reagan Percentage of Popular Vote 50. has won as many as 70% of the games that they played in the season. the distribution of ? approaches the normal distribution as n increases.
1) We substitute the point estimate X for 8 in each of the end points. We must. If wc know p .p ) 7 = = E [ X L.10. From the definition of the variance.2 p x 5 + p2] E ( X ? ).804) Example 6. i.9./A2 . wc deduce that V ( X ) a’= E [ ( X . however. a11 obvious estimator for = the variance is the average of the square of the deviations of the observations.10.9.6 I 5 301.10. This change is justified in the next few paragraphs. and the confidence interval becomes (6.ESTIMATING THE VARIANCE 97 interval is (6. (6.4.e.2 = [ E ( X ) ] ?+ V ( X ) (6. Suppose that the average o f n = 100 observations is I I 8 252( 1.2) X = 252.2. change thc denominator to n .196) 202.1. The 95% confidence interval is 252(0.1) If we do not know p . 0 A preferable method for small samples appears in section 6.p)’.9.2) . whence E ( X ? )= p’ I.2 p E ( X ) + pz E ( X 2 ). ESTIMATING THE VARIANCE We recall that V ( X ) E(X ..13. X. 0 6. we can substitute the point estimate.2p2 i pz = = E ( X 2 ).
8.10. Estimate V ( X ) . +3. It should be noted that the derivation of s’ in equation (0.10. and it is easy to calculate thc dcviations by subtracting X from t h c obscrvcd values. they are not independent because the deviations add up to zero. s n1 (6. For that rcason. Example 610. thc estimate of thc variancc is said to have n . Although S has n tcrms.10. (6. A sample of five observations is 1. X = S . l)2.10.. We estimate the standard deviation of X by the square root of the estimated variance.3) Then E ( S ) = n( p 2 + d) n = ( n .4. but the last deviation is dictated by the restriction that the sum is zero. and 6. and + 1. For this sample. and s = 16+ 9 + 1 + 1 + 1 = 2 8 . which leads to an estimate . The deviations arc .4) and so we estimate V ( X ) by s2= . The first rt . 4. 6. + 1 . .98 SOME STANDARD STArISrlCAL PROCEDURES and (6..5) docs not depend on the data heing normally distributed.1.1 degrees of freedom.5) S is called the sum of squares of deviations (of the observations from their average).1.1 of them may take any values.
ESTIMATING THE VARIANCE
99
'Z'hiit calculation worked smoothly because n is small and x is an integer. If .Y is not a "convenient" number, the squares of the deviations will not be so easy to calculate, and the method used before may introduce roundoff error. In practice, it is preferable to compute S by the formula
s=
In this example,
c xf

n
(6.10.6)
S  1 + 6 4 + 3 6 + 1 6 + 3 6  = 2 8 .
S
(2s)2
n
vations:
Zixumple 6.10.2. Estimate V ( X ) from the following sample of six obser
17.021, 17.028, 17.026, 17.024, 17.026, 17.022. The estimate ciin be obtained by following the procedures of the previous example. It is casicr t o subtract 17.020 from cach obscrvation first; recall from cyuation (3.10.3) that V ( X  b) = V ( X ) . This is only a shift of the origin and does not change the spread of the data, which now becomc
O.001, 0.007, 0.006, 0.004, 0.006, 0.002
Multiplying by 1 N changes the data to 0 M
1, 7, 6 , 4, 6 , 2 ,
which is an easier set of numbers to handle. Formally, we have made the transformation
Z = l O (  17.020) with V ( Z ) = 10'V(X) . O OX
We estimate V ( Z ) by
The ostimatc of V ( X ) is 6.07 X 10
'.
100
6.11.
S<)ME STANDARD STATISTICAL PROCEDURES
STUDENT’St STATISTIC
In section 6.7, we derived confidence intervals for the mean of a normal distribution when the variance is known. The intervals were obtained from the standard normal variable
z=xp
u/m *
Usually, the standard deviation is unknown and has to be estimated from the data. We substitute s for u in the equation for z . The resulting statistic is denoted by
(6.11 . l )
This statistic does not have a standard normal distribution because the known variance has been replaced by its estimate, which is itself a random variable. The properties of the t statistic were first investigated by W. S. Cosset, who was one of the first industrial statisticians. He worked as a chemist for the Guinncss Brewing Company. Since Guinness would not allow him to publish his work, it appeared under the pen name of “Student.” Thereafter, Gosset was known to statisticians as Student and his statistic was called Student’s t. ‘Thc graph of the density function of t looks much like the standard normal curvc, except that it is more spread out. To obtain confidence intervals, we repeat the earlier derivation starting with 1 instead of z. and replacing 1.96 by t*, the corresponding tabulated value of t in Table I1 (see Appendix). The value of t” depends not only on the level of confidence desired, but also on the number of d.f. in the estimate of thc variancc. With infinite d.f., the density of t is the same as the standard normal.
Example 6.11.1. In the third experiment mentioned at the beginning of this chapter, a sample of ten resistors was taken from the production line. The resistances of the items in ohms were
114.2, 91.9, 107.5, 89.1, 87.2, 87.6, 95.8, 98.4, 94.6, 85.4.
Obtain a 95% confidence interval for the average resistancc o f the items being produced. A dotplot of the data appears as figure 6.11.1 and a boxplot as figurc 6.11.2.
I
. ... . .. .
I
I
I
.
1
94.9
9. 00
9. 60
,
0
I
102.0
108.0
114.0
C I
Figure 6.11.1. Dotplot for example 6.11.1.
CONFIDENCE INTERVALS FOR THE VARIANCE OF A NORMAL POPULATION
101
I
I
+
9. 00
I
I
I
I
1 I
84.0
%.O
102.0
108.0
114.0
c1
Figure 6.1 1.2. Roxplot for example 6.11.1.
Since there is no indication that a onesided interval is needed, it is appropriate to use a twosided interval. The estimates of the mean and standard deviation are X=95.17 and s=9.36,
respectively. The value of I* for a 95% interval based on 10  1 = 9 d.f. is read from the table as 2.262. The confidence interval is 95.17 (2.262)(9.36)
m
Sps95.17+
5
(2.262)(9.36)
88.47 5 p
101.87
m
Nowadays, one rarely has to carry out these tedious calculations by hand. The printout for obtaining this interval using Minitab is given in the following. The data were entered in column one of Minitab’s work sheet. The command TINT 95 C1 called for a 95% t interval for the data in C1.
SET in cl 114.2 91.9 107.5 89.1 87.2 87.6 95.8 98.4 94.6 85.4
TINT 95 cl N = 10 MEAN= 95.170 ST.DEV. = 9.36
A 95.00 PERCENT C.I. FOR MU IS (88.4697, 101.8703)
6.12. CONFIDENCE INTERVALS FOR THE VARIANCE OF A NORMAL POPULATION
We showed in chapter five that the sum of the squares of independent Standard normal variables has a chisquare distribution. Hence,
has a chisquare distribution with n d.f.
102
n
SOME STANDARD srArisricAi, PROCEDURES
variance as follows.
It can be shown that S / v 2 also has a chisquare distribution, but with  1 d.f. This property can be used to obtain confidence intervals for the
Example 6.12.1. Obtain a twosided 95% confidence interval for the variance of the resistances in example 6.1 1. I . In this example, S = 789.141 and there are n  1 = 9 d.f. The distribution of chisquare is not symmetric. We read from Table I11 (see Appendix) in the row for 9 d.f. that the 2.5% value for chisquare is 2.7(X) and the 97.5% value is 19.023, i.e., with probability 0.055 chisquare will be less than 2.700. Then, with probability 0.95, we have
2.7005 5
S u
5
19.023.
Inverting this inequality leads to the desired confidence interval
s
19.023 Substituting 789.141 for S gives
sm?, S 2.700
’
41.48 5 m2 5 292.27 and 6.44 5 u 5 17.10 .
Our point estimate of thc variance is s’ = 789.141/9 = 87.68. The confidence interval is not symmetric about this value. One should also note that the interval is quite wide. It stretches from 47% of the point estimate to 333%. Admittedly, there are only ten data points, but this does indicate that we need considerably more data if we want to get a really reliable estimate of variance. 0
6.13. CHISQUARE AND EXPONENTIAL VARIABLES
We pointed out in sections 4.10 and 5.11 that the exponential and chisquare distributions are both special cases of the gamma family o f distributions. We show in this section how the chisquare tables can be used to obtain confidence intervals for the parameter of an exponential distribution. The nicthod depends o n the fact that if we have a random sample of n exponential variables, Y 2 C x , / B = 2 n i / B will have a chisquare distribution with 2n d.f.
= 1
CHISQUARE AND EXFONENTIAI. VARIABLES
103
Exumpfe 6.23.1. Suppose that only n = 10 lamps had been tested in the second experiment, and that X = 252, as before. Obtain a twosided 95% confidcnce interval for 8. In this exampti, 2 C x, = 5040. We read values of chisquare with 20 d.f. and obtain. with probability 0.95,
9.591 ~r
whence
__
5040
8
5 34.170 ,
147.5 5 8 5 525.5
The method that we used in section 6.9 gives a different inverval, 95.8 I 8 5 408.2 . For small n, the method of this section is preferable because it is based on the exact distribution of the statistic and not an approximation to it. For larger values of n, the chisquare distribution approaches normality, and the two methods give almost the same answers. This method can also bc uscd to obtain onesided confidence intervals (upper and lowcr bounds).
Example 6.13.1 (cont.). To obtain onesided intervals, we argue that with probability 0.95,
5040 < 31.4, 61 whence
8 > 160.5 ,
or that
> 10.85,
8
5040
whencc
8 < 464.5 .
104
SOME STANDARD STATISTICAL PROCEDURES
6.14. ANOTHER METHOD OF ESTIMATING THE STANDARD DEVIATION In setting up quality control charts, it has been customary to use a different method of estimating the standard deviation based on the sample range. The range is the difference between the maximum and minimum observations in the sample. The advantage of this method is that the range is much easier to calculate than s2; modern advances in computing have made that advantage less important. The estimate, s2, that was obtained in section 6.10 is more efficient than the estimate that is derived from the range. However, for the small samples that are normally used in xbar charts (n = 4 or 5), the differences in efficiency are trivial, and little is lost in using the simpler range method. The validity of the range method is very dependent on the assumption that the samples come from a normal population. The range estimate is obtained by multiplying the sample range by an appropriate constant. For example, it can be shown that for samples of four Observations from a normal distribution, the expected valuc of R is 2 . 0 6 ~ . The corresponding estimate of u is R12.06. The multiplier 2.06 is denoted in quality control work by d,. Table 6.14.1 gives values of d2 for some values of n. The standard deviation is then estimated by
(6.14.1)
The range of the first five observations in example 6.11.1 is R = 114.2  87.2 = 27.0. The value of d , for n = 5 is 2.33. The estimate of the standard deviation from that subsample by the range method is
27.0 u =  = 11.6 ; 2.33

the estimate from the other subsample, the last five observations, is
98.4  85.4 = 5.58. 2.33
We cannot say which of these estimates o r the previous estimate is closer to the “true” value for the population that provided this sample. They are each estimates of the unknown standard deviation. It is obvious that the highest resistance, 114.7, made the first estimate larger than the sccond. The
Table 6.14.1. Expected Values of the Range, d ,
n d,
1.13
2
3 1.69
4 2.06
5 2.33
6 2.53
7 2.70
UNBIASED ESTIMATORS
105
optimist should resist the temptation to calculate several estimates and choose the smallest one!
6.15. UNBIASED ESTIMATORS
Equation (5.8.3) shows that X is an unbiased estimate of E ( X ) , which means that if we use X to estimate E ( X ) , we will, on the average, have the corrcct answer. Before giving a formal definition of the property of unbiasedness, we define the term statistic.
Definition. Let x be a data vector. A statistic, t(x), is a function of the data, i.e., a number calculated from the data that does not involve any unknown parameters; X is a statistic. 0
Definition.A statistic, r(x), is said to be an unbiased estimate, or estimator, o f a parameter, say, 4, if
Unbiasedness is an attractive property for an estimator to possess, but it is not adequate just to have an unbiased estimate. In the first experiment, section 6.3, E ( X , ) = p. This says that if we throw away all the data except the observation in the first component, and say that p = I if that component is defective and p = 0 if the component is okay, we will have an unbiased estimate: unbiased but also unwise! How might one choose between estimators? Which is the best “recipe” for estimating +? One course of action is to use unbiased estimators wherever possible, and, when faced with two unbiased estimators, pick the one with the smaller variance. It is a good principle. In many problems, it is possible to find an unbiased estimator of 4 that we know to have a smaller variance than any possible competitor. Such an estimator is called the MVUE, minimumvariance unbiased estimator. There is an interesting mathematical theory behind the search for MVU estimators, but to investigate it would take us beyond the scope of this book. It suffices at this stagc to say that under most circumstances, the sample mean, X, is the MVU estimator of E ( X ) .
Definition.An estimator of 4 that is a linear combination of the observations is called a linear estimator. It can be written as
a‘x =
C aixi .
(6.15.2)
x. but we gave little indication of why they were chosen in particular. we equated the number of successes X to E ( X ) = np and solved for the unknown p. Are there perhaps better estimators available? In this section and the next. We are considering estimators of the form W(x) = c n. In t h c binomial experiment. This idea is best illustrated by a proof of a result that common scnsc would regard as obvious.3) BY equation (5. best lincar unbiased estimator.7.15.=1.1.3). MOMENT ESTIMATORS In the previous sections.16. V ( W ) = I u ~ v ( x and so we must minimize c a: : ).E(X) and = W). Thc unbiased linear cstimator of 4 that has the smallest variance is sometimes called the BLUE.4). a. hence. . by equation (5.106 SOME SI'ANDARD STATISTICAL PROCEDURES Linear estimators constitute an important class of estimators. That was the rationale tor the estimates that were used in the three initial experiments. are equal. They are used extensively when we fit equations by the method of least squares. c a. It consists of equating sample moments to the expected values of the corresponding population moments and solving for the unknown poramet e n . (6.15. . Show that the sample mean is the best linear unbiased estimator of E ( X ) . we have discussed estimates o f parameters. The sample mean is a linear estimator of E ( X ) . In the normal and exponential examples. Thc method of moments goes back a long way in the history of statistics.7. in which case. Since W is an unbiased estimator. = l / n for all i.15. we discuss two standard procedures for finding estimators: the method of momcnts and the method of maximum li kcli hood. Ca. R( W ) = E ( X ) . 0 6. The estimators that we chose were reasonable.. subject to the restriction of equation (6. Direct application of the method of Lagrange multipliers shows that this occurs when all the coefficients.51. ExampZe 6. a. we equated thc first sample moment X to E ( X ) to estimate p and 8.
2. 10.24. The probability density function is E ( X ) = 0/2. 80 If there are two parameters to be estimated. we equate the first two sample moments to their expectations and solve the simultaneous equations. 2 = 78.17 whence and X ' . W c denote the estimate of the variance obtained by the method moments by G2.= 0. and so 80 = 5 P and 5 p = . The number of items.MOMENT ESTIMATORS 107 Example 6. The tirst two sample moments are and so we write C.0625 . Suppose that we have five observations: 1.16) with r = 5 .O). u Example 6.91 . X. Estimate the percentage of defectives in the process. This estimate is S l n rather than S / ( n .r / y . The fifth defective was the eightieth item inspected. it is biascd on the low side.1. inspected has a negative binomial distribution (section 3. and so the estimate of 8 is 6 = 2X. Thcn X = 4 and 6 = 8. Estimate the variance of the resistances in example 6. 3.16.1.t G 2 = 9136. i l Example 6. A quality control engineer smnples items from a production line until five defective items arc found.16. = X = 56. which is not compatible with thc largest observation. Estimate the unknown parameter 8.I). and 10. Then E ( X ) . A sample of n observations is taken from a uniform distribution on the interval (0. 2.3. 0 This can sometimes give an unreasonable answer. 4. .16.11.
Definifion. MAXIMUMLIKELIHOOD ESTIMATES Another standard method obtaining estimates of parameters is the method of maximum likelihood. maximizes the likelihood of that particular sample. which we can write as P ( x .. We have The maximumlikelihood estimate of 4 is the value that maximizes and In(L) = .17.. This value is usually obtained by differentiation.1. . In(L) rather than L itself. The likelihood of the sample is defined as the product V ( x ) . denoted by 4. differentiating. 4). Suppose that we have a continuous variable and only one unknown parameter.108 SOME STANDARD STATISTICAL PROCEDURES 6. 0 whence. Example 6.). we have a probability mass function L(x.n In(0) . The density function can be written as f(x. 4). A. Find the maximumlikelihood estimate of the pararneter in an exponential distribution. and In the case of a discrete distribution.. . .. It is often more convenient to maximize the natural logarithm. x2. Consider a sample of observations x = (x. ttK d[ln(L)] d0 =s+i? n n5 and .17. . 4) to emphasize that it is a function of both x and 4. in retrospect. The idea is that we look at the data and pick the value of the parameter(s) that.x. which was developed by R. Fisher.
=X and c= 7 s n Notice that we would have obtained the same estimates if we had written u = CT’ and differentiated with respect to u rather than a. i. Usually.2. One is the gamma distribution. Since the largest observation is usually less than 8 and never exceeds it. The normal distribution. which appears as an exercise. In the data set that follows example 6. The likelihood function is Then Differentiating.16. by setting It equal to the largest observation.e. The maximum is obtained by taking 8 to be as small as we can. Another is the uniform distribution. The likelihood function is 6) in the Differentiating L does not find a maximum.27. Find the maximumlikelihood estimator of uniform distribution. this estimate is biased on the low side. it finds a minimum when It = +w.3. we have and Solving the simultaneous equations gives the estimates C.3.16. agree on the same estimator.. the maximumlikelihood cstimate is 10. moments and maximum likelihood. 0 Example 6. consistent with the data. There are exceptions. cl . the two methods.MAXIMUMLIKELIHOOD ESTIMATES 109 Example 6.17.3. See example 6.
10 10.2 6.102 1. EXERCISES 6.563 0.30 14.6 7. 6.061 . 6.5 7. however. A sample o f 25 devices i tested for spccd.80 11.50 10.4 7. The mean grain sizes in a sample of six thin f l s of AIl%Si on Si im measured in p m arc 0. A sample of 20 bearings has an average diameter o f 12.9 7. .4 Makc a dotplot of the data.80 14.70 10.20 10.30 10. omitting the eighth wafer.5 7.40 13. How many independent rncasurements must be taken for an cnginccr to know the octane number to within 0. 6.10 10.4 7. Kepeat exercise 6. Obtain a onesided 95% confidence interval that gives an upper bound on the rncan grain size.60 14.50 11.110 SOME STANDARD STATISTICAL PROCEDURES The method of maximum likelihood has several very attractive theoretical properties. however.1 numbcr with 90% confidence? 6.30.5 7. treated in most textbooks for a general course in mathematical statistics.4.0 10.20 10.2.50 14. Obtain a twosided 95% confidence interval for the mean speed. They are. Their speeds in MHz are s 10. Sixteen wafers are testcd for sheet resistance.50 14. the diameters are normally distributed with cr 0.30 10.10 11.50 10. 6.5.90 14.3.5 7.3.934 0.3 7.3 7.3 7.6.1.673 0. Obtain a twosided 95%) confidence interval for the sheet resistance.30 Make a dotplot of the data.22mm. a discussion o f thcm would take us beyond the scope of this book.50 14.3.863 1.4 7. The resistances arc 7.10 1.6 7.7 7. Find a twosided 95% confidence interval for the average diameter of the bearings being produced. A machine produces ball bearings. 10.61 mrn. The standard deviation of a measurement of the octane number of a gasoline is 0.10 10.5.
Obtain it twosided 95% confidence interval for the average lifetime o f a device using: (a) the central limit theorem.0 17.5 14.4 17.12. 6.06.5 13.7.10. Obtain a twosided 90% confidence interval for the variance of the shcet resistances from the sample reported in exercise 6.6 15. Use the normal approximation to obtain a twosided 90% confidence interval for thc percent defective in the batch. (a) using the sample ranges averaged over the five samples.2 12.11.0 19. (b) the chisquare distribution. How is that figure calculated? 6.9. and showing that the average range is close to 2. cr. 6. A sample of 47 items from a large batch contains 18 defectives.8 17. He says that the figure 55. The lifetimes are assumed to be exponentially distributed.0 13. computing their ranges. A quality control inspector takes four devices off the production line at thc end of each shift and tests them for speed. 6.5 13.4 10.6% in favor has a margin o f error &3%. The speeds for five such samples in MHz are (samples are rows) 16. 6. Obtain a twosided 95% confidence interval for the percent defective in thc batch.5 15.0 14.j ) ’ / 3 for each sample and taking the square root of the average of these cstimates of (r’. A television newscaster reports the result o f a national poll in which lo(#) people split 556 in favor of a question and 444 against.0 22.13.2 14. If you have a software package that draws random samples from a normal distribution. and (b) using the more usual : method of computing s2 = 2 ( y .2 16. The average lifetime of the devices tested is 1832 hours. You can do this with Minitab in the following way: .5.0 17. A sample o f 50 parts from a large batch contains six defectives.8. 6.06 by taking several samples of size 4. = 2.2 18. verify that thc value of d .EXERCISES 111 6.3 Obtain estimates of the standiird deviation. Twenty dcvices arc placed on life test.3 12.
+lo. A sample of six observations is 3. Usc the method of moments to find estimates of a and p from a samplc of n independent observations from a gamma distribution. +O). The moments of this distribution were calculated in exercise 4. 6. 9. 2. normal 0 1. has a uniform distribution on the interval (0. A random Z sample of 20 observations has . A variable. The tenth defective is the 267th device tested.16.12. 1 f(x)=2e =0 B<x<+e. Estimate the fraction defective.112 SOME STANDARD STATISTICAL PROCEDURES random 500 clc4.00.15.17.14. Use the method of moments to estimate parameters 1y and p. and +8. Use the method of moments to obtain an estimate o f 8.15. The beta distribution was introduced in exercise 4. Use the method of moments to obtain estimates of parameters a and p. IITI~X clc4 ~5 rmin clc4 c6 let c7 = c5c6 describe c7 6. An engineer tests devices until ten defectives are found. X. A sample of 100 observations from a gamma distribution has C y = 492 and C y 2 = 1024. y = 16. elsewhere . 6. t5. 6. . 6.0 and C y 2 = 13.18.
In these two examples. A chemical 113 . he decides to take action as if the process were out of control and starts looking for an assignable cause for the bad point. p. he is testing the hypothesis that the process mean is at its usual value. Hypothesis testing is also used in comparative experiments. In quality control work. The hypothesis was that p was cqual to some particular value. The number of defectives is the test statistic. John Copyright 0 1990 by John Wiley & Sons. Inc CHAPTER SEVEN Hypothesis Testing 7. If one of the points falls outside thc control lines. he rejects the lot. the engineer looked at a random variable that was taken to be (normally) distributed with a certain expectation. engineers test hypotheses and make statistical decisions without actually using those expressions. using the binomial distribution. If the point falls outside that acceptance region. Whcn hc plots a point on the chart. the engineer rejects the hypothesis and takes the appropriate action.1. M. the engineer rejects the hypothesis. accept or reject the hypotheses. of a binomial variable. otherwise. The engineer is testing the hypothesis that thc percent defcctives in the lot is at some predetermined acceptable value. is discussed further in chapters 11 and 12 under the heading of acceptance sampling. p. They postulate (set up) hypotheses. In the second example. This a fundamental part of the scientific method. the hypothesis was made about the parameter. p.). INTRODUCTION Scientists make decisions on the basis of data. if it falls in the acceptance region. the engineer accepts the hypothesis and decides that the process is in control. If there are too many defectives. if it is too high and falls outside the acceptance region. collect data. after analyzing the data. If the point falls within the control lincs. the enginccr accepts the hypothesis. An engineer looks at a quality control chart on which the averages of the daily samples are plotted.Statistical Methods in Engineering and Quality Assurance Peter W. Testing for attributes. In the control chart. and. A quality control inspector takes a random sample of items from a lot. he accepts it. the hypothesis being tested involved a statistical model.
In this case. AN EXAMPLE In the next eight sections. 7. One morning. We wish to test the hypothesis We decide (arbitrarily for the present) to take a random sample of n = 16 resistors. We wish to determine whether they meet the specification of 200ohms. The chemical engineer in the laboratory could set up a confidence interval for the difference between the two means. the engineer could accept thc hypothesis that the means were equal. the emphasis was on what are called decision rules. the engineer should calculate the interval and report that as well. This is t h e classical hypothesistesting situation. We can live with that amount of variability. In the previous examples. Our probability model for the process is that the resistance. If the interval included zero. A few runs using sodium hydroxide and a few using potassium hydroxide are made and their averages are compared. The engineer is testing the hypothesis that the means for the two reagents are equal against the alternative that they are not. a large batch is delivered. we develop the theory of testing hypotheses about the mean of a normal distribution in the framework of an example. The quality control engineer and the inspector both need rules to apply to take appropriatc actions.114 HYPOTHESIS TESTING engineer may have a choice of two alkaline reagents in a process. The information on the size of the difference bctwecn the means is important to future decision making. our experience with this supplier has been that their resistors have a standard deviation of Sohms (a number chosen to make the arithmetic simple). measure the resistance o f each. For the latter. the choice might be made on cost or other grounds. He wants a rule for saying yes or no. the hypothesis of equality would he rejected. The confidence interval and the hypothcsis tcst arc mathematically the same. If the interval did not include zero. The inspcctor does not want a confidence interval for the fraction defective. Suppose that we purchase 200ohm resistors from a supplier. and standard deviation 8. We do not expect that each resistor will measure exactly 200 ohms. Our problem is to decide whether the average resistance of the items in thc batch is 2OU ohms.2. The data are compatible with that hypothesis. X. and calculate the average resist .of an item is a normal random variable with cxpectation p. There is a strong connection bctwccn confidence intervals and hypothesis testing. The engineer may conclude that one reagent is clearly better than the other or else that there is not an important diffcrcnce bctwecn them s o far as the quality of the product is concerned.
There is strong evidence that the hypothesis is false. This will be our test statistic. the prisoner may actually be guilty. For a given sample size. many had batches would get through. The two intervals X 2 250 and X 5 150 would constitute the rejection (or critical) region. and reject if X fell outside. the hypothesis of innocence is accepted.3. We know that it is an unbiased estimator of p and that its standard deviation is 8 / f i = 2. If X is not close to 200. presumably. On the other hand. reject 92% of the batches for which p was actually equal to 200. The probabilities of making these errors are conventionally denoted by alpha and beta.8 < X < 200. On the other hand.8) = P ( z 2 0. refusing to buy that batch.10) = 0. Accepting the hypothesis means that there is nof strong evidence to reject it. The interval 150<X<250 would be the acceptance region. and the middle groundnot proven. but. We can make two errors in the decision process.2) = U(X c: 198. with acceptance limits so far apart. X. If we accept the hypothesis. respectively.THE DECISIONS 115 ance. The type 11 error is t o accept the hypothesis when it is false. How do we define close‘? Look first at two extreme possibilities. we are strongly persuaded by data with X outside those limits that the hypothesis is false.2. We can accept the hypothesis or we can reject it. Scottish criminal law used to allow three possible verdicts: guilty. we should. if the acceptance interval were 199. We might decide to accept the hypothesis if X fell within the range 150 < X < 250. Neither we nor the supplier would doubt that the batches we rejected were indeed substandard.46. but t h e evidence is not strong enough for conviction. Both risks can be reduced by increasing the sample size at the pricc of increased inspection costs. . THE DECISIONS There are two decisions that can be made on the basis of the test data. we argue that the hypothesis is false and reject it. How do we decide? If X is close to the hypothesized value of 200. When the acccptance limits are wide. Rejecting the hypothesis means deciding that the average resistance is not 200ohms. we decide that the batch o f resistors meets our specification. The type I error is to reject the hypothesis when it is true. acecpting the hypothesis does not really mean agreeing that the hypothesis is correct. Otherwise. one risk can only be reduced at the expense of increasing the other risk. Thc prisoner is declared to be guilty only if there is ii preponderance of evidence that causes the jury to reject the hypothesis of innocence. and we buy it. We reject the hypothesis if X is “far” from 200. we argue that the sample is compatible with the hypothesis. not guilty. and. Those limits arc unreasonably restrictive. We start by assuming that the prisoner is innoccnt. Hypothesis testing is like a judicial process. The choice of a suitablc acceptance region involves balancing these risks in some way. since P ( i 2 200. 7.
when we accept H.. and so it is defined by 200.1. In the example of the resistors. a twosided 95% .5..2200. In this example. If X I or 196 if X r 2 0 4 . from scrapping the whole batch to testing 100% of them. to contain X 95% of the time.0 and standard deviation 2. ACCEPTANCE AND REJECTION REGIONS The usual procedure in designing a test to choose the test statistic and the desired level of alpha. An engineer testing with (Y = 5Vo is arguing that a value of X. rejecting the batch may mean several possible courses of action. test them.96(2) < X < 200.: p. the alternate hypothesis is H. To calculate the acceptance region with u = 5 % tor the resistor example. is called the null hypothesis.0 + 1. The acceptance region is. and then to calculate the acceptance region.1 < X < 203. therefore. there is a very strong connection. Now we have a decision rule. The other hypothesis.. is strong evidence that the null hypothesis is false. If 196<X<204. Values of the test statistic that are in the critical region are said to be (statistically) significant. is called the alternate hypothesis. When we reject H. or test procedure: take 16 resistors at random from a batch. The acceptance region that we have just obtained looks like a confidence interval. or H.. we accept H. H.4. NULL AND ALTERNATE HYPOTHESES We actually consider two hypotheses.. accept the batch. 7.0. we argue that if H. then X has a normal distribution with mean 200. In this example. The rejection region is chosen so that X will fall inside it only 5 % of the time. The traditional choices of level for alpha are 5% and 1%.. denoted by H a .96(2) 196. The hypothesis that we have mentioned.. we reject H.). or the critical region. It is the twosided alternative.91 . We mean by the acceptance region thc set of values of the test statistic that result in our accepting the hypothesis.) is true. The other values of the statistic constitute the rejection region. The null hypothesis is to be rejected if the average resistance is either much less than 200 or much greater than 200. reject the batch.0 .116 HYPOTHESIS TESTING 7. Note that the null hypothesis gets the benefit of the doubtthe cases in which the old Scottish juries would have returned a verdict of not proven. and average the 16 readings.which is so far from the hypothesized value that it could only have happened less than one timc in twenty.
but does want protection against gasoline with an ON less than 90.0 by some “safety factor.: p > 90. i.) The buyer wants to have gasoline with an octane number of 90.92 > 200 .” In terms of hypothesis testing. 7.6. or some other figure.e. When the null hypothesis i s rejected.ONESIDED TESTS 117 confidence interval for p is X .e. Several samples of the gasoline arc drawn and their octane numbers are measured..0) and H. (In the real world. the buyer agrees to buy that load of gasoline.92 . the buyer would take for hypotheses H. A tank truck of gasoline arrives at the buyer’s garage. ONESIDED TESTS Suppose that a company that runs a large fleet of cars contracts to purchase gasoline in large quantities from a major oil company.. if X .92€ 200 and X f 3.92 . The vendor may look at the problem differently. Thc vendor would prefer a situation in which the customer had to buy the batch of gasoline unless it was established that the batch was below specification.3. and the buyer will buy the batch only if the data convince him that the gasoline is not below specification. i. which is identical with the acceptance region. p = 200.< p < X + 3. where c is a value calculated for this test. this would be very unlikely because few buyers have such test facilities. X > 90. The rejection region that the buyer chooses is the upper tail of the distribution of 2. and only if.i. the set of values defined by X > c. The buyer does not care if the oil company supplies higheroctane gasoline than specified. This is a onesided test.0 or higher.1.0.0 (or 190. The truckload will only be accepted if X exceeds 90. unless X .c.9. The interval contains the hypothesized value.3. 196.: p = 90..0.1 < X < 203.” so the buyer will use a decision rule such as “buy the load if.
O They are looking at the situation from two different sides. we reject the null hypothesis. The choice between one and two sides rests with the person designing the test. The vendor would prefer the hypotheses H. We make some runs with their new process and calculate the average yield. <W. we have improved the power of the test to spot departures from specification on the high side. different parties may choose different alternatives. = W . the critical region would have been X > 200 + 1. A good rule of thumb i to usc a twosidcd alternative unless a convincing case can be made for either of the singlesided alternatives. we would like to increase that figure. THE CHOKE OF ALPHA The choice of alpha depends on the consequences of making a type I (or a type 11) error. We can have either onesided or twosided alternatives to the null hypothesis.5 would call for accepting tI. but in a particular situation. each party is looking after its own interests. This example illustrates two points. the alternative had been N. only if X is too high. or only if X is too low.0. = 60. In the same situation.SIS TESTING was significantly less than 90. X > 203. The cost o f a typc 1 error is a few weeks' work on the pilot plant. We wish to test H.: p. By choosing the critical region appropriate to the onesided alternative. The chemists in the laboratory come up with a possible new process.0. 0 (or 290. The currcnt process converts 60% of the raw material to finished product. or if X is either too high or too low? The correct choice is usually clear from the s context of the problem. in the twotailed case and for rejecting it in the other.: p. A value X = 203. Suppose that we arc considering replacing our present proccss for making a chemical product by another process. This is illustrated in the following example. 7. The cost of a type I1 error is .118 HYPOTHI:. The cost that we pay is that we can no longer spot departurcs on the low side. This means that we decide to carry out further investigation into thc new process at the pilot plant level. we may not care about that. The essential question is this: Do we want to reject II.: > 200.0) and H a : p.7.6432) .0 If the average yield with the new process is significantly greater than 60..0 p with a = 5 % .0 against Ha: p > 60. If in the example of the resistors.3 Notice that the upper limit in the onetailed test is lower than the limit in the twotailed test.
The second virtue is that we know how the statistic behaves. X is a mapping of that point on to the real linc. a small beta would be nice. that is a matter of opinion.THE 2TEST 119 the loss of an exciting new process that could makc the company a lot of money. we will be allowed enough time and research budget to achieve that also. THE ZTEST There are two important advantliges in thc choice of X as the test statistic in the problems of thc previous sections. The null hypothesis is that the new drug is of no usethe two means are equal. We could just as well use any equivalent mapping. acceptance and rejection. We can regard X as a summary. The others are given the placebopills that do not contain any drug (although the recipients are unaware of this). 7. We want alpha to be small. At the next stage. Now we had better he sure that we do not erect a costly monument to our misjudgmcnt. Supposc that a pharmaceuticar company proposes a new drug that allegedly alleviates the symptoms of some illness. It is a good estimator of p. It is normally distributed. which calls for a low value of alpha. by most standards.8. The new drug will be approved only if thc paticnts who receive it have significantly fewer attacks per person than the control group. it is the best estimator. One group is given a course of pills made from the new drug. The responses of the patients are noted and a test statistic is calculated. With luck. That is a very serious step. Indeed. The last example in this section concerns drug testing. The usual procedure is to select two groups of patients. We know that X is the statistic that we would use to estimate j ~ and that X . there is the test. a very conservative test. We do . Thus.25: X is the minimumvariance unbiased estimator o f the parameter that is being investigated. the consequences of an incorrect decision are different. too. or reduction. This brings us back to the subject that was only briefly touched upon in section 6. Unless the scientists think that there is some danger that the new drug will actually make the patients worse. I t might be the difference in the numbers of attacks per patient in thc two groups. At this stage we can afford to take a fairly large value of alpha (and a corrcspondingly small value of beta). but now the consequence of rejecting the null hypothesis is that the company will choose the new process i d build a new plant at a cobt of several million dollars. Rejecting the null hypothesis means approving the new drug for use. As for beta. The hypotheses tested after the pilot plant runs are the same. of the set of n data points into a single value. the test has a onetailed alternative. Wc can rcphrasc the development o f the test in the following way. The real line is then divided into two regions. The mathematician sees the sample as ii point in ndimensional space. takes values (at least in theory) anywhere on the real line from oo to +a. that property was used to obtain the acceptance limits.
the value of beta is seen to depend on the value of p.(z) for the cumulative distribution of the standard normal distribution. p = 56%. Instead of focusing on X directly.. A plot with p along thc horizontal axis and the power along the vertical axis is called the Operating Characteristic (OC curve) for the test.0? In this case. in the null hypothesis.1. I has a standard normal distribution.0. The ztest is no different from the test using X itself as the statistic. of being more general in its statement and of pointing out more clearly the connection between tests and confidence intervals.0 and standard deviation 2. Beta is the risk of failing to reject the null hypothesis whcn we should do so. ‘The OC CUNC is the plot of . It measures the ability of the test to discriminate between p and the value p. With u = 8 and n = 16. and the corresponding z statistic is (X . is the type I1 error when p = 205. Thus. for example. When the null hypothesis is true. The OC curve for this example is shown in figure 7. It has the advantages.7 = 0. Thus. we focus on where the null hypothesis is H. wc have 203.85542). THE TYPE I1 ERROR The onesided test that was developed earlier for resistors tests p = 200 against the alternative.: p = fi.3 . What if the null hypothesis is falsc. the rejection region is X > 203.120 HYPOTHESIS TESTING not have to use 2.9. we accept the null hypothesis if 111 < 1. We can use any function of X that has a onetoone correspondence. at least pedagogically. the beta risk is 20%.0. Measuring in units of the standard deviation of X. Then The alpha risk is 5%. and the mean resistance is not 200.0.3. X is normally distributed with mean 205. So we may as well focus on a function of X that we can handle. Similar calculations show that when p = 204.p. 7.0. This is the rationale for the ztest. for example.205.205) 12. We write F.9. and when p = 203.96. we wish to have a twosidcd test with (Y = 5%. p = 3696. but has some other value? What. if. p > 200.0 = 1. The power of the test is defined as 1 .
The power requirement leads to ~ = 2 0 3 .6 u M U 2 .0 + ( 1. the graph is a straight tine. X is normally distributed about a mean 200. OC curve.2) x) .0ohms.0 ohms. 0 . Suppose also that the engineer wants to have a 90% power of rejccting the null hypothesis if the resistance is as high as 203. (7.0 and the same standard deviation. How should the test be designed? The engineer will choose a onesided test with the critical region defined by X > c. n. we have c = 200.8 . where the rejection region is X > c. To satisfy the requirement that a = 5%. Suppose that the engineer is prepared to take a 5% alpha risk of accepting a batch with resistance as low as 200. Two numbers have to be specified: c and the sample size.1.THE TYPE 11 ERROR 1 121 0. ) (7.1) .4 02 . It is then determined by only two points.( 1 . When it is plotted on normal probability paper. 2 88 ( ~ ) . X is normally distributed with mean 203. 0. I .9.0 with standard deviation 8 / 6 . When the null hypothesis is true.64 8 If p = 203. or ogive. t.9. 190 0 193 196 199 Mean 202 205 208 Figure 7.0. which leads to the next consideration.“ 2 e a h * 0.9. It is the cumulative normal curve.
In example 6.122 HYPOTHESIS TESTING Solving cquations (7.11 with confidence intcrvals. the estimate is obtained from a sample of n ' s observations. and so there are 11 . we get n =61 and c =201.262. the null hypothesis is rejected. The calculated value 1..9. The averagc resistance was X = 95.9.5% in each tail. .0.1. In this case.17 with estimated standard deviation s = 9. and the desired value of alpha.10. they were 1. is t* = 2. Let z and zg denote the values o f z that correspond to our choiccs of alpha and beta (in the example.645 and 1.0 9. Suppose that we wish to test 11": p = 90.f. The value of t obtained from thc sample is compared to t*.1 d.746 is less than 2.10. . and so the null hypothesis is c ) accepted.f. If the observed value falls in the critical region.1. Then the sample size rcquircd is given by (7.11. The normal test statistic z is replaced by Student's t : (7.0 against H.28. there is to be probability 2.: p # 90.f. Wc did this earlier in section 6. respectively). which is the tabulated value for n . Thc test statistic is I= 95. a samplc of tcn resistors was taken and their resistances were measured. Example 7.1) The t statistic has the same number of degrees of freedom as the estiniate.36. we have to estimate it from the data.262.I d.7 The derivation of a general formula is left as an exercise for the reader. and the appropriate tabulated value for I with nine d.3) 7. of thc variance.17 .361m = 1. THE tTEST If we do not know the variance of the hypothesized normal population. If a = 5% with the twosided alternative.9.90.746.1) and (7.2). . and let the difference between the means in thc two hypotheses be 6.10.
11 or highcr.2 87. The sum of squares of deviations is S = 789.9 107.68. 90. The Minitab printout for the ttest is as follows: MTB >set in cl1 MTB> 114.000 N MEAN STDEV SE MEAN T P VAI.UE Cl1 10 95. one cannot conclude that p # 90. 1. it contains the hypothesized value 90.12.12.170 9.4 94. There is much more information here. we might accept thc null hypothcsis at a =5%. We can construct tests in a similar way.746. As the sample size increases. 7.47 5 p 5 101.1 both deal with the saniple of ten resistors.11. The point estimate is s' = 87.6 85.E.141. and we calculated a 95% Confidence interval for the variance 41.364 2. We can see from the confidence intcrval that a larger sample with a correspondingly lower value of t' might have led t o rejection of the null hypothesis.75 0. .2 91. but only barely. and. the engineer had chosen to assign alpha a valuc of 0.6 95. in some problems. would be significant.ooO VS MU N.8 98. Clearly. The 95%Jconfidence interval obtained earlier is 88.1 87. therefore.75 and infinity and can now be easily calculated by modern computers.4 MTB > end MTB > ttest YO c l l TEST OF MU = 9O. f * decreases.27 . for some peculiar reason. Calculating and printing the P value clearly avoids having to look up t" in the table.1 I . I and 6.833. we used the chisquare distribution to construct confidence intervals for the variance of a normal population. That is not good enough. and reject it if we chose a = 10%. Thus. we should have had t" = 1.961 1. which is much closer to the calculated value of 1. The observed difference between X and p would have been declared significant if.0.75. Examples 6.5 89.87 . the data are compatible with the null hypothesis.48 5 a 25 292. It is the area under the density curve for the t distribution between 1.0.TFSTS ABOU 1 THE VARIANCE ' 123 The engineer could just report the bald fact that the t value was not significant. TESTS ABOUT THE VARIANCE In section 6.11 The P value is the smallest valuc of alpha at which the observed valuc of t . The interval gives us a range o f values of p . We can also go back to the tables and note that had we taken iilpha to be lo%.
THE EXPONENTIAL DISTRIBUTION Chisquare statistics can also be used in tests about exponential distributions.): fl = 160 against H. we compare this value to the 95% value. p.13. therefore.: 8 > 160.92. There were ten lamps in the sample. based on 31 observations. Since 13. we recall that 2 2: x . with alpha set at 594. We obtained a confidencc interval for the percent defective in the population by using a normal approximation to the binomial distribution. Even though the point estimate of 87. TESTS FOR THE PERCENT DEFECTIVE In example 6. 31.92. This is illustrated in example 7. and rejccT thc null hypothesis. Suppose that we We can rewrite the test statistic.f. . The same approximation can be used for testing hypotheses about the percentage.: v 2> 60. we accept the null hypothesis. Then the test statistic would have been 30(87. as (n .68)/60 = 43.l)s2/02. the statistic has a chisquare distribution with 20 d. only ninc d.88 is nearly 50% higher than the hypothesizcd value.05.124 HYPOTHESIS TESTING Suppose that we wish to test the hypothesis H.f.f.13. which is grcatcr than the tabulated value of 43. had the same estimate.68.8.f.00. the null hypothesis would have been rejected. In example 6.12. from thc tables.13. we compare the calculated value 31.f. we cannot say. 7. 7. that the hypothesis is false. The reason is that there was not enough data. and their average lifetime was 252 hours. of which 20 were defective. 16. For a test with a = 0. we considered a sample of 100 components.1.1.00 against H. In that case.77 with 30 d.4.). To test the hypothesis H.84. / B has a chisquare distribution with 2n d.1. Under If. If lfo is correct. 87.: 02=60. we obtained a 95% confidence interval for the parameter of an exponential distribution. With a = 596.153< 16.5 to the tabulated 95% value.153 has a chisquare distribution with nine d. S / 0 2 . the statistic S/W = 13.
The chisquare test statistic is In this example.33 . respectively.25 against the alternative H.1.13. We consider a model with two “bins”: defective and nontlefective. the expected numbcrs are 25 and 75.75. we discuss these goodnessoffit tests. Test at the null hypothesis: H.13.1 Defective Observed (0) Expected (E) Deviation (dev) = 0 .14.25 = 5 and 80 . In 1900.1.33.14.: p =O. Karl Pearson introduced chisquare tests in the context of binomial and multinomial distributions.25 either. and the evidencc is not strong enough for us to reject the null hypothesis. CHISQUARE AND GOODNESS OF FJT We have been using the chisquare distribution as the distribution of the sum of squares of standard normal variables.E 25 5 Nondefective 80 20 75 +S . the history of chisquare goes back carlicr. has a normal distribution with expectation np = 25 and variance npq = 18. However. A sample of 100 components has 20 defectives. The deviations in the bin numbers are thus 20 . L7 7.CHISQUARE AND GOODNESS OF FIT 125 Example 7.75 = +5. the number of defectives.: p f 0 . This is shown in table 7. beginning with another look at the binomial test of the previous section. The test statistic is a =O.05 xn z==. respectively. The observed numbers in the bins are 20 and 80.14. Example 7. Table 7. It has led us to tests about the variance of a normal distribution and the expectation of an exponential variable.p m 5 4.).. We argue that under the null hypothesis. respcctively (notice that the deviations sum to zero). 25 u = . It would not be strong enough for the onetailed test with the alternative H.155.1. X.: p c0. In this section.+ 25 = 25 75 1.1 (cont. 2 5 .
0 Example 7.84 is equivalent to comparing 1.33 c 3 . accept the null hypothesis that p = 0.+ 2 =4. The extension is illustrated by two more examples. Out of 160 components tested. There is only one d.99.1. In table 3.84.126 HY PorfIESIS 'rES'rINO We compare the value of U to the tabulated value. and we compare U to the tabulated value of chisquare. because the two deviations arc forced to sum to zcro. 5.m 75 225 which exceeds the critical value. we listed the frequencies of the digits 1 through 9 in a sample of 320 digits. 3.f.155)2and the tabulated value of chisquare is 3.15.2. therefore. 7.15. It is essential that the test be applied using the actual numbers in the bins and not the percentages. Suppose that an electronic component undergoes a twostagc testing procedure before being put in service.99.84 = (1.33 to 3. with 65% passing both stages. In the example. 3..f.: p . Comparing 1. the percent defective would have been the same. Example 7. and the others passed.155 to 1. Stage 1 Stage 2 PIlSS OE 0 E (J 24 6 = 18 20 32 .15.+ 1 9. the derivation appears as an exercise. pz = 0. two d. Had there been 300 components with 60 defectives.. MCJLTINOMIAL GOODNESS OF FIT 0 This test can be cxtended to thc multinomial distribution with several bins.96. 8 4 . It had been expected that IS% would fail at the first stage and 20% at the second. but the value of U would have been tripled. Since U > 5.65 against the alternative that those are not the correct values. = 0.12 32 104 122 104 + 18 36 144 724 = .25. 20 failed at the second stage.15. U = 1. We wish to test the hypothesis H(.84.20. 18 failed at the first stage and were rejected. and since 1. 225 225 lJ= .+ . Thc frequencies were . for chisquare with one d. we reject the null hypothesis.f.12 24 We notc again that the sum of the deviations is zero. the ztest and the chisquarc tcst are mathematically the same. Therc are.96)2.2. This eyuivalcncc is always true for the binomial test.33 = ( 1. and p3 = 0.
8 4 3 1 3 2 0 2 I 6 and U = 144/32 = 4.80 12.5.6 80.2 The standard deviation of an octane determination is 0.7.30 8. the alternative p # 11. the expected number of occurrences i 32.90 11.20 11.. A samplc of tcn devices was tested for speed. t here fore.19 0.90 8. and a = 5%. 0 EXERCISES 7.50 CY the hypothesis p = 11.60 9.11 0.1. A sample of nine of their films gives the following thicknesses: 0.90 = 5% 7. how many measurements should be made if 10% when the octane number is 80.17 0. That is considerably less than the tabulated value of chisquare with nine d.92.17 ‘Tcst the hypothesis that the actual average thickness by their method is indeed 2.f.EXERCISES 127 freq uency 0 40 1 36 2 29 3 33 4 29 5 34 6 32 7 30 8 31 9 26 A quick glance might suggest that thcrc are too many zeros and not enough nines. 80.0 79. In exercise 7. 7. which is 16. 7 3 Fivc nieasurcments o f the octane nurnbcr of .2 80. Thc deviations arc.2.27 0.0? p is to be .2 p m . Is the data compatible with the assumption that each digit occurs 10% of the time in the parent population? s For each digit.25 0.00 Test at 10.4.17 0. 7. The frequencies of the deviccs in MHz were 10. A research group is trying to produce silicononsapphire filnis with a thickness of 0.0 p m against the alternative that it is not.26 0.00 8.25.3.8 ii gasoline are 79.7 against the alternativc that it exceeds 79.25 vs.20 0. Test at a = 5%) the hypothesis that the octane number is 79.3.
12. test the hypothcsis HYPOTHESIS TESTING tive p # 12. = 2400 against the alternative H. Test the hypothesis that the average MTBF is the same for both vendors against the alternative that it is not. 7. Is this sample compatible with the claims of the nursery that white.3.: 6. the fraction of students exempt was p = 0.5. Out of 1300 voters polled. . 81 are white.11. We wish to test the variance of a normal distribution. A company buys components from two vendors. how many observations should we take in order for a 50% increase in the sample variance. In previous years. H. Following exercise 7.: p = 12.2. A sample of 50 components from vendor A has a mean time between failures (MTBF) 402 hours. at a! = 5%. 7. 32 were exempt from the first semester of calculus. 20 are red. and the rest are particolored. In a sample of 80 freshmen engineering students.00 against the alterna 7.0 against ’ ’ the alternative hypothesis a > 4. test the hypothesis H. particolored.10.: H.31.50. Let p denote the fraction of Democrats in thc electorate.7.: p > 0. a sample of 50 from vendor B has MTBF = 451 = hours. Is an apparent SO% increase in the variance significant if it is based on only seven observations? 7..: 6 > 2400. s’. using a = 5%. When they bloom.6. the hypothesis a = 4.13.50 against the alternative H.5 against the alternap = 13.0 at a = 10%.9.128 7. Assuming that the lifetimes are exponentially distributed. What is the value of /3 if 8 = 3OOo? 7.8.04.5.: p = 0.7. In exercise 6. over the hypothesized variance to be significant? 7. The company gardener plants 120 flowers in a large bed outside the main entrance to the factory. A sample of 18 items are put on test. Test.0. A poll is conducted on the eve of a closely contested congressional race. Test H. test the hypothesis tive p > 13. In this an unusually good year? 7. A sample of seven items from a normal distribution has a sample variance s2 = 6. In exercise 6. 703 are Democrats. The average lifetime is 2563 hours. and red flowers ought to occur in the ratios 9 : 6 : l ? 7.
y . We turn now to comparative experiments in which we compare two populations and ask whether they are the same or whether they differ. a normal distribution from a sample of n observations. . 8. M. p l . u2. John Copyright 0 1990 by John Wiley & Sons. X is normally distributed with expectation p. Since the first sample has m independent observations. consisting of y. comes from a normal population with unknown mean p2 and known variance v : . The second sample. COMPARING TWO NORMAL MEANS WITH KNOWN VARIANCE Suppose that we have two random samples. The first. or to test the hypothesis H. We wish either to estimate the difference. for example. y . We estimate p.2. . . first.: p. .k .. . comes from a normal population with unknown mean p. and known variance (T:. and variance. x2. . when their variances are known. when the variances have to be estimated from the data. and variance crTf/m. Inc CHAPTER EIGHT Comparative Experiments 8. . 129 .x. The chapter ends with a discussion of comparing two binomial populations and the use of chisquare for analyzing 2 X 2 tables.. . estimating and testing hypothof eses about the mean. by 2. . p . later. I . we discussed somc common statistical procedures in which the emphasis was on drawing conclusions about a population from a single sample.Statistical Methods in Engineering and Quality Assurance Peter W. Then we discuss comparing the variances of two normal populations and comparing the means of two exponential populations. We considered. which does not assume normality. . and. Is the yield of our plant the same whether we use hydrochloric or nitric acid as reagent? Is the yield of wafers the same with two different etching regimes? Do we get less variability in our product when we lower the operating temperature? We begin by comparing the means of two normal populations. = k . Y. which are similar problems. This is followed by Wilcoxon’s nonparametric test.. INTRODUCTION In chapters six and seven.
The two sample means are X = 71. A or B .9 .t.2.8 72.2.96qu:lm + &n < p.1. l m + c ~ ~ l n . X j and . t . . If.p2 = 0 is obtaincd by setting (8. and the variance for 13 is 4.62 < p.1 72. 0. we replace X by . A chemical engineer has a process in which either of two reagents.8 71.04 The z statistic for the hypothesis that p.j + 1 .130 Similarly.2.6 75.913 .2) take the simpler forms ' + 1 ln < p.93.2.3) (8.9 69.2 72.2. is an unbiased 2 2 estimate of p I.1 75.0. in equation (6.p2 with variance v . 'I'hcir diffcrence.7 74.1) and (8. 9 6 d 1 lm Xy If a: = a: = u.  < 0.1 .p2 < .7 + 1 .1.2.3 70.2) z = v m .4) Example 8. = p2 is z=which is barcly significant. d = x . can bc used. X .1 72.4 68.3 66.y and alv5 by we obtain the following twosidcd 95% conlidence intcrval for p I.1.83 = 2.83 + 1.h:  COMPARATIVE EXPEKIMEN'I S v m i . means is A dotplot o f both samples on the same scale (see Figure 8. 9 6 u m r n + 1 I n (8.1. 9 6 ~ v ~ l +n : l n .2.79 3. r v pI (8.1 and If we assume that the variance of A yields is 6.10 .1 74.y.96V6l12 + 4/12< pI .00.8 70.1).): < X .1) .5 74. a 95% confidencc intcrval for thc difference between the population 71.0 70. y = 72.7 68. Twelve runs with each of the reagents are made and the following yields are observed: A@): B(y): 71.ailn). 1.7.72.0.3 74.5 72. equations (8.pi < X .1 73.6 73.2 74. a ztest for the hypothesis I!. Similarly.1) suggcsts that the yields with B might bc higher.2.93 . j N( p2.2 70.
!’..1.513 and ss = 4. + s.1 (cont.3.2 d. 0 I I .2.UNKNOWN VARIANCES I I .). If we d o not assumc that the variances are : known. Example 8.2.3. 60 68. ___5__ 6..IaL has a 2 chisquare distribution with n . the test statistic becomes in (8. . ) = ( m .2 m is an unbiased cstimatc of c*with in + n .p7. .2. . we calculate estimates from the two samples: s = 6. ) = ( n . . where f* is the tabulated value for t with n + n .2) This statistic has Student’s t distribution with m + n .f.f. z X  t*sVl !m + I In < F .592.X)’ and Then S l / uhas a chisquare distribution with m . 00 . .f.. In a similar way. . s. . I 131 A I I I 72. UNKNOWN VARIANCES Ilie situation when thc variances are unknown has to he separated into two cases. The pooled estimate of the common variance is s2 = 5. .2 d. . . .1. . (8.1 d.f.. Dotplot for example 8. the confidence interval is obtained by substituting s for [T and t* for 1.Y)* . We deal first with the case in which it is reasonable to assiimc that the two variances are equal to u’.3).I d. S < v t*sd\/lI m +l/n.3) We illustrate this by returning to cxample 8. and S.2.0 1 I 76.1)a’ and so s..3.0 I I B Figure 8. E ( S . Lct S. .2.0 7.4).541.2. By substituting s for equation (8.= 7 and E ( S .l)a2 . . . 8..2.96 in equation (8.  2 ( x . S = 2 ( y .0 1 I 74. m + rt .2 d.
132 COMPARATIVE EXPERIMENTS The 5% valuc for I with 22 d. and so the data are compatible with the hypothesis H. 2.4. the Minitab program gives the If= 1.f.074.91 .1..83 0. 117 8. is t* = 2.: p. which is less.90 . before the days of computers. you are not going to get into serious trouble by acting as if equation (8.4. but now it is routinely available in many software packages. If m and n are greater than about 20. t ~ 2~ o~ i 6 . The ttest statistic is (== 1.3.= s 2 ’ n1 and we can then write the test statistic as (8. in which case a 95% confidence interval is given by Exumple 8. m1 and ’ s. This interval contains zero. = p2. value of I ’ as For this data set.1) has a normal distribution.917 . p < . in absolute value.1) The mathematical problem is that this statistic does not have Student’s distribution and the actual distribution cannot be dcrived in a tractable form. UNEQUAL VARIANCES When it is not reasonable to assume that the two variances are equal. this approximation was awkward to use.1 (cont. and so the twosided 95% confidence interval for the difference is citlculatcd by equation (8.4. than the tabulated value for alpha = 5 % .074. and so we do not reject the hypothesis of equality. 1 = s. There does exist an approximation.).3): . We can certainly estimate the two variances by s.2.3 . . the situation becomes difficult mathematically and is known to statisticians as thc BehrensFisher problem.
therefore. The data have been rewritten to show the pairs and the differences: . )= V( y .1 (eont. (T 2 and ci.. In the earlier treatment of the reagent data. This is called a paired experiment. is The batch component has vanished and V ( d .5.k . denotes the contribution from being in the ith batch. which we assume to Now we add the batch term and write have the same variance u2. where p. = p2 + ezl where pl and k arc the unknown constants whose difference we are trying to measure. took 12 batches of raw material and divided each into two halves. e2.. for that stage and that stage alone. The engineer. The value of I ’ in this case is not much different from the r value of the previous section. in each pair and carrying out a onesample Itest on the differences. = p. THE PAIRED &TEST We treated the data in example 8. which had canceled out.4 d. To illustrate this.” perhaps 24. the other sources of variation have canceled out.2. . free of the batch effect. x .. and the appropriate analysis is called the paired ttest. and e. ) = c r i + cr’. we take another look at the data of example 8. 0 8. and y . . )= 2 ~ ’ . The difference. of wafcrs is split into two halves for one step of the processing. Other expressions are also used.2.THE PAIRED ITEST 133 which should be compared to the tabulated value of f with approximately 22.arc independent random errors. the variance between batches. Suppose that the engineer was aware that there were two sources of variation in the process: random noise and also variability between batches of raw material. there are 12 wafers receiving treatment A and 12 receiving trcatment R .y . In semiconductor work. we had a simple model: X.2. . from x. V ( x . therefore. Then the difference between the two observations made on the same batch was a measure of p .1 as if the 24 observations were all independent of each other.The test procedure consists. splitlot experiments are run in which a “lot.). Example 8. + e. of subtracting y . The variance of a single observation has two components. onehalf received reagent A and the other received reagent B.f.1.
Fewer degrees o frecdom means that a higher value of f* must be overcome to establish significance. but if thc variance between batches is considerable.3 70. I 0.6 74.6 6 70. If the medians are the same.f. One chooses whichever .8 1. Mann and I). the averagc rank of the X variables should be about thc same as thc Y average. they add up to (rn f n)(m t n + 1)/2. the twosamplc [test is the “best” test when the conditions of normality and equal variance are met..6 3 74.1 75.7 75. I I 2 68. Theoretically.8 Pair A B d 69.2 72. Whitney. and so on. From now on.7 .3 2.6.3 8 72.53 with 11 d. we work with the ranks o f the observations.. a test that does not depend on assuming that the populations are normally distributed. we are testing the equality of thc population medians rather than their means.9 66. the superiority of the ftest over Wilcoxon’s test is not great even then.5 72.8 74.8 11 70.5 2.1 12 68. H. developed a similar test that bears thcir names and is mathematically equivalent.6 The onesamplc [test on the differences gives t = 4. the increase in precision from eliminating it will more than compensate for that increasc in the critical vdue of r.1 72. we give each of the tied observations the average of thcir ranks.. The smallest observation has rank one. Formally.0 0. an industrial statistician.0 10 74.7 5 71. which is clearly significant.2 I) 4 72.1 3.2 2.134 COMPARATIVE EXPERIMENTS B d Pair A 1 71.4 74. . Wilcoxon’s test can be as good or better than the rtest.2 7 73.9 2.n observations in increasing ordcr of size. If the X median is appreciably lower than the Y mcdi.in. When there are departures from normality. The tcst statistic is t h e sum of the ranks of either the X or the Y observations. working at Ohio State Clnivcrsity.2 3. n 8. However. Frank Wilcoxon. WILCOXON’S TWOSAMPLE TEST During World War 11.6 0.1). the MannWhitney version is more popular among social scientists. developed a nonparamctric test. Notice that the number of degrees of freedom f has dropped to 11 (12 . the X obscrvations will tend to he clustered to the left and the Y obscrvations on the right. the largest has rank m + n .8 1. At about the same time. If there are tics.e.1 73. R. i. The first step in this procedure is to arrange d l m I.1 70.
11 The sums of the X ranks for the 20 sequences are W = 6 7 8 9 8 13 12 13 14 15.2 4 Y 5 70.1 6 X 70.4 7 X 70. usually the sum that has the fewer observations to be added.9 8 Y . a = I /20 + 1 /20 = 0.2. We remove the balls from the box one at a timc and note the order of the colors. When m and n are ten or more. W = 6 or W = 20.. In the MannWhitney test.WIICOXON’S TWOSAMPLE 135 is mow convenient. and only if. namely. f Tables o the critical values of W are available for various values of rn and rz. reject the hypothesis if. They are XXXYYY XYXYXY YXXXYY YXYYXX XXYXYY XYXYYX YXXYXY YYXXXY 9 10 XXYYXY XYYXXY YXXYYX YYXXYX 10 XXYYYX XYYXYX YXYXXY YYXYXX 12 9 10 XYXXYY XYYYXX YXYXYX YYYXXX 11.5. the sum o the Y ranks has E ( W j = n(m t n + 1 j12 and f the same variance. Exunipfe 61.5 X 69.1 und ignoring the remarks about pairing that were made in section 8.3 2 X 3 68. Note that we are not making any normality assumptions about the actual data. The 24 ohscrvations in ascending order are 1 66. each of those scores has probability 1/20.0 Y 70. ‘I’hcrc are 20 possible arrangements of six balls.3 X 68.c. 11 12 The lowcst value is 6. For this test.). Consider the simple case where m = n = 3. We can now formulate a simple test for the hypothesis that the two populations have the same median.I (cont. i. three red and three white. We suppose only that we have 12 observations on A and 12 on H . The theory o f the test can be derived by thinking of a box with three red balls and three white balls. and the test statistic is the sum of those scores.2. the highest 15. we can act as if thc sum of the X ranks is approximately normal with E( W ) = m(m + n + 1 ) /2 and V( W ) = mn(m + n f 1 j / 2 .05. if X has the three lowest observations or if X has the three highest observations. We are mercly applying that approximation to the orderings of m red balls and n white balls as they come out of the box! This is illustrated by going back to the original data o f exiimplc 8. and so there are 20 possible orderings (ignoring ties) of X and Y. each X observation is given a score equal 10 thc number of Y observations that exceeds it.
John Tukey (1959) presented a simple test that he had constructed to meet "Duckworth's specifications.7 Y 22 74. 10 is the 1% value.3 Y 17 73.2 Y Y 21 74. W is normally distributed with p = 150 and (+' The = 300.8 Y 16 73. we subtract 1 from D ." Duckworth wanted a quickandeasy test for comparing the medians of two populations that did not need any tables and could be used for a wide range o f m and n .136 COMPARATIVE EXPERJMENTS 9 71.1 X 10 71. m 2n. or at least that neither mln or n / m should exceed 2. (It is a good idea to check your calculations by adding the Y ranks. Suppose that the smallest observation is from thc X population and the largest from the Y population.7. The test only wofks if the largest and smallest valucs of thc observations come from different populations. D.8 Y Observations 11. 13 is the 0. in his words. Under these circumstances. that total is 179.1) X 14 72.6 13 72. is the sum of the two overlaps. we reject the hypothesis of equal medians at a! = 5 % . 10. too. the number of X observations that are smaller than the smallest Y plus the number of Y observations that are larger than the largest X.7 X 15 72. the two should add to (24 X 25)/2 = 300. Tukey produced a test that is. They each receive the same rank.5 Y 24 75.) Under the hypothesis of equal medians. (11 + 12 + 13)/3 = 12. and 13. The sum of the X ranks is I + 2 + 3 + 6 + 7 + 9 + 10+ 12+ 14+ 17+ 18+22=121. "not only quick but compuct. thc table of critical values consists o f the three numbers 7. It is preferable that the two samples have approximately the same number of observations. If D 2 7.1 Y 12 72. .8 X 23 75. The test statistic.150 z= m = 1. test statistic is 121 . but it is easy to use and the table of critical values can be carried in your head." This tcst is not as powerful as some of the more elaborate procedures. THE DUCKWORTH TEST In the first issue of Technometrics.67 0 8. and 13 in order are all equal.6 X 74. 12.2 X 11 (72.1% value. If either 3 + 4n/3 I I or vice versa.1 Y 20 74.1 18 X 19 74.
Suppose that two samples have the following estimates of the variance sf = 80. each divided by its degrees of freedom. which we can denote by 4.8.f.8. (8.1) If I. The distribution of F depends both on the number of d. l u ~ ) / ( . which implies that only values of F greater than unity need to be given. To derive the distribution mathematically. we hypothesize that thc two variances are equal.0.2 (cont. a complete table would occupy : many pages. s: 7 .2.= 2. The F statistic will sometimes be denoted by F(+. 40 .8. we note that there are three X observations below all the Y's and one Y above all the X ' s . Table IV (see Appendix) lists critical values of the F statistic. How do we : test the hypothesis a: = a:? The tcst statistic is the variance ratio F= s .we obtained estimates s and s$ of the variances.). Using the Duckworth test for the data of example 8. The total is D = 4. The use of the table is illustrated in this example.. +2). COMPARING VARIANCES In section 8. .8.4. in the denominator.f.f. ExumpZe 8. and m so F is the ratio of two independent chisquare variables. is close to unity. in the numerator and the number of d. In the present situation.1).0.COMPARING VARIANCES 137 Example 8. we start with the ratio The numerator is ( S . I). some books economize by letting s be the Iurger of the two estimates and s$ the smaller. If F is large or small. and +2. we accept the hypothesis that the variances arc equal. which is not significant at a = 5%.2.1. and F reduces to the simple variance ratio si/si of equation (8. 15 d. 15) = . respectively. 0 8.2.0. Because there are two parameters involved. 10 d. we conclude that the variances are different.f. : We compute the variance ratio with the larger estimate in the numerator: 80 F( 10. the denominator is (S2/ui)/(n l). and s 40.
can be a long way off. dividing through by . I ) One more difficulty remains.05 in the tenth column and the fifteenth row: F*(10. 8. and estimates based only on a few n d. we get Unfortunately. that difference is not significant. the valuc o f /+’thatappears on the righthand side o f equation (8.9.s:/s:. we want to compare F with the value that would be exceeded by chance 5% of the time (not ~ 1 2 ) The appropriate value is . It is perhaps discouraging to see that even though thc one estimated variance is twice the other. therefore. The value on the left side is not in the tables.1) is found in the table under a = 0. Since we have already taken care to put the larger estimate in the numerator. construct a (1 . found in the table under a = 0.9. It can bc obtained by using thc following equation: (8.9.a)% confidence interval by writing then.2) .54.138 COMPARATIVE EXPERIMENTS The table gives the value o f F that corresponds to the given percentage in the upper tail. CONFIDENCE INTERVALS FOR THE VARIANCE RATIO We saw in the previous section that the statistic that actually has the F distribution is We can. In fact. good estimates of variance need plenty of data. and so we have to switch the subscripts throughout. The confidence interval that we are seeking is thus (8.9. this i s upside down.15) = 2.f.025. For a 95% interval.
1 (coat. . 2n). The first sample.10. comes from a population with expectation 8. . respectively. .102) Exampfe 8. We assume that . their average lifetime is 178 hours. . . . respectively. = by comparing the test statistic F. comes from an exponential population with pararneter 8. in two ways.x.S)(3. x2. .= Y X (8.52) 8.5)( A)< 4< U(72 (2. Method 1. The second. x .10. 30 components from supplier B have an average lifetime of 362 hours. and 8. . y. y .. and then follow the procedure of section 8.f.10. There is an alternative way that has the virlue o f not k i n g an approximation. . If we divide each of them by their d. can be obtained by the methods o f the previous section as (8. wc can act as if X and 9 are both approximately normally ' distributed with variances X and fz. We can now obtain a Y5% confidence interval for the variance ratio: (2. COMPARING EXPONENTIAL DISTRIBUTIONS Suppose that we have samples from two exponential distributions.a)% confidence interval for 8. we have an F statistic: has the F(2rn. y .. test the null hypothesis H. We can.1) to the tabulated values of F(2rn. A ( I .8. and 2 C y/O2 have chisquare distributions with 2m and 2n degrees of freedom../8.). First.1. . 2n) distribution. We estimate 6. We recall that 2 C x/8. by X and 8.: 8.COMPARING EXPONENTIAL DISTKIBUI'IONS 139 Example 8... We can then compare 8.2. and take the ratio. . . therefore. Twenty cornponcnts from supplier A are put on test. by j .
.11.362) .184 t 151. A 95% confidence interval for the difference between average lifctimes is (178 .140 COMPARATIVE EXPERIMENTS ( 178)2 V ( X )= = 1584. Under the null hypothesis that p .40)= 1. I _ (362)’ = 4368.8? < .1. . 30 V W ) + V( Y) and at (Y = 5 % . A random saniple of 200 parts from supplier A contains 30 defectives.64.8.1 .1 < 8. The confidence interval is 362 I78 8. 0 3 . .2 335 < 8. the common value p is estimated by pooling the two samples: .) = = 2. Arc the two suppliers producing at thc same rate of defectives? We can act as if the fraction defectives in the two samples are normally distributed and construct a test and a confidence interval along the lines of section 8.05(60. < 33.362 V( I.96d1584. = p t = p . COMPARING BINOMIAL POPULATIONS A manufacturer receives supplies of a certain part from two suppliers and wishes to compare them. B random sample of 100 parts from supplier B contains six defectives.2 + 4368. Suppose that the first sample contained x defectives out of m parts and that the second contained y defectives out of n parts.2. Method 2 F = = 2 . the estimates of the binomial parameters are We recall that the variance o f thc number of defectives in a sample of size n is npy.38. we would decide that the reliabilities differed.2 20 Z= and 178 . to be compared to F.
m = 200. .0. @ l q .1. dcf.Y6V"DZ € pI .06) . + m n Example 8.11. n = 30.OOW54. / m 0. Pooling the samples.1) For the confidcnce interval. = @ = 36/300 = 0.02 € p .96v(V(x/m) V ( y / n ) ) .11. m ) ( 1 / 2 0 0 /loo) For the confidence interval..1.16 . we no longer have to assume that p .12. For this data set. Z = 0.1. y = 6 .15.12. whence (8. 8.p 2 < 0.p 2 + m n X Y < .15 . A 170 30 B 94 6 .06 = 2..15 and b2=0. n = 100.26.0.. and the = interval is (0. $&/n = O.0006375.07 0 0. +1 ~ ( 0 . = p 2 . 1 2 ) ( 0 . CHISQUARE AND 2 X 2 TABLES A commonly used alternative to the ztest of the previous section is to set up the data in a 2 x 2 table (a table with two rows and two columns) and make a chisquare test. 0. We take the best estimates of the variances: v (i) "4' m = and V ( y )= A n d 2 A twosided 95% interval is given by X _ _ _ _ 1 .09 + 0. we take V(x/rn)= @ d / mand V(yln) = a a / n .p z < 0. 9 6 ~ ( V ( x / m )V ( y / n ) )< p l . For the data at the beginning of this section.06.CHISQUARE AND 2x 2 TABLES 141 In constructing the z statistic. @. the table would be Supplier nondef.
wc can denote thc cell frequencies by a u .5 48. +6.7 55. nondef.1. After the treatment. 'I'wcnty received the standard deposition.9 46.h ~ ) ~ ~ ' EXERCISES 8. respectively. there is only one degree of frecdom. E=176. with deviations +6.+ .7 63. we have for that cell 0170. and the probability that it came from supplier A is 200/300. the probability that it would be a nondefective is 264/300 = 0.4 48.+ .7 43.5 31.2 44. h .3 Control 59. The chisquare statistic may be calculated by the formula: U= (u + h)(c + d ) ( u + c ) ( h + d ) N ( u .6 Similar calculations show that the expected values in the other cells are 88.1 54.h c ) / N .1 1 = z L .9 52. all 40 wafers were processed together and the probe yields were measured. 0 E = . 176 88 12 24 In the general case.+ . dcf * A 170 30 200 H 94 6 100 264 36 N=3W If the probability of a nondefective were independent of supplier. and d: c cr+c h+d d h a+h c+d N=a+h+c+d The deviation in each of the four cells is r ( u d . The deviations all have thc same absolute value.88)(2/3) = 176.6 31. thc expected number of nondefectives from A would be 300(0.I42 COMPARATIVE EXPERIMENT'S We can add the marginal totals and see immediately that if a part were to be chosen at random from the combined sample.3 66.3 45.4 62.88.1 55. Thus. Twenty wafers were sent to a vendor for a boron transplant.4 46.2 34.4 . The value of the chisquare statistic is I/ = 36 36 36 36 .7 57. c. 'I'hc yields were 54. 6. and 24.= 5. 12.
1 13.6 12.9 12.3 12.4.9 17.7 12.2 12.7 12. s = 2. Test the hypothesis that the percentages of voidfree units for the two vendors are equal. 3(Hx) Angstroms thick. The specification calls for 0 to be 325 hours o r longer. Eighty failurcs were investigated and the mean time to failure was 282 hours. During wafer prcparation. An o i l company is considering introducing a new additive in its gasoline that it hopes will increase the mileage per gallon.~JS that they run on thc regular gasoline with additive. In a comparative tcst of two vendors.7 13.6 Boron 52.6 48. speed (frequency of operation). Test the hypothesis H. the following results were obtained: For vendor A .: 0 =: 325 against the alternative H. The mean times between failures (MTBF) of an electronic component have an exponential distribution. 8.8 13.1 13. Deviccs from two manufacturers were said to have the same nominal Their observed frcqucncies in MHz were A: A and 25 from manufacturer B were tested at room temperatures. < 325. Denote the MTBF by ti. 32 out of 45 units tested were voidfree..0 14.2.5 56. Twentytwo dcviccs from manufacturer 16. s = 1.9 Do the devices from the two manufacturers have the same avcrage speed'? Make dotplots of both samples on the same scale.: 6.7 12.2 55. A summary of their results is: Control average mpg = 31. and Duckworth's test.9 13.5.3 16.8 17.5 12.8 13.3 52.5 52.1 34.4 84.7 14.9 17.6 R : 13.8 13.0 13.2.7 12.7 51.7 13.4 49. Wilcoxon's test.3 Was the boron implantation effective? 8.8 75.6 14. 39 units out of 76 were voidfree. The engineers in the research group take ten cars and run them on the company's regular brand and another ten C.9 16.EXERCISES 143 54.5 9.0 16.2 13. For vendor B.0 13.0 13. 8.7 10.8 82.5. 8.4 Is the new additive effective in increasing the mpg? Is there greater variability in mpg with the new additive'? .6 9.2 13.1 61.4 18.2 17.2 84.9 12. a layer of gold. lJse thrcc methods for this exercise: the ttest.7 57.8 Additive average mpg = 32.1 14.7 11.4 12.3 54.6 16.3 15. is sometimes sputterdeposited onto the back of the wafer.5 74.6 13.3 49.8 49.5 16.3.4 13.
Components in the production machinery wear out. Some of the noise we cannot eliminate. which we can call noise. John Copyright 0 1990 by John Wiley & Sons. Indeed. which is statistical jargon for saying that a deviation of a few degrees in the oven temperature will not seriously affect the palatability of the product. it will not be cost efficient to invest in the highestcosting ovens. We also assume that we have some measure of the quality of each widget. These measures may be discrete.Statistical Methods in Engineering and Quality Assurance Peter W. out o f which comes a product. as in chapter 144 . we can eliminate almost all thc noise. Whether we like it or not. and there is variability in the raw material.1. if you are selling cake mixes to be made into cakes in the buyers’ own kitchens. INTRODUCTION A great deal of the industrial production in this country comes from production line worka method of production that calls for carrying out a repetitive process again and again and again. No mattcr how hard we try. that are identical with one another and each conforming exactly to the standards that we had in mind. you would do well to have a mix that is robust to the fluctuations in ovens from house to house. we cannot construct a process that will continue forevcr t o produce absolute clones. or within a tenth of a degree. although we can reduce it. we cannot persuade the hens to lay eggs that are absolutely uniform in size or the cows t o produce cream that does not vary in the percentage of fat from day to day. We are going to assume in the next few chapters that we havc a production process. It is convenient to say that wc are producing widgets. M. ‘The latter cost more. both in initial capital investment and in maintenance. In making a cake on a production linc in a bakery. Perhaps. Unfortunately. if we try hard enough. each production process has built into it a certain amount of natural variability. It would be nice to think that once a production line has been set up. called the product. If the cake recipe is robust to temperature changcs. We can construct ovens that maintain a desired temperature within 5 degrees. and out of the other will come a steady succession of objects. all wc have to do is to put in raw material at one end. Inc CHAPTERNINE Quality Control Charts 9. life is not likc that.
If the process has gone out of control.2. there is a stir in the shop. What happened in those days in too many shops. They continue to produce widgets as long as the items look satisfactory and there is not a preponderance of complaints from the clients. The basic idea is that wc take a sample of four. Production continues until it dawns on someone that they are now producing unacccptable widgets. at least. Perhaps they introduce inspection of the outgoing product so that they can. “tweaking” this knob o r that. which seems to improve things. These charts are then used to tell us whether or not the process is in control. is that people set up the machinery and start production. the manager gets a signal to take action. he wrote a book entitled Economic Control uf Quulity of Munufuctured Product. In 1931. and still does. that are difficult to quantify. In this chapter. gradc 2 goes to the grocery stores. . In the jargon of quality control. Perhaps they make some ad hoc adjustments to the machinery. Then each widget is classified as either acceptable or not acceptable. We then prepare two charts. That is a more complicated situation than we want to discuss at this time. such as palatability. On the other hand. we could grade each cake on a scale of 1 through 4. We can get into arguments when we try to measure characteristics. QUALITY CONTROL CHARTS Statistical quality control charts werc introduced about 60 years ago by Dr.QUALITY CONTROL CHARTS 145 three. we discuss quality control charts for continuous variables. cut down on the number of unacceptable widgets that leave the premises. We plot the average length of each sample on one chart and the range of each sample on the othcr chart. What Shewhart did was to introduce a simple graphic procedure for monitoring the output of the process that enables the manager to tell whether the process is in a state of control. The measure of quality is something like thc weight or length of the widget. widgets off the line at regular intervals and measure their lengths. The clients in this sense may be the actual end users or they may be the people in another plant some miles away who are using the widgets in the construction of some more complex pieces of equipment. At that stage. defective or nondefective. It is often rather haphazard. or five. it is the subject o f the next two chapters. and grade 4 is inedible. we clearly have the attribute case. grade 3 is sold at a discount to college residence halls. 9. conforming or nonconforming. that situation is described as sampling and classification by attributes. a scientist at Bell Labs. Walter Shewhart. If we either accept each cake or reject it. Grade 1 cakes we sell to the best restaurants at the premium price that they deserve. the research arm of American Telephone and Telegraph Company.
such as every hour. we say that the process is under control. The upper control line is at a distance 3 a / v 5 above the center line. Under the null hypothesis of control. or at our estimate of that distance. or the variability may be too great. If a point lies either above the upper line or below the lower line. or once a shift. From the point of view of hypothesis testing. I t could have happened by chance. we carry out a test. is normally distributed with expectation p and variance a 2 / n . Some cngincers call each sample a rational subgroup. We stop production and. The piece of paper on which the sample means arc plotted has three horizontal lines. the probability that a point falls outside those limits is only one in 400. Thc lower line is 3 a / G below it. How do we make adjustments to get the process to meet specifications? That is the topic of proccss improvement (offline experimentation). Is it moving along steadily? 9. The first we have mentioned earlier: a certain amount of random noise occurs naturally in any production process. We plot the sample averages. In the TJnited Kingdom. 'The center line corresponds to the process mean. that the product being manufactured and the customer's spccilications are some way apart. The other is that in a random process.3. it is morc usual to place the outside lines at a distance 3 .along the vertical axis. in the jargon. It may only have been a passing phenomenon that is easily explained and is not expected to recur. The area outside the outer control lines is the critical area. we seek an assignable cause. For the present. therc is a ccrtain amount o f regularity. along the horizontal axis. when we look closely. we have a null hypothesis that if the process is under control. As long as the plotted points fall between the two outside lines. we take a random sample of n widgets from the production line. we are concerned with the stability of our process. Commonly used values of n are 4 or 5. This cause. usually denoted by X. may turn out t o be something serious that calls for correctivc action such as replacing a part of the machinery or an operator who has not bccn properly trained. With a normal distribution of noise. 0 9 u l f i above and below the center line. which is to be discussed in the final chapters. X. with time. A process is said to be in (statistical) control if it is performing within the limits of its capability. We may find. that the capability of the process is not what we would wish. If the probability . but only about one time in 400. or rejection region. the random variable. With each sample. THE XBAR CHART At regular intervals.146 QUALITY CONTROL CHARTS Shewhart's ideas depend upon two statistical concepts. We can now give a definition of control in this context. only one time in twenty will a variablc differ from its mcan by more than two standard deviations. we say that the process is out of control. or sample number. The average length of our widgets may not be what the customer wants. when found.
and a simple action.06 for n = 4 and 2. We recall that d2 = 2.729 0.1. Everyone is involved. can see how their process is doing. It provides our estimate of p .880 1.933 . We now have a simple graphical procedure that shows how the process is behaving.14.1 Table 9.fi The upper and lower control lines are then set at Values of d 2 and A .059 0. the procedure is a little different.we are talking about the mean and variance of the process as it actually is and not about the specifications.4. as in section 6.. Factors for the xRar Chart n 2 1. It can be divided by d . and called xbarbar. or decision.023 0.975 S 6 7 2. are given in table 9.326 0.992 2. the probability is 0. rule. One of the workers is assigned to make the entries. The standard deviation has traditionally been estimated from the ranges. to obtain an estimate of the standard deviation.704 0.4. LINES 147 distribution of X is really N( p. . Their grand average is denoted by R . In practice. We define another multiplier. A. The ritnge.SETTING THE CONTROI. The chart can be kept on the shop tloor.= 3 d.534 d43 0.419 0.9% 2.33 for n = 5.4. Everyone can focus on the problems of quality.0025 that we will mistakenly stop production and that there will be no assignable cause to find. a’ln). and fixes the center line of the chart. Its simplicity is one of its great virtues. Then we can multiply that estimate of the standard deviation by 3 / f i .911 Eff * 1.OOO 3 4 A: 4 2.128 1. The average of the individual X is denoted by f.693 1. Everyone. workers as well as management. R. A pilot set of about 20 rational subgroups usually suffices to get started. We have to estimate those two parameters from a pilot set o f data taken from actual production. SETTING THE CONTROL LINES When we speak of X having expectation p and variance cr2/n. of each subgroup is recorded. 9. That figure is the alpha risk.577 0. The two steps are combined into one.
where c. The unbiased estimate is SIC.7758~’. is now shown for the case of n = 4. Values of D. For each sample. (upper line) or D3 (lower)..5.5.148 QUALITY CONTROL CHARTS Alternatively.12.12. but the concept of range as a measure of variability is casicr to understand than the concept of standard deviation.183~’ and sd(w) = 0 . in an instant. one computes The square root. The sample range. This is a diagnostic tool.1. The upper and lower lines are obtained by multiplying K by D.282 0 2.004 D. The center line is set at R. The asymptotic value of l/c. Those efficiencies are also given in table 9. 0 3.5. A similar chart of the sample range keeps track of changes in the variance. It can be shown that if the observations come from a normal population. The chart is meant to give everybody a picture of what is happening. is not an unbiased estimate of the standard deviation.1. Not only can the worker on the floor easily calculate the range of a sample of four observations. The simplicity should not. R. however. It can be argued that this advantage is less important today than it was 25 years ago. for u in the latter expression gives V ( w )= 0.114 . there are three control lines. are shown in table 9. Thc primary advantage of using the range to estimate u is its ease of calculation. every modern engineer carries a hand calculator that can compute s’ and SIC. s.12. The range estimate for u has very high efficiency by comparison to s/c4. Again. Charts based on S are the topic of section 9. we want everybody to be able to relate to it. 9.4)’. E ( w ) = 2 . ‘Table 9.1.267 7 0. is a multiplier whose values are given later in table 9.and D.076 1. is plotted along the vertical axis. be dismissed. Denote the range by w .574 4 5 6 0 2.. and D. 0 5 9 = d 2 aand that ~ V ( w ) = 0.Substituting w l d . Multipliers for the Range Chart II 2 3 0 2. The rationale for computing multipliers D. for larger n is 1 + ( 4 n . 4 2 8 ~ . one could obtain an unbiased estimate of the variance from the squared deviations in the usual way. D.1. the average from the pilot set.924 0 2. THE R CHART The xbar chart keeps track of changes in the process mean over time.
428) w .82) = 8.3sd(w) are the correct limits and that w is normally distributed.729)(1.50.2.3(0. the average range is 1. The charts of the pilot set are in figures 9. To compute D. However. the upper and lower control lines are set at 10. this is the rationale that has traditionally been used.82.1. The data are in table 9. Had that value been. which is not true.729)( 1. The grand average X for the 20 groups is 10. and a process that produces outliers has problems. The R chart is obviously sensitive to outliers. if the variability of the process is out of control. whether you like it or not. even though inflated ranges may make the control limits on the xbar chart far enough apart to include all the points.13 . 9.1 and 9.we should take D. The only point that comes close to an outer control line is the seventh. A single bad observation can make the range of that sample so large as to exceed the upper control line.284 and the actual value of D. It can be argued that the engineer should look at the R chart before looking at the xbar chart.= w . For the xbar chart. Remember that. It is possible that the search for an assignable cause will find a reason that could justify the engineer in throwing out that point and recalculating the estimate of CT without that sample. The difference between the factor 2.6.6.428)w = 2 . AN EXAMPLE The following pilot set of 20 subgroups of four observations each was used to set up xbar and R charts for a process. That should only be done if the engineer is very sure that it is justified. There is a tacit assumption here that w 3. The process is clearly rolling along steadily.6. 2 8 4 ~ .13 + (0. but this quantity is negative. which has a sample mean 11. it is an observation. instead.82) = 11.(O. the process is certainly not under control. Following the same procedure as for the xbar chart. = w + 3sd(w) = w + 3(0.80 . as manifested by the R chart..13. and so we replace it by zero.6. even when the observations are normal.A N EXAMPLE 149 where sd(w) denotes the standard deviation of w. 11.45 and 10.30. we .282 is due to roundoff in the calculations. we set the upper control line at D. The distribution of w is certainly not symmetric. = 2.
9 11.4 10 10.43 A 10.6 10.4 3 10.10 1.7 9.6 10.28 2.30 0.8 17 11.7 10.63 1.2 10.4 10. Twenty Samples 1 10.0 X  9.9 9.4 0.1 QUALI’I’Y CONTROL CHARTS 2 10.4593 11.5 10.9 10.1 9.8 10.Y 9.90 1.6 11.20 2.3 10.5.4 11.50 10.9 0.1325 9.2 8.85 1.6 10.6 115 11.4 10.6 8.9 9.0 8.2 X R 10.3 2.9 13 10.53 10.8 8.6 10.3 9.1 14 10.00 2.2 8.7 10.7 9.6.0 12 11.0 11.3 9.1 9.3 11.5 11.2 9.0 9.50 3.8 8.7 8 10.8 4 10.5 8.1 11. ‘Ihc xhnr chart. .60 0.1.1 9.9 7 11.1 10.4 1.9 15 9.8 8.6 18 11.3 9.6 10.2 11 11.7 10.01) 1.2 20 10.3 8.5 0.2 10.1 95 0.0 8.5 10.1 5 8.7 11.8 9.2 11.70 1.1 16 8.6 9.8 10.1.6 9.80572 Figure 9. 1.3 9.83 2.6.8 0 8.9 11.5 H 2.7 YM .6 11. 10.2 11.8 7.0 10.0 9 9.8 2.7 10.5 10.5 10.5 10.7 11.9 11.8 9.3 10.5 10.13 1.15 1.45 9.5 19 10.7 9.8 i v1 fd 10.7 10.150 Table 9.
be due to a new operator who was a bit heavy handed until he acquired the knack.7.282)( 1. the operator .15 . an example of a process that is manifestly out of control. on the othcr hand. 16 would have noted that it was above the upper control line and proceeded to look for an assignablc cause.ANOI'HEK EXAMPLE 151 0 4 8 12 Samplc number Figure 9. There are numerous points outside thc limits on both thc high and low sides.6. It could occur by chance. for examplc. It could. the lower limit is zero.7. The raligo chart. All thc points fall between the limits. in this case. the (supposed) problem sccms to have bccn temporary. ANOTHER EXAMPLE The process that is charted in figurc 9.82) = 4. The process has iin unfortunate tendency for the mean to increase. One outsidc point in a string of 20 points may not signify anything serious.2. As soon as this is noticed. The upper limit for the range chart is taken at (2.1 is. 9. Ccrtainly. One might suspect the following scenario.
it could also include differences between batches of raw material. . In some cases. seemingly.1 9. who might adjust the machine in different ways when they start their work turn. The xbar chart. and usually overdoes it.152 12. But the variability between shifts could include differences between operators. It is a measure of the variability of four items made on the same machine on the same shift. What if the outside points had not occurred in such a systematic pattern.1 0 4 8 12 Sample number 16 20 Figure 9. tweaks one of the variables to bring the value of the mean back down. at random among the 20 points recorded? How might such a phenomenon be explained? It is important to recall how the estimate of the proccss variance was obtained. Wc obtained it from the ranges of the individual samples.1 10. but came.1 QUALITY CONTROL CHARTS 11. The process drifts upwards and is “adjusted. It does not take into account variability between shifts.7.1. and other external sourccs of variation.” One piece of evidence to support this theory is that points above the upper line are immediately followed by points that are markedly lower. the variance that we have measured will be relatively smallthe variation between four items produced on a precision machine under virtually identical circumstances.1 8. Then another cycle starts.
8.23 11.93 10.3 8.28 10. After the thirtieth sample (the 20 in table 9.23 10.88 11.8. I. 11.90 11.25 10.70 11.8 .60 11.1 and 9.1 and the first ten in table 9. Thirty more samples have been taken and added to the previous 20. DETECTING SHIFTS IN THE MEAN Figure 9.18 10.2.8.6. Values of X for 30 More Samples 9.3 10.18 10. we discuss alternativc stop rules and a criterion for assessing them: average run length.8. something happened to the process.60 11.50 10.75 10.1).58 11.30 11.00 9.35 10. is that an unrcasonable delay in detecting a change'! In the sections that follow. That happened at the fortyfifth sample.6.6.63 10.13 10.DETECTING SHIFTS IN THE MEAN 153 Table 9.30 10.08 11. there was a value of X that exceeded thc upper limit.88 10. After a while.95 9.1 is an xbar chart for the same process that produced figures 9.8 11.8 9.1.95 9.8.65 9. There was an increase in p of about one standard deviation. Their averages are given in table 9.25 11.65 10.3 9.8  P (] I P 10.00 10. 15 samples after the change had taken place.23 10.
8. Rule 2. These lincs are used for the second rule. This leads to two classes of rules based upon runs. ALTERNATIVE STOPPING RULES The term stopping rule is used by statisticians working with stochastic (random) processes. In the data of table 9. If our process is behaving as we expected.1. 0 There are also rules that are based on the simple coin tossing model.6. One point outside the three sigma lines. Since they should occur at random. 9:lO. we spoke of the alpha and beta risks. points. Too few runs might indicate a slow cyclic trend in the process average. A run of seven. there is a run of five points below the center line from sample 8 through sample 12. either all above the line or all below it. AVERAGE RUN LENGTHS In hypothesis testing. Two consecutive points. A run is a string of successive points. In figure 9. this rule would have been triggered after the thirtyfifth sample.016. roughly half the points should be above the center line and half should be below it. we would fail to detect it. The probability of a run of seven is 2(1/128) = 0. 0 Some engineers draw two more lines on their charts at a distance 2.9. either both above the upper warning line or both below the lower warning line.1.154 QUALITY CONTROL CHARTS 9. alpha is the probability that we (wrongly) trigger a stop rule when thc process is still in control. too many or too few. Onc could have too many runs in the extreme case of two operators working alternate shifts if one operator consistently made the widgets too long and the other consistently made them too shortthe sawtooth effect. the situation is the same as a sequence of coin tosses with heads for points above the line and tails for points below. We choose the . There are also rules based on the number of runs. beta is the probability that if there had been a shift in the process average. or more. In the context of control charts. The R chart has a run of four above the center line from 9 through 12. either all above the center line o r all below it. We have been assuming that we are dealing with a random normal process with mean p and standard deviation LT. Rule 3. and call them inner control lines. Our only stopping rule so far has been Rule 1 . A stopping rule is a rule for making the decision that the process no longer satisfies the original conditions. or warning lines.Oulv'Z above and below the center line.
As long as the process remains in control.rtion of X. the process average changed between the thirtieth and thirtyfirst samples.4 0. Now the proccss is operating about a new centcr linc that is only 2.0 1.1. In figure 9. on the average.1. How long a run will it take. the number of samples until one exceeds the limits has a negative binomial distribution with p = O.3 1.02275 and thc average run length (ARL) is 44.Y 2.AVERAGE RIJN LENGTHS 155 test that has the smallest value of beta for a given alpha.00135). We assume that our estimates of the mean and thc standard deviation are corrcct. the emphasis changes slightly to a diffcrent criterionthe average run number. but is 4 t r l f i above the lower line.8 Rule 1 310 200 161 122 Rule 2 799 349 238 165 93 72 56 44 35 28 22 18 117 84 62 46 0. since the distance to the upper line is now reduced to one standard devi. Using rule I . the chance that a point will be outside thc control lines is 2(0.4 1 . The chance that a point will now fall below the lower line is virtually zero.1 1.0vlVii below the upper line. but the chancc that a point will fall above thc upper line is 0. Here it refers to the number o f samples that are taken before thc rule is triggered. Thc change was dctected in the fortyfifth samplea run of IS samples.10. the avcrage run lcngth whcn thcre is a zero shift in the mean is 370.0 12 I0 9 7 6 27 22 17 14 12 LO 8 7 6 . Table 9.H 35 IS 1 .7 l.2 0. to detect a change in the mean? Wc consider rule 1 in some detail. With control charts. For a shift of 2ulv'Z.3.7 0. the ARL is reduced to 6.6 1.9 1.5 0. .6 0. Average Run Lengths to Detect a Shift d(ol\/sz) in the Mean d 0. The word run is used in a different sense from the previous section. Suppose that the mean shifts from p to p t ~ + l v ' X while the standard deviation remains the same.8.s 1. If the samples are independent.(X)27. Its expectation is 1 lp = 370.2 1.
. the relative efficiency of the range estimate to s is close to unity. Thc center line is set at the average of the sample deviations. It can be shown that the variance of s is given by the formula V ( s )= (1 . using rule I and rule 2. and so the standard deviation of s is 0.1. are obtained from tables. for (r and set the upper control limit at 0.1. l N . 9. three standard deviations above and below the center line.3.266. However. we have used the R chart for the control of variability. The formula is ( I + p ) / p 2 . but on the sample standard deviation.. the engineer sets the limits at UCL = B. the sample standard deviations.5 that charts can be based. B.4). are given. from the N samples in the pilot set. c. now We substitute f/c.12. for several values of the shift in the process mean.S and IXL = R 3 i . 9.. The control lines are drawn. 0. . arc plotted along the vertical axis.11. are given in table 9. such as n = 4 or tz = 5 . as usual.9213.12. = 0. This is shown later in table 9.156 QUALITY CONTROL CHARlS The ARL for rule 2 can be calculated by setting r = 0 and m = 2 in formula (13. if we wish to detect small shifts in the mean within a reasonable average run length. S= C s . For n = 4. the calculations are casily made. we can use larger samples. where values of c. not on the range.' . and S. where p is the probability of a point outside a warning line.f. s CHARTS So far in this chapter.Ci)(. s. and B. Some values of c. With modern computing facilities. and little is lost by using the method that is simpler arithmetically. We mentioned in section 9. Table 9.10.9213 In practice. One can also make charts based on the sample variance in a similar way.12. For small samples. in which case the relative efficiency of the range decreases... SETTING THE CONTROL LIMITS FOR AN s CHART In these charts. s.389 UCL = s + 3( )s = 2. where the values of B.1 contains average run lengths.389~.
the upper and lower 0.552 1.510 B.0148 1. 0.0157 1.0854 1.354 0.716 1.992 0.9400 0. 0 0 0 0 0.5% values of chisquare.0252 1.0638 1.594 1.406 0.9835 0. after taking square roots and simplifying.751 0.0510 12 & .766 0.975 0.955 0.9869 9.482 0.239 0. therefore.266 2.9693 0. assumes that the distribution of s is normal. ALTERNATIVE CONTROL LIMITS In the previous section. calculate better limits in this case because we know the exact probability distribution of s when the data are normal.321 0.f.831 0. respectively.12 the procedure for obtaining a confidence interval for the variance of a normal population.9754 0. derived from the statement 0.9213 0. 3 I 1.8381 . with 3 d.0194 1. As in the R chart.503 1.185 1.970 1. it is the standard procedure..267 2.Y794 0.12.13.1284 1.0317 1.646 1.118 0.1546~ s < 2 . With n = 4.466 0.568 2.9862 0.OM8 1.572 1. The reader will recall from section 6.8381.738 0.725 0. A twosided 9Y% confidence interval for s is.2533 1.7979 0.534 1.Y823 0.679 1.0180 1.0363 1.407 0. 0 6 8 ~ < . this procedure tacitly.518 1. the control limits were calculated as the center line plus or minus three standard deviations.0210 1.185 0.781 0.ALTERNATIVE CONTROL LIMITS 157 Table 9.0140 1. Nevertheless.9594 n.490 Efficiency 1.8W 0.9727 0.9845 0.850 0. Constants for s Charts n 2 3 4 5 6 7 8 Y 10 11 12 13 14 15 16 17 18 19 20 c 4 llc.96so 0.911 0.9854 0. and wrongly.Ooo 0.07172 and 12.07172q y U 3s2 < 12.098 1. are 0.1.882 1.712 0.761 1.869 0. however.448 0. Then.284 0.814 0.428 0.8862 0. 3.1133 0.700 0. 1.9776 0.0229 1.9810 0.797 0.9515 0.0133 B. We can.618 1.0281 1. . We now develop these limits for the case of samples of size 4.030 0.382 0.
24%.158 QUALITY CONTROL CHARTS We now replace u by . thus. 9. One should not. The robustness of control charts under nonnormality has been investigated by Schilling and Nelson (1076). We saw in chapter five that the distribution of the sample mean X approaches normality as n increases. the distributions of R and s are not normal. The situation with the R and the usual s charts is different. and it still serves as a powerful (and simple) diagnostic tool. however. UCL= 2. Our emphasis in process control has been that the process should be rolling along steadily. the xbar chart is robust to moderate departures from the normality assumption.245i and LCL= 0.168. 9. A process can. The control limits are. has to fall in thc interval . one in 200 rather than onc in 400. Even if the actual alpha risk is a little cloudy. PROCESS CAPABILITY Nothing has been said yet in this chapter about specifications. closer to the center line than the usual limits. How cssential is this? It is not serious as long as the distribution is roughly normal.f. Even if the data are normal. the risk of a falsc stoppage may increase twofold or more. Even for sample sizes as small as n = 4 and n = 5. They are measures of proccss capability. the principle of the R chart is sound. Instead of having only one chance in 400 of a point outside the limits under normality. and so the rule o f thumb of drawing the control lines at plus or minus three standard deviations does not pass muster mathematically. be under control but be way off specification! Two criteria tic process control and specifications together. in theory. The measured characteristic. such as the length of the item. and the lower limit is positive.14.168s < s < 2.15.fl0.?lc4 .0213 and obtain 0. We consider the case in which the specification is twosided and symmetric. Notice two things: the limits are tighter. overreact to this. NONNORMALITY The derivation o f the control charts depends on the assumption that the observations are normally distributed.
C.. only 84.73%)of the items will fall in the intcrval ( p . C. = A desirable aim for a process is to have C. respcctively.1JSL .73% of the items will be within specifications....6. but 2. the percentage of items within specifications will fall below 99.. in some Japanese . p Z &. If p = + 2 c . so that the process mean is indecd at the target value..  6ff ' where USL and LSL are the upper and lower specification limits.k ) . the specification limits will be LSLp40.& 1.k C. p + 6).. CPU) = C. howcvcr.p CPU= 3a and CPL= p LSL 3ff If p = p o = (USI. where If the process is correctly centered..LSL. as CT decreases. k = 0 and C/.u.PROCESS CAPABILI'I'Y I59 p. and USL=p+2c Virtually every item produced will be above the LSL.( 1 . both these indices are equal to The second criterion is C. + I. If.73%.27% will fall above the USL... called the target value of the characteristic and 26 is the spccification is width. p = pn+.. = minimum (CPL. It was reported a few > years ago that the minimum acceptable value of C. If. If the process is on targetif p is actually equal to the target valuc p. The process capability index. increases. if thc process is off target.. is defined as the ratio of the specification width to the process spread: . and only 97.. measures the potential cability of the process.. We now introduce another measure of process capability that includes the failure to he on target. We denotc the actual avcragc value for our process by p (as represented by thc ccnter line on the control chart).13% will meet specifications. We define two new indices USL.SL)/2. the process is not properly centered.0 has fi = 3a and is capable in the sense that 99. = 1.)a process with C. C. for cxample.0.
Supposc that the sample averages in table 9. and the center line is at x = 10.50 9 9.85 10.80. Then rr = is the range of the ith pair. 5 ~ .13.5.4 defective parts per million.50 13 10.50 = 1. is called the sequence of moving ranges. are r .30 10. The control lines are drawn at a distance t 3 J l 1. denote the value of the ith individual.63 19 11. which corresponds to a fraction O. in the semiconductor industry. n = l? Such a chart is called an I chart (individuals).26.6. The average of thc moving ranges is F = 0..46.128 = 2.20. LCL=8. I CHARTS Thus far we have discussed setting up xbar charts when the sample size is greater than n = 1.83 9.0 and Cp.60 The first two values of r.+. .160 QUALITY CONTROL CHARTS industries is 1.28 10.10 17 18 10.90 6 7 8 11.1 had been reported as single observations without any information about (r. control limits are set at UCL= 10.45 11 12 9.l28.43 9. which represents about 3. The standard deviation is estimated from the moving ranges of pairs of adjacent observations.10. 4 Example 9. Therefore.28 = 0.15 10. There are two choices for estimating the standard deviation. Nowadays.70 . b..e.ooOO6 (6 in lO0. The estimate of Q is given by i / d 2 = F/l. manufacturers are striving for processes with C.70 i 1 2 3 9.42 and rz = 10.70 .l = 1 . What can be done when the samples are individual observations.53 10.13 4 5 9. We have a set of twenty individual data points: X 10.629. 9.OOO) of nonconforming units. The sequence of the values of r.66(0.LSL = 12a and reducing the deviation of the mean value from target to l p . This can be achieved by reducing the standard deviation until USL .629)= 11.60 15 16 10.13+2. = 10. Denote its average by J. = 2.00 20 9.20 10 9. Let x . i.9.p. either to use the sample range or the sample standard deviation.. = 1.00 14 10.33.66f from the center line.2.16.
22. each with three observa tions.01 25. 25.06. 24.80 25.O 20.84 25.24 25.58 24.35.32.456 I I I i It will be recalled that the control limits in the original xbar chart were set at 11.80. 21. The following data set shows 30 samples.10 25.11 25.61 27.27 25. 25.15 24.76 25.70 2851 23.83 28. 20.45 and 8.00 0.57.EXERCISES 161 UCL = 11.71 23.05 21. 26.Yt) 23.75 *  * * * * * * * MU = 10.86.82.22 23 .54.14.31 25.76 23.51 22. 23.26.29 21.13 22. 25.02 22.39 23.13 LCL = 8.02. 23.17 25.32 23.50.04 26. 26.75. 22. 26.16.12 23.76.95.96 23. 23.62.23 26.80 11.52 25. 25.77 20.10 24. 22. EXERCISES 9.55 18.70 26. 22.60 22.83 25.10 23.77 22. 26.23. 21 33. Make an xbar chart and an K chart for the data and mark any points that are out of control. 24.25 * * * * * * * 2 V u I G 2 m 10.42. 23.73.48 23. 24.22 25. 28. 28. 27.64 23.99.48 24.19.53. using all three stopping rules.82 21. 23.94 24.52 22.84 25.41.69 25.02. 24.38 29.72 23.1.1.86 23.61 21.55.88. 0 The Minitab I chart for this data is shown in Figure 9.78 25.16.59 26.76.10. 22.9 1 25. 22.54 23. The first three samples are in the first row.39 23.37. 22. .57 25.
6 63. 65.0 17.8 67.5 18.6 65.7 17.6 17.7. 15.3.0 19.6 17. Samples of three items are taken from the production line at thc cnd of each shift for ten days (30 samples).8 62.8 62.6.0.0 18.5. 18.7 18.4.4 18.1 18. 67.5 18.8 64. 62.4 65.0.Y 16. Make an xbar chart and an R chart for the data and mark any points that are out of control.6 17.2 17. 64.6 65. When the process was under control.0 60.4 17.6 63.2 64.2 66. 67.0 63.2.9 17.4 66.5. 62.2 18.6 63.4 16.6: 9.0 67.4 62.0 65.0 64.7.0 QUALITY CONTROL CHAKTS Repcat exercise 9.8.0 69.0 65.8 18. 37.1.2 64. the process average was 64. 20.8 62.2 16.7 20.2 17.4 65.2.0 18.9. 64. 16.8 62. and so on.4 19. 66.6 66.6 61.5 17.5 18.8.8 64.1 20.5 16. 65.4 17.8 62.8 17.6 63.0 18.8.6 64. using all three stopping rules.2 61. 64. 18. 62.2.1 assuming that you were given standard values and = 1.4 65.8 63.4.0.0 and the standard dcviation was 1.2 63. 16.1 18. This data set contains 100 observations taken from a production line in 25 samplcs o f 4 each. 64. The first two samples are in the first row. 17.4 64. 18.8 20.0 16.8.4.4 18.4 16.8 60. The first row contains thc first three samples.4 65.5 18. 18.6 64.2 16. 17.2.8.2.5 17. 63.2 64.6 18. 67.0 64. 64.0 65.6 67.4. 18. f .8 16. 59.4 18.5.0 62.0 65.1 17. 18.9.8 17. 65. 65.9.6 19.6 18.9.2 65.4 62. 65.5.4 17.9 17.8 66.3 17.0 66.5.2 64.5 20.4 17.0 65. 63.6 18.4 63.4 62.6.6.3 17.3 19.6. 17. 18.6.8 66.0 67.7. p = 24.7 18.3 18.5 17.2 19.6 62.8. 67.4 16.6 64.0 66.4. Do your conclusions about points being out o f control change? 9. 65.0 65. Make an xbar chart and an R chart o f the data and notc which samples are out o control.5.4 64.3 19. 16.2 64. 17.6 65. 17.6 15.8.0.0 66.4 62.0 17. 64.6 65.2 65.7 9.1 18.9 18.2 62.2.8 62. This data set consists of W observations in thirty samples o three f observations cach. 15.8. Make an xbar chart and an R chart and mark any points that are out of control.2 16.0 18. 65.5. 19.162 9.0.7 18.8 17.0.2 66.6.8 63.4 64.0.4 66. The measurements of the items are given below with the first three samplcs in the tirst row. 64.8 68. 18.2.4 60.3 15.0 63.4.
15 40.65 39.43 39. .84 39.76 40.57 41.14 40.87 39.66 39.01 39.95 40.30 39.2s 40.55 39.20 41.32 38.95 40. I4 40.57 40.1.53 40.97 40.68 41.16 40.53 41.13 40.1.14 39.92 40.84 39.55 40.93 38.54 40.15 41. Assuming that E(s) = c4w. 9.34 9.49 40.80 39.83 40.9.92 40.86 39.31 39. that E ( s ) # (r.90 41.44 40.6.38 40. Make an s chart for the data of exercise 9.01 40.c.07 39.EXEK CI SB S 163 40. prove that V ( s )= (1 .86 41.36 38.02 40..31 38.c:)w2.20 40. 9.41 40.67 40.73 41.80 41.97 40.79 39.60 40.69 40. Prove that s is a biased estimator of (r.10 38.10.79 40. Make an I chart for the data in table 9.47 39. 9.29 39.47 39.80 39.56 40.18 39.8. 9.94 40.56 39.48 39.16 40. Make an s chart for the data in exercise 9.53 41.46 37.94 40.38 40.54 41. i.92 40.95 40.98 39.98 39.7.58 38.17 41.33 40.77 39.19 40.35 38.52 39.33 41.68 41.8.38 40.07 40.99 40.4.
in which case it is rejected. Inc CHAPTERTEN Control Charts for Attributes 10.2. Then we introduce another series of charts that use the Poisson distribution. in which case it is accepted. nogo: cither the itcm is acccptablc (go) or it is nonconforming (nogo). However. 10. or tcsting. We recall from sections 3. Nonconforming units replaces the use of defective as a noun.12 and 6.of 164 .. we can reasonably act a if the number of nonconfornis ing units has a binomial distribution with parameters n and p . BlNOMIAL CHARTS We assume a process rolling along under control. wc present control charts based upon thc binomial distribution. INTRODUCTION In chapter nine. we discussed control charts in which the measurements were continuous random variables. John Copyright 0 1990 by John Wiley & Sons.S. An item is examined and classified as either conforming.Statistical Methods in Engineering and Quality Assurance Peter W. Another term used is go. producing parts of which a proportion p do not conform. In this chapter.3 that if the inspector takes a random sample of parts. where n is the number of parts sampled. terminology more in line with international standards. An inspector draws a sample of size n. Those are known as control charts for variables o r control charts for mcasuremcnts. or as nonconforming. Wc arc sampling. Many engineers use the terms interchangeably. if the sample consists of only a small fraction of thc batch. and p is the proportion of nonconforming units in the batch. In order to bring U. Either the item has the attribute of acceptability or it does not. One can picturc an inspector drawing a random sample from a large batch of components and counting the number of nonconforming units. the number of nonconforming units theoretically has a hypergeometric distribution. by attributes. M. In the first part of the chapter. We can associate with each item tested a random variikk x that takes the value zero if the item is acceptable and one if it is nonconforming.1. we look at the situation in which no measurements are recorded. the words nonconforming and nonconformity have lately replaced in some industries the older words defective and defect.
3. There are two equivalent methods of making a chart. we will need samples of lo00 items! The value of p varies from application to application. The data are given in table 10. is estimated by (10. and think in terms of upper and lower control lines at a distance of three standard deviations from a center line as before. When we get the fraction of nonconforming units down to 0. Some semiconductor mariufacturcrs are working in some stages in the region of one defective unit per million.1) The upper and lower control lines are drawn at UCL = x + 3 v ~ ( . *I (10.3. Two examples arc enough to illustrate the general procedure. over a 1Zmonth period.005. is an unbiased estimate of p.1 shows a control chart for an inspection scheme in which samples of n = 150 were taken from each batch. In the second. is normally distributed. in which case it is called a p chart. It is a random variable and its variance is p4In.1. Then the variance of x. are nonconforming. The Numbers Nonconforming in Samoles with n = 150 6 4 1 9 0 3 8 8 1 5 3 8 4 6 5 1 6 5 2 4 3 4 8 9 5 4 . The process is in statistical control. printed in the Austin American Statesman on December 24. The center line is set at xbar. BINOMIAL CHARTS FOR A FlXED SAMPLE SIZE Figure 10. we can act as if p .2 )/ n n 4 (10. We will construct an np chart.3. or else one can plot x i and obtain what is called an np chart. On the other hand. i l n is our estimate of p.1) where p . The fraction nonconforming for that sample is Pi=<.BINOMIAL CHARTS FOR A FIXED SAMPLE SIZE 165 which x .3.1%. This means that if the fraction nonconforming is thought to be about 5%. One can cithcr plot p i along the yaxis. an Associated Press report. quoted an Air Force report that a certain defense contractor had. the sample size varies.1.2) Table 10. How large should n be? A commonly used rule of thumb is to take a prior estimate of p and choose n so that np > 5 . reduced its overall defect rate from 28% to 18. the average of all the x .3. If n is large enough. 1989.2. 10.3. a sample size of at least ICW) is recommended. . The first has a fixed sample size.
30 and so the lower line is set at zcro. (10.53%.76. In these two extreme cases. The center line is set at this value. but there are two close calls. Equation (10.33) If the value given by equation (10. CHAKIS FOR AI’TKIBUTES 12 B 9 8 6 3 0 Sample number Mgure 10.1. the center line would be labeled 3.7%.2) as 12.166 15 CONTHOL. which is often the case.3. The average number of nonconforming units per sample is 5.3. In the p chart. . which corresponds to approximately a 4% nonconformance rate. The fourteenth sample had 12 nonconforming units. Hinomial control chart. In this example. the percentage of nonconforming units in the samples was as large as 8% and as small as 0. The UCL is calculated by equation (10. The only difference between the np chart and the p chart is in the scale of the vertical axis.3) is negative.8. we plot the percentage nonconforming. and LCL = X 3 v m .3.84% and the UCL 8. LCL is set at zcro. All 25 points lie between the control lines.3.3) gives a value of 1. and the seventeenth only one.
Figure 10. or actually. If the method of testing a fuse is to expose i t to increased levels of current until it blows. especially when testing is destructive. VARIABLE SAMPLE SIZES Another strategy is to sample every item in the batch: 100% sampling. 100% sampling can be virtually.41. but in practicc it docs not.1 is a p chart for a process to produce electrical parts.4. and l(K)U/o testing would leave the engineer with n o product. However.VARIABLE SAMPLE SIZES 167 10. The test can merely be a matter o plugging a component into a test machine and f seeing if thc light goes on. Serious questions have been raised ahout the practical efficiency of 100% testing. or of having ii piece of automatic testing cquipment installed online. The p chart (uncqual sample sizes).4. Especially when the nonconforming rate is low and the procedure is tedious. The important difference between this example and the preceding one i s that the sample 0 5 10 15 Sample number 2 0 25 Figure 10. automalcd. with batches of about 20()0 items subjected t o 100% inspection. . thc test destroys the item. inspectors tend to become inattentive. Sometimes ICH)% sampling is clearly out of the question. in some cases. It would seem that this method ought to spot all the nonconforming units.
4 23 21 254 1.2% to 11. The usual procedure is to compute the valucs for thc outer control lines from equations (10. This gives values 0. o/o nonconf. 2015. The percentagcs of nonconforming parts ranged from 8. Control lines based on the average batch size are adequate to get the general picture.2 23 17 74 14 9 10.3 2 0 27 26 25 0 9. 18 . A chart prepared under these conditions is shown in figure 10.1. batch size nonwnf. we calculate the limits (10. 08 8.4.4 1 8 13 63 I4 5 9.6 21 2047 l2 Y 9.4 2451 6 15 90 222 1.1.2. but for the close cascs. Table 10.4. The batch size varied from day to day for various reasons.4 1901 2 2 11 98 16 7 9.1004. although point 13 on the high side and points 16 and 29 on the low side come close. For each batch. = 0. the average of the p i (some engineers use P = 1 0 0 ~ ) .1206. 19 5 20 13 17 7 8.6 19 11 1 4 21 6 1. 10 1 0 28 33 224 9. batch size nonconf. The center line is set at 0.3 Average batch size = 2015. % nonconf.4.2 8.0 2 8 18 54 19 76 29 0 1. the lower line is set at zero if this formula gives a negative value.0802 and 0.6 1 2 15 84 204 1. In this data set.9 15 7 8 10 94 12 7 9. using the average batch size. are smaller batches. the largest batch had 2488 parts and the smallest 1549. Point 13 is now out of control because it was the largest batch in the group. 06 1 2 18 84 23 1 11. 07 17 28 06 22 4 11. 100% Sampling of Batches of Components batch size nonconf. 14 26 3 21 39 I W 8.0 14 82 11 5 8. The lowest points. Thc outer lines arc no longer horizontal. 10 75 10 8 1.4 15 75 1so 8. No points are out of control.3.2 11 89 213 1.2) and (10.1004 p 2 9 2 7 18 93 24 3 1. We can make a chart that ignores the variation in sample size. we must calculate the limits using the actual sizes. 17 25 3 0 2374 171 16 5 17 9 1.2 16 2153 29 4 11.168 CONTROL CHARTS FOR ATTRIBUTES size is not constant. % nonconf.4. This variation makes the simpler np chart inappropriate. 14 14 59 13 4 9. 16 13 28 48 25 9 1. The data are given in table 10.9%. 06 9 27 11 12 8 8. 9 nonconf.1) Again.2 2 4 28 13 21 0 9. % batch size nonwnf.3).6 4 2322 25 4 1. 16 and 29.3.
INTERPRETING OUTLYING SAMPLES In section 9.2.5. Nothing was said about samples in which R was too small. If a batch has an unusually high fraction of defective parts.INTERPRETING OUTLYING SAMPLES 169 1004 0 5 10 15 Sample number 2 0 25 30 Figure 10.5.4. we made a passing reference to the question of handling outliers in the data set that is used to set up a control chart. . such as a malfunctioning piece of equipment that is replaced. because with the usual sample sizes of ri = 4 or 5 . the lower control limit on the R chart is zero. That point can then be dropped from the data set and the control limits can be recalculated. the search for an assignable cause may lead to a reason. Similar problems arise with p and np charts. The p chart (unequal sample sizes). The interest then was in R charts. An unreasonably large range can give the engineer an inflated estimate of (T. 10. This will lead the engineer to set the control lines for the xbar chart too far apart and to be lulled into a sense of security when all the averages fall comfortably between the unrealistically liberal limits.
One of the standard reasons for low points is that they have been reported by an incxpericnced inspector who fails to spot all the defects. The last sample is now bclow the UCL and is under control.170  CONTROL CHARTS FOR ATI'KIBUTES Table 10.5.1.5.0 t 20. The UCL is at np = 32.94 I * * * * np = 18.00 10.1. The np chart lor table 105.1. The np chart is shown in figure 10. The last sample has 31 defectives and lies above the LCL.1. This difficulty also occurs with ccharts. 19. Numbers of Defects in 25 Samples ( n = 150) 17 13 24 I9 17 18 24 17 20 17 23 21 15 16 31 22 19 22 19 26 2 20 3 21 4 Points that fall below the LCL may present a different problem. The average value (center line) is np = 18.0. it often happens that the low point is too good to be true. Thcrc is a temptation to greet such points with applause. np = 20. Suppose that the three points below the LCL are dropped from the data set and that the control lines are recalculated from the remaining 22 points.0 t 0 0. and 20. The observations are shown in table 10. Look how well we can do when we really try hard! It is just as important to search for an assignable cause for a low point as for a high point. The new center line is at the new average. They are the numbers of defective items in 25 samples of n = 150 items each.0 5 10 15 Sample number 20 25 Figure 10.5.1. The following data set illustrates the difficulty.0 UCL = 29. Unfortunately. There are three points below the LCLnumbers 18.5.55. . 30.05.
THE AKCSJNE 'IKANSFOKMATION 171 10.3281 . The control limits are UCL = 0. where arcsin(u) is the angle whose sine is u.3672 The center line is at 0. l h a t should not be interpreted as meaning that all points below the line are false.6294 and LCL = 0. When the arcsine transformation. ') Nelson (1983a) suggests an alternative transformation: The difference between the two transformations is a matter o f taste and i s not important.3672 0. and the control limits are set at x + 3 sqrt( n . is applied to thc data in table 10. x 2 arcsin(). 0.6.2838 0. This is one of the reasons why the arcsine chart has been proposed as an alternative to the p chart when the sample size is constant. THE ARCSINE TRANSFORMATION In thc last section.4661 0. expressed in radians. Another reason i that the arcsine transformation makes the s plotted variable independent o f p.1395 There arc no p i n t s outside the limits.1635 0.4661 0.1. .3281 0.3672 0.4027 0.3.4027 0.3281 0. The LCL for the np chart is set at the larger of zero and nfi3<$(1 p). One can argue that when the LCI.3844. In applying the transformation. the percent nonconforming. is zero.3282 0.2838 0. the engineer loses power to detect genuine improvements in quality.3281 0. in the sample and replaces it by x =2arcsin(0) .4661 0. one calculates p .4661 0.4027 0. It is also claimed that the transformed variable has a distribution that is closer to normal than does the binomial.5735 0.4949 0. doubts were cast about the importance of observations that fell below thc LCL.4949 0. The variance of x is l l n . the transformed values are 0. It can he shown (and it appears as an exercise) that the LCL will be set at zero unless np >9(1 1)).5223 0.2838 0.3672 0.3281 0.
55) = 34. are called ccharts. . The average number of nonconformities per day during that period was c = 56.3.55.1 is a cchart made from the first 21 points.7. Charts for the total number of nonconformities.95 5 3(7. there are. and the control lines were drawn at 56. The charts shown in figures 10.6.1. 79. The center line of the chart was set at that level. the standard deviation of thc proccss was taken as 7. That is a nonconformity. and the total number of nonconformities in the ten reels was recorded.1 and 10. Thc data arc given in tablc 10.7.7. If the number of nonconformities per reel has a Poisson distribution with p = 5 .7.95. that total will have a Poisson distribution with p = 50. Figure 10. At various places in a spool.1.172 10. The cchart. The letter c is used instead of p to denote the parameter of the distribution. Since the mean and variance of a Poisson variable are the same. CCHARTS CONTROL CHARTS FOR ATIXIBUTFS Suppose that we are looking at the output of a machine that makcs yarn for knitting and weaving. on the average. five nonconformities per reel. based on the Poisson model.2 are taken from a process similar to that described before. ten reels were inspccted in the course of each day.7. Sample number Figure 10.7. If we take a sample of ten reels and count the total number of nonconformities. the yarn will be weak enough to break. The output of 50 days’ work was inspectcd.
2. Investigation over the weekend found an assignable cause in a part of the machine. Her duties were taken over temporarily by another employee who did not know exactly what to look for. The cchart.7. Assignable causeinadequately traincd workcr. It would be nice to conclude that the process had suddenly pulled itself together and that those points might set a standard for future performance. As a result. Unfortunately.7. A temporary repair was made and production resumed on Monday. he missed quite a few weak spots and reported too few nonconformities.CCH ARTS 173 0 10 Sample number 20 30 40 Figure 10. That was not done because a new part Table 10. closer investigation revealed that the regular inspector was sick on those two days.2. Points 15 and 16 are below the lower line. There are three points outside the lines. This occurred on a Friday. and the engineer noticed it immediately without the need for a chart. Number of Nonconformities in Ten R e s of el Yarn 77 59 54 64 54 49 75 41 49 56 38 59 51 41 54 76 60 43 45 24 40 41 61 45 59 51 22 46 43 59 49 60 43 50 49 65 49 50 59 59 45 45 55 56 53 07 43 42 85 53 . Throwing out those three points would have increased the average number of nonconformities to 59.1. Point 10 is too high.
and xIn.. and ten points.+ 5x. Some defects. The total number of demerits is then y = x. + 5x... we have assumed that one nonconformity is no better or no worse than another. This might be a more appropriate procedure in the evaluation of the overall performance of a defense contractor. have a Poisson distribution. 10. DEMERITS In constructing ccharts. some compiinies use a system of demerits. and let them have Poisson distributions with parameters c. We can. During the last 29 days. the average numbcr of nonconformities per day dropped to 49. xs. Figure 10. + lox. Charts for demerits are similar to ccharts. CHARTS FOR ATTRlRUTES arrived on the afternoon of the twentyfirst day and was substituted next morning for the part that had been temporarily repaired. Although Y is a linear combination of Poisson variables.7. ‘The replacement resulted in an immediate improvement. That may not be so. Suppose that a process has defects with three demerit levels..174 CONTROI. no doubt it would have been eventually.2. A more serious defect may carry a penalty o f ten demerits. such as a scratch in the paint.2 shows all SO points plotted with the same control lines that were calculated from the first 21 points. A small defect may carry one demerit. five points. That represented a 10% reduction in the nonconforming ratc. What justifies our concluding from the chart that there has becn an improverncnt is thc preponderance o f points below the old center line. and so it docs not. but il still lcft room for further improvement. it is not the simple sum o f them. respectively..2. are merely cosmetic. . one point. others are functional. however. Let thc numbers of occurrences o f these defects be denoted by x I . + lox. than computing an unweighted pcrcentagc figure. c. To handle this situation. itsclf. I t is interesting to note that rule 1 of the previous chapterseek an assignable cause if there is a point outside the outer control lineswas not triggered.8.. The improvement in quality after the new part was installed is obvious. respectively. A chart of the last 29 points shows a process under control at thc new level o f c. c5. . = y. particularly the run of 13 consecutive points below thc line. to which rcference was made in section 10. readily calculate its expectation and variance as We then set the center line from the pilot data at y= x.
n= X' 1953 70 2448 83 2154 90 1446 47 1753 1441 55 2549 90 2340 82 1954 78 2248 75 1746 49 1450 51 2450 66 1949 65 1554 57 1552 54 2049 84 1852 80 2449 120 2149 74 2038 89 2552 88 2450 85 1850 64 64 1950 80 Batch 16 is clearly a bad batch. The average number of defectives is 73. EXERCISES + 252. t lOOX.5. Thc hatch sizes and the number of defectives are given in the following. but it is next to thc smallest hatch. both with and without the three low points. 10. Make a (Poisson) cchart for the last 29 observations in table 10.3. Make a p chart and a np chart for this data set. Is it out o f control'? 10.1. 4 3 7 2 8 8 4 4 7 8 1 2 1 1 7 3 8 5 8 7 7 5 Is the sample with 13 defectives out of control'! 10.7. which were made after the new part was installed. Do your conclusions agree with those of section 10.6. Is the process under control at the new level of defects? 10.5? . Twenty samples o f 200 items each were taken from a production line. Make a cchart for this set of data assuming that the average number o f defects is 9.2. Use p = 49. Batch 2 has the fewest defectives. Make a p chart of the data. The number of defectives in the samples arc given in the following.EXERCISES and the control lines at y =yo ? 3v2.3.1. Make arcsine charts for the data in table 10.. .2.4. The average hatch s i x is 201.5. Twentyfive batches of devices were subjected to 100% inspection.5 and mark any points that are out of control: 13 9 10 9 8 14 18 17 12 14 12 13 12 19 9 15 15 8 10 11 10 10 11 14 8 10.
when they are bad. Five hundrcd devices are inspected and the number of defects in each device is recorded.176 CONTROL CHARTS FOR AlTRIBUIES 10. they arc dreadful. and so the defects cannot be regarded as occurring at random.. five defects pet item.97. defects. they are indeed good. . on the average. A summary of the data is x= n= 0 274 138 1 2 37 11 3 4 9 5 31 (Le.0. When the devices are good. This corresponds to a process that produces “good” items with an average rate of one defect for every two itcms and is contaminated by about 10% “bad” items that have. Let x . >4) The average number o f defects per device is 0. denote the number of dcvices that have n. Show that the data are most unlikely to have come from a Poisson distribution.6.5 and 50 from a Poisson distribution with p = 5. This data set was constructed with 450 observations from a Poisson distribution with p = 0.
Statistical Methods in Engineering and Quality Assurance
Peter W. M. John Copyright 0 1990 by John Wiley & Sons, Inc
CHAPTERELEVEN
Acceptance Sampling I
11.1. INTRODUCTION
In developing the theory and practice of acceptance sampling, we introduce two people whose attitudes have traditionally been those of adversaries, the vendor and the buyer. The vendor manufactures widgets in a factory. When a batch of them is made (we henceforth use the jargon of acceptance sampling and use the word lot instead of batch), the lot is presented to the buyer, with the claim that it is o f satisfactory quality. Before accepting the lot, the buyer calls for an inspection to determine whether the lot meets requirements. If the lot passes inspection, it is accepted. If the lot does not pass inspection, it is rejected. The reader will already have rcalized that we are going to talk about a special form of hypothesis testing. In practice, the role of buyer can be played by the actual purchaser or by the vendor's own quality control inspectors who stand between the production line and the factory gate. The emphasis in this chapter is on single sample plans. This is the simplest case. The inspector takes a random sample of n items from a lot and rejects the lot if the sample contains more than c defective items. The numbers n and c are predetermined. The discussion is, for the most part, confined to sampling by attributes. Chapter twelve begins with multiple sampling plans and sequential sampling. It ends with a description of the use of one of the major sampling schemes, MILSTD105D. that is commonly used in defense work. The field of acceptance sampling covers a very broad area and we have space for only a fraction of the topics that could be covered. The reader who wishes to pursue the matter further is referred to the books by Schilling (1982) and by Duncan (1986).
11.2. THE ROLE OF ACCEPTANCE SAMPLING
When a buyer is offered a lot, there are three choices: (i) accept the lot as satisfactory without any inspection; 177
1711
ACCEPTANCE SAMPIJNC; 1
(ii) inspect every item (102)0/0 inspection); (iii) use an acceptance sampling scheme in which only a fraction of the items are inspected. ‘I’here arc several arguments against 100% inspection. Among them are the following: (i) if testing is destructive, e.g., if testing consists o f twisting an item until it breaks, the buyer will be left with no items; (ii) testing introduces a chance of damaging good items; (iii) the cost of testing every item may be excessive; (iv) when thc fraction of dcfective items is low, 100% inspection can become a tedious task, with the result that too many defects are missed. A desirable goal would be a manufacturing process in which the first optionzero defectszero samplingis valid. That is the aim of manufacturers who seek firstclass quality and firstclass competitiveness in world markets. This is already happening in situations where a major manufacturer has built up such a \trong, close relationship with a supplier that the expenditurc of time and money on acceptance sampling is no longer warranted. In such a case, nonconforming units are so rare that the occurrence of one will justify a very serious investigation t o identify the error and to eliminate it. There is also a mathematical point. Whcrcas only a few years ago manufacturers talked about defective rates of I % to lo%, w e are now seeing in the microelectronics industry processes that have a defect rate as low as 500 or even 50 parts per million. If a manufacturer is only producing one nonconforming part in lO,o(H), there is only about a 1% chance that a sample of 100 parts will contain a defective part. If the nonconformity rate is onc in lOO,OOU, only about one such sample in I 0 0 0 will contain a dcfcctive. The situation described in the last two paragraphs is a worthy aim. However, many processes are far from that ideal. For that majority, acceptance sampling remains a major feature of the production system. It is there to give protection both to the consumer and to the producer. Schilling (1984) argues that acceptance sampling has other uses than mcrely deciding the fate of individual lots. He argues that “when applied to a steady flow of product l‘rom a supplier it can be made a dynamic element supplementing the process control system.” It can he t h e buyer’s quality control chart for the vendor’s production.
11.3. SAMPLING BY ATTRIBUTES
Each item in the sample is tested by an inspector who rates it as either defective or nondefective. Thc words conforming and nonconforming are
SINGLE SAMPLING PLANS
179
often used instead of the older terms defective and nondefective. We stay with the older terminology, although we sometimes find it convenient to call items “good” or “bad.” The quality of a lot is measured by the percentage o f defective items: the lower thc percentage defective, the higher the quality. Deciding whether a lot should be accepted or rejected is called lot sentencing. When a lot is rejected therc are several choices: (i) The entire lot can be scrapped, which may incur a large financial loss. (ii) If the defects do not render them unusable, the items in the lot can be sold on an “as is” basis at a lower price. This can happen in the garment industry, where batches of shirts can be sold at discount outlets as “seconds.” (iii) 100% sampling can be made of the items in the condemned lot and the defective items replaced by good items; thc lot is then said to have been rectified. These choices are sometimes referred to collectively as lot disposition actions. The standard inspection system for sampling by attributes that is used by the U.S. Department of Defense is MILSTD1OSD. It is discussed in chapter twelve.
11.4. SINGLE SAMPLING PLANS
We mentioned in section 11.1 that, in the simplest case, an inspector draws a random sample of n items from the lot and notes the number of defectives. He works according to a sampling plan that tells him how large the sample should be and gives an acceptancc number, c . If the number of defectives found in the sample is c or fewer, the lot is accepted; if there are more than c defectives, the lot is rejected. Some use Ac (short for acceptance number) instead of c and He (short for rejection number) for c + 1. The main question is how does one choose the values of n and c ? How large a sample size do we need‘? We will be taking samples that are only ii smdl proportion of the lot, so that we can assume that the number of defectives has a binomial distribution. A small handful of items will not give cnough power to discriminate between good quality and bad. Samples of about 100 or 200 will be the norm. It will not matter whether the lot size is 2000 or SOW, what will matter is t h c sample size. A common mistake is always to test a fixed percentage, say, 10%, of a lot. This can lead to ineffective samples from small lots and unnecessarily large samples from big lots. Cynics have suggested that an unscrupulous vendor, who realiled that the fraction dcfectivc in his manufacturing process
180
ACCEPTANCE SAMPLING I
had increased, might make it a point to present the substandard material in smaller lots. Then, since small samples have less power to reject null hypotheses than large samples, he could increase the chance that the poor material would pass inspection! The quality level of a process is the percent dcfcctive when it is under controlthe ccnter line on the control chart. We denote it by p o . The vendor wants a situation in which only a small portion of the product will be rejected when the process is under control. That is the vendor’s risk, alpha. How small alpha should be depends upon the consequences of the rejection of a lot. If having too many lots rejected means loss of business to competitors who have better quality, or if it means large amounts of money spent on scrap or rework, the vendor is under pressure to work at a low value of alpha. The buyer wants good product, with a low percentage of dcfectives. Zero dcfectives is a nice slogan and an important target, but it is not always practical. In some situations, one can obtain zero defectives by very slow and cautious workmanship at greatly increased labor cost. Perhaps it is in the buyer’s interest to acccpt a small fraction of defectives to obtain a reduction in unit cost. This will depend on the consequences to him of rcceiving a defective item. If your office manager buys pens by the gross and the pen that you pick does not work, you can throw it away and take another without too much loss. O n the other hand, if the lot consists of vital parts whose failure is a matter of lifc and death, that is a different situation. The buyer has learned that he can manage comfortably with a certain level of defectives. l h i s is the target level. It is convenicnt if this target level, which we denote as the acceptable quality level (AQL). coincides with the vcndor’s production level. O n the other hand. the buyer also knows that the fraction defective in the lots received from even the most cornpetcnt of vendors will vary; there is some upper bound 011 the fraction defective in a batch that can be accepted without extreme inconvenience or annoyance. This level is called the lot tolerance percent defective (LTPD). The buyer will want a sampling plan that gives 1 low risk, beta, of accepting so bad a lot.
11.5. SOME ELEMENTARY SINGLE SAMPLING PLANS
A single sampling plan is fixed by two numbers, ti and c. Why should one choosc one plan over another? The decision should be made on thc basis o f the relative alpha and beta risks for certain values of p. In this section, we look at four plans with small values of n. These plans are quite unrealistic. ‘They are included only to give the reader some simple illustrative examples. In the next section, we look at some more realistic plans. For purposes of illustration in this section, we make the following basic assumptions to make the arithmetic easy. The vendor has a process that,
SOME ELEMENTARY SINGLE SAMPLING PLANS
181
when it is under control, produces items that have p = 10% defectives. The buyer has learned to live with this rate of defectives and is prepared to accept lots that have a slightly higher percentage, but does not want to accept lots that have as much as 20% defectives. Thus, alpha, the vendor’s risk, is the probability that a lot with p = 0.10 is rejected; beta, the buyer’s risk, is the probability that a lot with p = 0.20 is accepted, an unfortunate situation for the buyer. We use the symbol P,(T) to denote the probability that a lot with percentage defectives equal to 7c will be accepted. We can write
a = 1  Pu(O.l)
and
p
=
fu(0.2).
We assume that the number of defectives, d, has a binomial distribution with the appropriate values of ri and p ; a is evaluated at p = 0.1; p is evaluated at p = 0.2.
Plan A. n = 3 , c = O
Draw a sample of three items and accept the lot only if all three are good. If = 0.10, q = 0.90, the probability that a sample will pass inspection is (0.9) = 0.729.
p
(Y
= 1.0  (0.9).’ = 1  0.729 = 0.27
p = (o.$o)3 0.51 . =
Approximately a quarter of the vendor‘s “good” lots will be rejected. On the other hand, if the vendor presents lots that are 20% defective, about half of them will pass inspection and the buyer will have to accept them.
Plan B. n = 5 , c = O
a = 1 . 0  (0.9)’ = 0.41, p = (0.80)s = 0.33. This plan is tougher on the vendor. but it gives the buyer better protection against bad lots.
Plan C. n = 5, c = 1
(Y
= 1.0  P(d = 0) 
f(d = 1)
= 1.0  (0.9)s  5(0.9)4(0.1) = 0.082
.
p = (0.8)’ + 5(0.8)‘(0.2) = 0.738.
This plan gives much more protection to the vendor than the previous plans, but it is completely unacceptablc t o the buyer.
182
ACCEPTANCE SAMPLING I
E
Plan D. n = 10,
=1
a = 0.26, p = 0.38. The two risks are more nearly equal, but they tire clearly too big for practical purposes. We need to take larger samples.
1 I .6. SOME PRACTICAL PLANS
In this section, we consider three plans that are o f morc practical use. Two of them have IZ = 100; the other calls for rz = 200. In comparing the plans, we assume that the vendor and the buyer have agreed on a target rate o f p = 2 % defectives.
Plan 1. n = 100, c = 4
Draw a sample of 100 items and accept the lot if there are four, or fewer, defectives. ' o cvaluatc alpha, we need to compute, for p = 0.02 and q = 0.98, l
This quantity is not easy to compute because of the number of decimal places required. During World War 11, it was very difficult hecdUSe electronic computers were not yet available. Statisticians turned to the Poisson approximation to the binomial, which is close enough for practical purposes ttnd is, nowadays, easy to compute o n a hand calculator. Thew ;ire now good tables of cumulative probabilities available, from which we can read a = 5.1%. To obtain the Poisson approximation, we equate the parameter, p , of the distribution to np. In this example, p =2.0, and
P(d 5 4) = e'
Ly
= 7 e  = 0.947, = 5.3qJ
i+
1
p
P2 P.' + + + 6 24 2
"4)
.
Table I 1.6.1 gives both the binomial probabilities of accepting a batch for several values of p and the Poisson approximations.
Table 11.6. I . Values of P, for Different Percentages Defective (n = 1 0 , c = 4)
P
Binomial Poisson
0.02
0.040 0.047
om
0.818
0.815
o.(w

I
0.629 0.629
0.0s 0.436 0.440


_
_
_

_
~
n.06 0.277
0.285
0.07 0.163 0.173
0.08 0.090 0.100
CtIOOSING THE SAMPLE SIZE
183
Table 11.6.2. n = 200, c = 7
P P',
0.02
0.951
0.03 0.7&
0.04 0.450
0.05 0.213
0.06
0.083
0.07 0.027
0.08
O.oc)8
Plan 2. n = 200, c = 7 This plan also has a at approximately 5%. but it is more favorable to the buyer. The values of beta are less than those for plan 1. The sample sizc has been doubled. The values o f P,, are given in table 11.6.2. Plan 3. n = 100, c = 7 This plan is the most favorable of the three to the vendor. It gives him enormous protection with an alpha risk of only 0.1%; the buyer will accept 87% of those lots that arc 5% dcfectivc. The sample size is the same as plan 1, but the acceptance number is the same as in plan 2. The prohabilitics are given in table 11.6.3. 11.7. WHAT QUALITY IS ACCEPTED? AQI, AND 1,TPD The buyer detines two particular levels of quality. The acceptable quality level (AQL) is the worst average level of quality that he is willing to accept. Often this is a levcl of percent defective that the buyer has learned to live with over the years, and both the buyer and the vendor may agree that such a percentage is reasonable for both sides. There is also a lower level o f quality (higher value of p) that the buyer finds particularly offensive (unacceptable), even in single lots. This is the lot tolerance percent defective (LTPD). The buyer wants a plan with a low probability, beta, of accepting a lot at the LTPD. The vendor wants a plan that has a low probability, alpha, of rejecting a lot that is at the AUL. Example 11.8.1 shows how one might design a sampling plan with the AQL at 2% and the LTPD at 7%. 11.8. CHOOSING THE SAMPLE SIZE In chapter seven, we showed how to calculate the sample size and the cutoff point for a onesided test of the mean of a normal population. The same principle is used to construct a single sampling plan. Suppose that we wish to
Table 11.6.3. n
Y
= 100. c = 7
0.03
P"
0.02
0.9W
0.989
0.04
0.953
0.OS (J.872
0.06 0.748
0.07 0.599
0.08
0.447
using the normal approximations to the binomial distributions involved. Two measures are used: the average outgoing quality (AOQ).07: fi= 1. 0 11. and /3 = 10% when p = 0.: p = po against the alternative H. Then. THE AVERAGE OUTGOING QUALITY.. = 11.02.P o Example 11. The average outgoing quality is the quality that one gets by applying a sampling plan to a process in which the average percent defective is p and rectifying all the lots that fail inspection. whence fi= z a m iz p m PI . When a lot is accepted. if p = p .280.n items in the lot may include some undetected defectives. If p = p . we make the following assumptions about the inspection procedure.n ) p defective items. each lot accepted by the buyer contains ( N . We can calculate n and c for the test (sampling plan) in the following way.645mt1. Thus.q. AOQ AND A o Q L Both the vendor and the buyer may also be interested in the average quality of the lots that pass inspection with a given plan. and u 2= np. Suppose that the engineer wishes to design a sampling plan with risks a = 5% when p = 0. In calculating the AOQ.: p = p 1 at given levels of a and p. This means that it is subjected to 100% inspection in which every defective is (we hope) detected and replaced by a good item. Thus.05 n=124 and c=5. When a lot is rejected. and the average outgoing quality limit (AOQL). if p = P o . and.138. it is rectified. the defective items in the sample are replaced by good items and the rectified sample is returned to the lot.184 ACCEPTANCE SAMPLING 1 test the null hypothesis H. on the average. the number of defectives is normally distributed with p = np. .1. . Given the problems that exist with 100% inspection. .9. the rejected lots are replaced by lots that have zero defectives.8. The remaining N . this assumption may be too optimistic.
will be rejected and rectified. if the quality drops to 8% defectives.07 0.083 0.1. it is 2.24 0.06 0.6 arc shown in table 11. The average outgoing quality limit (AOUL) is the maximum of the AOQ for a given plan.” We are saying that for plan 1.1. When these are combined. bad lots come out better than good lots. is thus only fractionally lower than p.03 0. A O Q for Plans 1 and 2 Plun I P pa AOQ (%) 0.52 0. few lots are rejected.24% at p = 0.n)P N If nlN is small. the buyer will receive.008 .06 0.04 0. will occur in the case when the vendor produces 4% defectives.450 1. Then.52% at p = 0.50 0. On the other hand. with p = 2%. and so few are rectified. Note that we spcak of “longterm average” and “expect fo endure.45 0. In plan 1. 1.02 0. AOQ A N D AOQL 185 A fraction.746 0.08 0.19 0. The AOQL represents the worst longrun average percent defective that the buyer should expect to endure when that plan is used.9.72 Plun 2 P pa AOQ (%) 0.818 2.90 2.THE AVERAGE OUTGOING QUALITY.03 0.090 0.05 2.04 0. 1 P . Note well “on the average”Table 11.24% defectives.04.06 0.949 1. (1 1. the average number of defectives per lot is seen to be and so the AOQ is P A N . with this sampling plan and rectification. this quantity is wcll approximated by the formula AOQ = P u p .02 0.951 1. the AOUL is 2.90 0. and for the second.277 1.2) Average outgoing quality values for plans 1 and 2 of section 11.07 0. of the lots will pass inspection.163 1. . on the average. on the average. P a . Note that under this scheme.1.08 0.98 0. The AOQL for plans 1 and 2 are underlined in table 11.9.629 0.027 0. A fraction.14 0. 91% of the lots will be rectified.213 1. virtually all the lots will be subjected to 100% inspection and the AOQ falls to less than 1%. For the first plan.07 0.66 0.9.18 0. the AOQ.9. 2.9096.03.05 0. the worst fate that can bcfall the buyer.436 2.
In 1977 the Journul of Qualify Technology devoted an entire issue to his memory.11. The average total inspection (ATI) is then given by Pun + ( I .n) ..l) In section 1 I . this procedure gives values o f the AOQ and AOQL that arc very clost: to the values for the scenario that we considered.1 I . and the overall quality of the product that the buyer actually receives will improve.INC. who pioncered the field of acceptance sampling during his career at Bell Laboratories. 1077). SCHEMES We have confined our discussions to single sampling plans for sampling by attributes.10. F. N items will be inspected. In the remaining sections o f this chapter. All four of thcsc schemesrectitication sampling plans.)(iV . we chose a sampling plan based on desired values of A Q L and LTPD: the lot size N did not appear in the calculations. we consider multiple sample plans and sequential sampling. We have only discussed one way of defining AOQ. Dodge. Our scenario iniplies that there is a supply o f good items that can be substituted for the defectives. we make a few remarks about four other sampling schemes. the rcadcr is referred to the books by Duncan ( 1986) and Schilling (1982) that were cited earlier. and continuous samplingare associated with H. Each accepted lot will have n items inspected.106 ACCEPTANCE SAMPI. RECTIFYING INSPECTION PLANS We assumed in the previous section that every rejected lot would be subject to 100% inspection. (11. which in turn almost implies that the inspection is taking place at the vendor’s plant before shiprncnt. 11. we will detect (and rectify) more of the bad lots.. We can estimate the cost of inspection using this procedure. reprinting in that issuc nine of his most important papers (Dodge. chain sampling.8.P. There are other possible scenarios. A differcnt procedure for testing at the buyer’s site would be for the buyer to carry out 100% inspection o f the rejected batches and cull the bad items (which would not be paid for). OTHER SAMYI. in each rejected lot. For further details.lNG J some days the buyer will fare better and some diiys the buyer will fare worse! The interesting fact to note is that if the production line deteriorates and the vendor produces more than 4% defectives. Dodge and Romig (1959) published tables for choosing values of n and c based on . 11.)N = n + (1  P. skiplot plans. When the percent defectives is small. In the next chapter.
perhaps. A pressure that is. That is appropriatc if the vendor is erratic. As long as the vendor’s past record is free of blemishes. and set c = 0.14. the lot is rejected. The constants that define the plan are not n and c.12. The idea here is that after the vendor has established a past record for good quality. we have envisioned the vendor presenting the itcms to the buyer in (large) lots.CON’rlNIJOIIS SAMPLING 187 other considerations. 11. The vendor must start again from scratch. In chain siimpling. should. test only the oddnumbered lots that arc presented. CHAIN SAMPLING PLANS We have built up a theory for handling individual lots. Which lots are chosen to be tested is decided by a process of randoniization. which could be applied to a situation in which the inspector saw a continuous flow of items passing by on the production line. which appeared in 1981. F. CONTINUOUS SAMPLING So far. . A listing of continuous sampling plans is given in MILSTD12353. we will still accept the lot. H. SKIP LOT SAMPLING There is another way of rewarding past good quality. They assumed that the objective would be to minimize the AT1 for given values of the AOQL. we take that into consideration. provided there have been no defective items in the previous i samples. Dodge (1943) proposed a scheme for sampling by attributes. ‘That is very different from telling the vcndor that thc buyer will. in future. ‘These plan4 were subsequently refined by Dodge and Torrey (1951). we should not insist on inspecting every lot. that claim o f previous virtue is lost. and N. wc should skip some lots and test only a f fraction o the lots that arc presented. If thcrc is cxactly one defective item in the sample. But if thcrc is a defect. If the number of defectives exceeds one. which was introduced by Dodge (195Sa). 11. But if he is fairly steady. p. It is skip lot sampling. 11.13. but n and i. not so subtle is put upon the vcndor. or can. a singlc defect will be excused. building up a record of pcrfcct samples until there is again a backlog of at least i without defect. which also was introduced by Dodgc (1955b). The buyer might agree to test roughly half the lots i d choose thcrn by tossing a coin each time a lot arrives. Rather. we make substantial reductions in ri. and by doing so save money on inspection? This is the rationale for chain sampling.
compared to a critical value. SAMPLING ACCEPTANCE SAMPLING I BY VARIABLES The proccdurc for sampling by variables is similar to the procedure for sampling by attributes. in turn. When it is appropriate. Design a single sampling plan that has a = 5 % when p = 0.1. which implies a multivariate testing situation.4. The set of n resistances is reduced to a single statistic.06.6. The engineer accepts the lot if there are four or fewer defectives. sampling by variables has thc advantage of requiring smaller samples.03? What is the probability of failing to reject a lot with p = 0.188 1I . 11. . say.03. and /3 = 15% when p = 0. 11.2. the resistance. Onc draws the sample and measures a characteristic. Find the AOQL for plan 3 of Section 11.08? What is the AOQ for p = 0.06? 11. EXERCISES 11. such as the mean or standard deviation.03? What protection does the buyer have against accepting lots with p = 0. for each item. A single sampling plan has n = 150. What is the probability of accepting a lot with p = 0. There can be difficulties when thc conformity of the item calls for its being satisfactory in more than one characteristic.3. A single sampling plan has n = 200 and c = 5. which is. 15.03? What is the AOQL for this plan? You may use the Poisson approximation. What is the probability of accepting a lot with p = 0. for example.
we take another sample to settle the issue. THE COST OF INSPECTION Except for defining AT1 and mentioning the DodgeRomig rectifying inspection plans. If it is a very bad sample. It used to be customary in military procurement. we rcjcct thc lot. We take an initial sample from a lot. This is more expensive. In our examples. In theory. the buyer gets 100 fewer items per batch than otherwise. 12. items and the second n2. If it is a very good sample.Statistical Methods in Engineering and Quality Assurance Peter W. One way of getting around this problem is to adopt a double sampling procedure. DOUBLE SAMPLING We will investigate sampling plans with two samples. M. and c 3 . John Copyright 0 1990 by John Wiley & Sons. the buyer could argue in favor of plan 2 because that was the plan that gave him the best protection.1. but is not a condition o f the modern MILSTD105tables. It seems a foolish waste to keep on inspecting the whole of a large sample if it is obvious that the vendor is producing at excellent quality with a percentage defective well below the target value. I n many double sampling plans. O n the other hand. it can be argued that it also called for inspecting n = 2 0 items rather than n = 100. The first has n. n z = 2 n .. we accept the lot. either as a direct charge for the inspection at the buyer's site or as an overhead item figured into the vendor's bid for inspection at thc vendor's factory. This is not a necessary condition. The ultimate refinement is the method of sequential analysis developed by Abraham Wald (1947). Inc CHAPTERTWELVE Acceptance Sampling I1 12. c 2 . The extra inspection cost will be borne by the buyer. If the testing is destructive. thus far we have ignored the cost of inspection in the f discussion o sampling plans. 189 . If the number of defectives is neithct so largc nor so small as to make a decision obvious. one could continue this procedure to a third sample or even a fourth. There are three acceptance numbers: c.2. .
0710 0. 5 5.0222 0.4315 0. .0902 0. requires a samplc size about 190. inasmuch as only twothirds as many items have Table 12. Wc draw another random \ample of riz items. = 9. = 6 d. and so on.ooo1) d.859)(100) t (0. = 4 d.’ 0. (Y = 4.2. If rf i c 3 .028)(100) + (0. c.1. If c . thc lot is rejected. 5 1 d2 2 0 0. we reject the batch outright.02.6288 0. 0 Exumple 12..028.275. and cb2= c. wc acccpt the batch with no further >ampling.02. we use the binomial tables.0176 . = 3 and d.2%. When p = 0. The vendor who is producing good quality items is rewarded by a marked decrease in inspection costs. = 5.02. If d .0114 0 .0049 Ow 1 S H7 O. The possibilities arc listed in table 12. We then sum the products of the probabilities of d . If d . so that a = 4. = 0.1 %. 5 c.2. < d .05.714)(300) = 243. 0 A single sample plan with p. = o d. with n = 100 and p =0. For thc first sample. =7 d.1 First Saniple d. :>9) = 0.0353 0. d .(MH) 1 o. 5 c ? .IYO ACCEPTANCE SAMPLING I1 We draw the first random sample iind note the number. Suppose that 1.0989 and the probability of acceptance overall is 0. and let rf = d . and /3 = 27. I 3 d. we give the lot a second chance. Wc accept the lot if d = c 3 .n 2 = 200. The double sampling plan may require either 100 or 300 items to bc inspected. we note that P(d.7867 0. Summing. 5 s Probability Product 0. The expected number of items to be tested is E ( n ) = (0.54 d.958.258 + 0. We illustrate this method of double sampling by an example.05. =5 d. the probability of acceptance on the second round is 0.141. and so E ( n ) = (0. and d 2 to obtain the desired probability.23s 1 (1.1. Such a plan requires us to take n = I00 every time.05. If p = 0. + d . of defective\.2579 + 0. = 0.(HH)2 Second Sample d. = 3. Let n . = 8 Probability 0. there is probability 0. I 2 n.cK)31 0 .02. Acceptance after the second sample can take place in scveral ways. and find that the probability of acceptance is 0. or d .5%. = 1 .ow7 O.2. Let d L denote the total number of defectives 111 the second sample.0175 = 0. > c 2 .859 and the probability o f nonacccptance is 0.141)(300) = 128 When p = 0. . (1894 0.859 that we will only need the first samplc. p . We can have d . d 2 5 4 . we can show in the same way that the probability of acceptance is 0.
After each observation is made.x.. The line starts in the nodecision rcgion. accept H. k..: simple alternative. . THE SEQUENTIAL PROBABILITY RATIO TEST Wald's sequential probability ratio test was developed during World War I1 and published in 1947. We keep on sampling until the line crosses into either the acceptance region or the rejection region. ?'he bitsic ideiis have been given in the previous section. The total number of defectives at each stage is plottctl against the sample number on ii chart thiit has two lines drawn on it. rejection. if L < k. = 4. . if L > k. identically distributed random variiiblcs: x ..THE SEQUENTIAL PKOBAUILI'rY RATIO TEST 191 to be sitmpled. . The plotted points are joined by a line as in a control chart.3. . . to correspond to the preassigned values of alpha and beta.< L . . we calculate the likelihood ratio: We calculate two numbers. I ) we take another observation. . and no decision. 12. H .4..: 6. ) . reject it. If k.=c k . the United States during World Witr 11. we accept we . It would seem that with bad luck. we count the total number of defectivcs so far and make onc of three decisions: acccpt the lot. and k... Suppose that the density function for X is &(x) under the null hypothesis and f l ( x ) under the alternative. . Elo: 4 = 6... Suppose that we wish to test a simple null hypothesis. just as it is theoretically possible that we can toss a coin for months before getting the first hcad.4. The vendor with poor quality has to pay for inspecting more items. . . but Wald showed that in practice.against a about a parameter 6. SEQUENTIAL SAMPLING The procedure of scqucntial sampling was developed by Abraham Wald for sampled one by onc. or keep on sampling. 9 (12. and H. We observe it scquencc of independent. H .. we could kecp sampling forever: this is thcoretically possible. 12. . These lines divide the paper into thrcc zones: acceptance. The items After each item. his proccdurc called for testing many fewer items than the traditional fxcdsamplesize procedures described earlier. x 2 .
= 1a and 1P k.: p = p ... i 5 n.n6 ' 6 2 cr21n(k.5. are very difficult. in the second. After some algebra. we assume that 6 = p. =  a ' which are very good for small values of alpha and beta.5.4.2) We now consider two examples. the bounds on Y given in equation (12.) < ln(L) < In(k.5. We have lf. .p"). . X will have a Bernoulli (binomial) distribution.1) become a pair of parallcl lincs: (12.. the inequality of equation (12.5. (12.4.4.) < Y n .) 6 . THE NORMAL CASE Supposc that X is normally distributed with known variance and unknown mean. X will be normally distributed. 12.: p=p. = C (xi.) . we continue to take more observations. It is also useful in practice to replace the expressions in equation (12.192 ACCEPTANCE SAMPLING I1 H I .2) and (12. for the sum of the first n deviations from k.: Y. and H. In the first. corrcsponding to sampling for variablcs. 1 C.5.po> 0. The calculations of the correct values for k. We write Y. The testing stops when the path crosses one of the two boundary lines.1) If we plot the data with n along the horizontal axis and Y vertically.2) is reduced to < u ln(k. (12.3) As long as the path of Y remains between the two lines.1) by their natural logarithms and to write In(k. For convenience. so that in practice we use the approximations B k. and k .
0.H.) 6 ’ (12. . Again.5. AN EXAMPLE Supposc that V ( x )= 125. and we.95 0. This is particularly true if (Y = /3.ind lower limits are Z = 3 6 and Z . we write f z . Consider the following sequence of observations: I1 X Y 1 47 7 23 41 50 2 17 3 53 30 13 51 4 46 36 8 56 5 SO 46 3 61 6 47 53 2 66 67 54 7 8 48 75 12 76 9 49 84 5n28 5n + 36 18 46 7 71 17 81 The path of Y goes across thc upper boundary at the ninth observation. 12. = x. = 50.w 0..25.. therefore. = 50.18.) = 2. = z.0. = . Then 0. the two boundaries become the two horizontal lines Z= a 2ln(k. k .= PI + Pu 2 6 2 and Z.6. = . accept the alternate hypothesis that p = p . . The boundary lines of the critical region are Y=Sn28 and Y=Sn+36.28. .105. and /3 = 10%. Thcn the boundary limits are at the same distance above and below the center line Z = 0.10 k. p. . a = 5%. po = 40. we can write X z z 47 2 2 50 5 7 53 8 15 46 1 16 50 5 21 47 2 23 54 9 32 48 3 35 49 4 39 The upper .AN EXAMPLE 193 If instead o considering deviations from po. In(k.4) and the plot looks like a Shewhart control chart.89. 6 = 10. thc graph enters the critical region at the ninth observation. Alternatively. .) 6 and Z= v2In(k.05 In(k.x.0.) = 2.
3) Again. If we write z. In(k. = 0 if the item is nondefectivc. In(k. has a binomial distribution.194 ACCEPTANCE SAMPLING I1 12.: p = p.) = Jr2.06 = p = 0. 12. The two hypotheses are €lo: p = p. + g .2) where we have written and (12.96 = ) K= In( 94/98) h = 0.7.) = 2. the boundaries are two parallel lines. : y =0. . Y. its Wc continue to take more obscrvations as long (12. and f?. = C x.7.197 ' 0 06' 0 94 h In( 0...197. 0.10.8.14. 1 5 i 5 n.7.02 with (Y vs.. AN EXAMPLE OF SEQUENTIAL SAMPLING BY ATTRIBUTES Supposc thal we wish to tcst H o : p =0.. the boundaries for 2 are horizontal lines. H .7. .0365. = 1 if the ith item is defective. = x.1) which reduces to (12.02 In( I)1. and x. SEQUENTIAL SAMPLING BY ATTRIBUTES We now turn to thc case of sampling by attributes and let x .
the limit is 2. and we reject H.: p z &. The lower limit is negative if ri < 52. Y = 0.7.a l 2 ) and (1 .19. and so on. 0.I . We again plot Y = C ( x . The bounds k. are set at p l ( 1 .79.9. we accept ii. and the path of Y will cross the upper line if there are thrcc defectives before n = 29. . are broken into two singlesided tests. ..).p(.036Sn = 3 when n 29.927 + 0.39.0365n The path of Y i5 a horizontal line when there are no defectives.3 items. respcctivcly.. If the first two items are defective. have to be considered in terms of integers. if = I (i) (ii) (iii) (iv) the first two items are defective.0365n 1. we continue sampling.93 + 0.927. each defective causes a junip of one unit. we will have Y = 2 when n = 2. Corresponding lines in the pairs intersect on the center line.036Sn 5 Y Sr 1. Similarly. 12. becomes with the risk of type I error set at a / 2 o n each side. We accept H . TWOSIDED TESTS Tests of a simple null hypothesis against a twosided alternative H*: p = p* vs. N.927 + 0. The lines mark off a Vshaped region on the graph paper. The inequalities for the likclihood ratio define two pairs of parallel lines. there arc five defectives in the first 84 items.and k . The limits.53. the path will not cut the lowerlimit line until n = 5. there arc three defectives in the first 29 items.. It is unity for n = 80. (ii) if thcre arc n o defcctives in thc first 53 items tested. ‘Thus. For the upper limit.000. (iii) if there is exactly one defective in the first 5.TWOSIDED TESTS I95 We continue to take more observations as long as . 1.39. H . We therefore accept Ii. there are four defectives in the first 56 items. if n c. if there tire no more defectives until n = 81. wc note that for n = 2. and so o n .. * (i) we cannot decide to accept H ./3)/(cu/2). This means that if we have an initial string of nondefectives corresponding to the line Y = 0.
We should have V ( x )= 12S. 6 = 10. corresponding to 5% in each tail. and /3 = 10%. = 40. ‘The pair of parallel lines above the center line are Sample number Figure 12.25 and In(k.) = 2. .I% ACCEYrANCE SAMPLING 11 ‘The lines above the center are The lines on the bottom half of the paper are their reflections in the center line: 12. a = 10%.89 as before. fi. AN EXAMPLE OF A TWOSIDED TEST Suppose that we had elected to carry out a twosided test in the example of section 12.) = 2. but with cy = 1096. Twosided sequential test for section 12.10.10.6.10. l h e n In(k. I .
= 5n + 36. the American National Standards Institute and the American Society for Quality Control produced a civilian standard. The 1981 revision of this standard contained the change in terminology for dcfcctivc to nonconforming and similar rewordings.MILSTD. In 1971. The plot of this twosided test is shown in figure 12. Department of Defense issued a system for sampling by variables. MILITARY AND CIVILIAN SAMPLING SYSTEMS The standard for attribute sampling used in defense contracts in the United States is the official system called MILSTDl0SD. A complete account of these systems can be found in the latest edition of the book by Duncan (1986).6. crosses the line betwecn n = 8 and n = 9. B and C. the graph would be the upper half of figure 12. double sampling plans. MILSTD414. in 1957. This system.1 and y = 5n + 28. MILSTD1235B is a collection of continuous sampling plans for attributes.10. called ANSI/ASQC Z1. MIIdSTD1OSD The MILSTD1OSD system contains numerous tablcs o f plans for sampling by attributes.1. A modern civilian version of MILSTD414 was brought out in 1980. = Sn . It includes the CSP1 plans of Dodge (1943) and the CSP2 and CSP3 plans of Dodge and ‘I’orrey (19Sl). The two parallel lines below the center line are Y.12.28. The tables consist of single sampling plans. which appeared in 1963. This sampling system contains a collection of sampling plans. ANSI/ASQC 21. In it. .S. It closely paralleled MILSTD10SA. indicating that the null hypothesis p = 40 should be rejected in favor of p > 40.1 .. It has not been changed to kccp pace with the changes in MILSTD105 that came with the current version.4. Y. and some multiple sampling plans involving more than two stagcs. The 105 system first appeared in version A in 1950. no tablcs are included. D.9.1. Although sequential sampling is mentioned in the accompanying handbook.1 .10. wc see that the cumulative sum o f the deviations. before major changes were made in the current standard 105D. 12.1 5 0D 197 Y. parallels MILSTDlO5D. 12. For the onesidcd test o f section 12.. The U.36.1 1.1 and y = : Sn . It went through two revisions.
This can produce hard feelings. reduced.” The handbook itdds that “AQL plans are dcsigned to protect the supplier from having good lots rejected. or values. The producer then finds the sarnplc sizc code from a table and looks in another table to find the critical value. If single sampling inspection is to be done. but they are not in common use. The decision whether to use single or double sampling may be left to the producer. reasonable costs for inspection are includcd in the price agreed upon beforehand. can be considered satisfactory as a process average. for the b’ e n w AQL. The buyer contracts for a quantity of items that meets a sct of standards. . The producer of poor quality has higher costs. day after day or week after week. (the probability o f acceptance) is 50%. The buyer’s inspectors must iilso bc allowed access for occasional supervision of the actual inspection carried out by the producer. THE STANDARD PROCEDURE Once the buyer and t h e producer have agreed on the AQL to be used. The buyer itnd seller agree upon the AUL and the general inspection level (which is a function of the importance of the item being inspectcd). the producer begins by looking in the master table for normal inspection. Consider. is introduced. indiffcrcncc quality.198 ACCEPTANCE SAMPLING 11 The tables are based upon the AQL. A new term. The procedure calls for three levels of inspection: normal. then the buyer profits too by receiving the bcttcr quality product. But if the producer is thus nudged to improve quality. the Ml LSTDIOSD procedure follows several steps. and tightened. We discuss the situation in which the producer is making batches of the same s i x . Kcduccd inspection means increased profits. The implication is that the producer is paying for the inspection. for example. This is the ‘ quality for which I. 12. a producer who is producing lots of 1500 items under general inspection level 11. for the purpose of sampling inspection. Some people design pliins based on the indifference quality. If the inspection is to be carried out by the producer in his plant. Code letter K calls for a sample of 125 items. which is defined therein as “the maximum percent defective that. good records must be kept to which the buyer must have access. The consumers’ risks of accepting product of inferior quiility are only indirectly considered by studying the OC curve for the AQL sampling plans. That lot size corresponds to sample code K. The producer of good quality product is rewardcd by reduced inspection costs.” The term lot tolerance percent defective (LTPD) is replaced by limiting quality (LO).13.
5%. tightened inspection is required. TIGHTENED VERSUS NORMAL INSPECTION Tightened inspection usually calls for the same sample sizes. the lot will be acceptcd if there are three or fewer defectives and rejected if there are seven or more.5%) Sample Number 2 3 4 1 Sample Size 32 32 32 32 32 32 32 Saniple Size 32 64 Cumulative Acceptance No. Continuing with the example o f an AQL o f 2. Suppose that thc AQL is 2.13. Multiple Sampling for Sample Size Code K (AQL = 2.TIGHTENED VEKSUS NORMAL INSPECTION 199 Table 12. which is used when there is no reason to believe that the quality lcvel is either better or wor5c than has been specified by the AQL. (Ac) Rejection No. If there are three or fewer defectives. reduced. the criteria for tightened inspection .5 and sample size K .1. but the requirenients for acceptance are more rigorous. the lot is accepted.14. and tightened.1. = 7. If the producer chooses double sampling. IS. is three. Criteria are given for switching from normal to tightened inspection and back. = 80. and from normal to reduced inspection and back. 12. 12. c (or Ac). After the second sample. t i I = n. (Re) 4 (1 1 S h 7 90 128 160 192 224 3 5 I 10 13 10 11 12 14 6 8 If the AVL is I % . the lot is rejected. SEVERITY OF INSPECTION There are three degrees o f severity o f inspectionnormal..13. This is as close a s MILSTD1OSD comes to sequential sampling. c. each of 32 items. if therc are four or niore defectives. the supplier may be allowed to go to reduced inspection. Thc multiplesampling choice allows up to seven samples. If the AQI. If it is clear thiit the quiility level is very good. the critical number. If it becomes apparent that the quality level has deteriorated. the lot will be accepted if there are eight or tewer defectives in the 160 items and reject if there are nine or more. Inspection usually begins at the normal level. This is a change from the older procedure in which nz = 2r1.5%). is 2. The acceptance and rejection numbcrs after each sample are given in table 12. After the first sample is drawn.
3 1 Rejection No.1. This denotes that one cannot accept a lot on the basis of the first sample.5% Type Single Double Cumulative Sample Size 50 Acceptance No. This also occurs in table 12.5% TYpc Cumulative Sample Size 125 160 Acceptance No. the sample sizc is reduced to approximately 60% of what is rcquircd for normal inspection. AQI.16. In tablc 12.1. one can only reject it or take a second sample. with a corresponding change in the cost of inspection.15. The use of this table and the apparent Table 12. 6 5 7 4 32 64 13 26 39 52 65 78 91 4 Multiple 5 6 7 8 9 10 . We continue with our example and give the criteria for rcduced sampling in table 12.1. 6 Single Double Multiple 80 32 2 6 5 7 4 5 8 64 96 128 160 * 2 5 7 9 1 3 6 7 192 224 9 1 0 are given in table 12.200 ACCEPTANCE SAMPLING 11 Table 12. Reduced Sampling for Sample Size Code K. AQL = 2. Tightened Inspectlon. Sample Code K.1.1. 12.16.16. REDUCED INSPECTION In reduced inspection.1 the acceptance number for the first sample under multiple sampling is replaced by an asterisk.15.16.15. = 2. Rejection No.
the number of defectives exceeds the acceptance number. and a search is made for an assignable cause. Part B has plans based on the use of the range to estimate the (unknown) standard deviation. If two lots in any sequence of five are rejected.17. in any lot. If the number of defectives falls between the acceptance and rejection numbers. ACCEPTANCE SAMPLING BY VARIABLES The MILSTD414 document has four parts. We are speaking here of rejection on the original inspection. The producer may continue on reduced sampling as long as the number of defectives in each lot is less than or equal to the acceptance number in the table.ACCEPTANCE SAMPLING BY VARIABLES 201 discrepancies between the acceptance and rejection numbers are explained in the next section. the producer may return to normal inspection. therefore. the producer must return to normal sampling. Part D handles the situation when (r is known. SWITCHING RULES We begin with normal inspection.9. The morc modern civilian system. 12. If after ten lots have been inspected. the inspection is stopped. Tightened inspection then continties until either of the two following events occurs. but inspection goes back to normal anyway. and D.18. the process has not won its way back to normal inspection. 12. A lot that is rejected on the original inspection and then accepted after rectification is regarded as a rejected lot. If five consecutive lots are accepted under tightened inspection. If. Inspection and production will resume when it is thought that proper quality has been restored. Notice that in reduced sampling. There itre. three serics of plans covering the situations in B. and (ii) if the total number of defectives in those ten lots is less than or equal to the acceptance number for the gross sample that is given in a second table. Part C has plans for use when cr is estimated from the sample standard deviations. the producer must switch to tightened inspection. Part A is the general dcscription. the lot is accepted. C. the acceptance and rejection numbers are not consecutive integers. has switching . the lot is rejected. If the number of defectives is greater than or equal to the rejection number. ANSIIASQC Z1. Plans are given for onesided and twosided specification limits. The producer is allowed to go from normal to reduced inspection if (i) the preceding ten lots have bccn accepted.
6 103.9 101. 17.5 108.0 95. The standard deviation of the characteristic is u = 6.2 102.0.9 109.4 99.3 103.3.2 113. and 67 are defective. 20.06. 12. For further details.9 98.0 107. When would you make a decision and what would it be? (i) (ii) 98.7 100.8 108.3 99. 12. When would you make a decision and what would it be? (i) The first threc items are defective. and p = 0.6 104.0 109.1 95.6 105.05 when p = 0.4. EXERCISES 12. = LOO.8 104.8 106.03? .8 12. and rejects it if d>7.3 98.0 1w.4 85. (iii) Items 75.4 107.3 97.8 109. The target average for the characteristic being measured is 100.3 87. 174. 63. The lot is accepted after the second sample if the total number of defects is eight or less. 135.05 when p = 100.8 92.7 87. 37.3 98.2 93.2 101. (ii) Itenis 1.8 96.2.5.3 9. and 202 are defective.8 101.6 100.8 102.o 11 1.9 105.9 106.0 108.3 99.15 when p = 103. 42.3 to the following three data sets.202 ACCEPTANCE SAMPLING If rules between normal inspection and reduced or tightened inspection that are similar t o those of MILSTDIOSD.4 10'2.2 103. Apply your sequential sampling plan of Exercise 12.7 107.3 108.1 to the following data sets.5 108. Apply ~ O U Jplan from exercise 12.2 109.0 98.1 103.8 02.4 102. 175.9 09.8 104. The engineer accepts the lot o n the first sample if d s 3 . We wish to have a =0. Design a sequential sampling plan for sampling by variables with the following specifications. 12.5.9 07.2 96. and p = 0. 12.15 when p = 0.4 103.3 105.5 102. 101.7 101.6 93. 91.7 98.3 99.5 101. 109. A double sampling plan has 11. the reader is referred to Duncan (1986).1.6 102.1.4 103.2 105. = n. I 111.4 96.4 96.1 103.03. 190. What is the probability of accepting a lot with p = 0.9 92. Design a sequential sampling plan for attributes that has u =0.7 103.0.7 102.
a shift in the mean o f the process occurred after the thirtieth run. cl Rule 2. but at sequences of several points. How can we speed up the detection of changes in mean? It is donc by looking not just at the iiidividual points on the chart. on combining rules 2 and 3 with rule I . since we needed 15 more runs to trigger it. It became clear that the traditional ruletake action if there is a point outside the outer control linesdid not do a very good job of detecting that shift. in particular. and the geometric movingaverage or exponentially weighted movingaveragc (EWMA) chart. We are suggesting that the engineer should keep wveral charts simul tancously . John Copyright 0 1990 by John Wiley & Sons. 203 . we will rarely (one time in 400) stop the process when we should not. The Shewhart control charts are conservative in the sense that as long as the process is rolling along with the process mean equal to the target mean. INTRODUCTION In the first example of an xbar chart in chapter nine. We discuss three of these charts: the arithmetic movingaverage chart. Inc CHAPTERTHIRTEEN Further Topics on Control Charts 13. we now repeat the three rules that wcrc proposed. We are not advocating replacing the Shewhart chart with these new chartsfar from it.Statistical Methods in Engineering and Quality Assurance Peter W. we can use several other charts whose purpose is to enable us to detect more quickly changes in the process mean. One point outside the thrcc sigma limits. For convcnicnce. R u k 1 . 'The alpha risk is kept small. We begin the chapter with another look at the stop rules that were considered in chapter nine. M. which is the risk thiit wc will not stop the process when we should? In nddirion to the standard Shewhart charts. Two consecutive points either above the upper warning line 0 (twosigma line) or below the lower warning line. That problem was even more conspicuous in the example of the cchart in chapter ten.1. How about the beta risk. focusing. the cumulative sum control chart (Cusum).
so that the mcan changes from p to p + c r / f i . COMRINED RULES Rules 1 and 2 are rarely applied separately. Using the results that are derived in the next section. the word run in ARL stands for 1. More generally.30.204 FURTHER TOPICS ON CONTROL CHARTS Rule 3.8. But we need to ask a different question. rule 2 calls for a run of two points. In developing the A R L for the combined rule 2a. 13. A run of seven points either all above the center line or all below it. the ARL for the rule of seven is given by thc formula E(t)= 1 . The probability that the next seven points will constitute a run that triggers rule 3 is Y = (0. The probability that a point falls above thc ccnter line on the chart is now increased to 0.1587)(0. we follow the argument of Page (1954). we can show that when thcre is a shift of f l in the mean. . the probability that the next seven in a row will be above the line is p’. whose expectation i s the ARL when this rule is used. f is a random variable.8413)’ = 14.) What can be said about the probability distribution of f? The derivation is beyond the scope of this book.8413. . 0 The table given in chapter nine considered only one rule at a time. (Beware of the double use of the word run in this sentence. (0.8413)7 which we round to 15. rule 2a: Kufe 2a. Suppose that there has been a shift of + a / f i in the mean.(0. but we can derive a formula for its expected value. Suppose that it takes f points until a run of seven occurs. Rule 3 calls for action if there i a run of seven points. Take action if either there is one point outside the outer control lines or there are two successive points outside the warning lines and on the same side. if the probability of a point above the line is p.3. a 1 . DECISION RULES BASED ON RUNS 32 Rules 2 and 3 are both based on runs of points above or below a line other s than the outer control lines.8413)7 = 0. The normal prcmxiure is to combine them into a single rule.
(13.3. and we stop with L' = 1 . which we can call good. (iii) the next point is in the action region. The corresponding formula for the ARL is 1 . q (warning). there are three possibilities: (i) the next point is good. We combine these possibilities into a second equation: L = p( 1 + L)+q + r .3. and we expect L' more points. (13. (iii) the next point is in the action region. We combine these possibilities into the equation L = p ( 1 + L ) + q(l + L ' )+ f .3.2) Eliminating L' between the two equations gives (13.1) If we start from a point in the warning region.COMBINED RULES 205 The set of possible values of X can be divided into three regions. where p + q + r = 1 . and we stop with L' = 1. and we have the expectation of L more points: (ii) the next point is in the warning region. the good region consists of values of X between the two warning lines. let L' be the ARL when we start from a point in the warning region. and L ' = 1 + L . with probability p. the warning region consists of values between the warning lines and the outer (action) lines.4) .3) A more general form of the rule would require either a run of m points outside the warning lines or one point outside the action lines.q" r + pq'" ' (13. and the action region consists of values outside the action lines. warning.3. (ii) the next point is also in the warning region. Let the probability that a point falls in each of these three regions be p (good). If we s t a t with a point in the good region. and r (action). in which case we start again. For rule 221. and we stop. Lct L denote the ARL when we start the sequence of points from a point in the good region. The process will stop if there are two successive points in the warning region or one in the action region. there arc again three possibilities: (i) the next point will be in the good region. and action.
3085. r = 1 . A TABLE OF VALUES OF ARL Table 13. r = 0. calculate the AKL for the two rules We in ttic following way. MOVING AVERAGES The charts that are presented in this chapter for detecting shifts in the mean come under the general heading of moving averages.1 gives values of thc A R L for various values of the shift in the mean.F( 1 . Take action if there is either a run of seven successvie points on the same sidc of the center line or a point outside the action lines. Rule 3a is triggered by ii run of seven or a point outside the outer limits. 4 = .4. The values of AKL for rulc 3a arc shorter than those for rule 2a. we let the "good" region consist of the negalive values of X and Ict the warning region run from thc center line to the upper threesigma line.206 FURI'HEH 'I'OPICS ON CONTROL CHARTS We can also combine rules 1 and 3 into rulc 31: Rule .'Z r + py For rulc 3a. p = F( 1 . rl To calculate the average run length with rulc 3a for a positive shift in the mean. AN EXAMPIAE Let the shift be dcrltrii = O.S j = 0.5. but rule 3a has larger valucs o f the alpha risk.F(2.5.P4 13.1.103. = 1 y' 7 = 33.Sa/L/Si. We discuss three such charts.5) .06 r1.0062. 4 = F(2. we derive the formula that was used for rule 3 in section 13. I L = . . If we let r = 0 and m = 7. For rule 2a.0062.0606 ..?a.6853.3.S) = 0.2.6.4) with m = 7. 13.S) = 0. The basic Shcwhart chart with rule 1 uses only the single latest . whcnce I.393 In this example. 13. Then we apply equation (13. 0.9332. and p = 0.
for the ith sample. ARITHMETICAVERAGE CHARTS The first of our charts for detecting shifts in thc mean is the (arithmetic) movingaverage chart.0 5 observation to make the action decision and pays no heed to any of the previous data.I___. Arithmetic and GWMA charts have applications beyond traditional quality assurance. If k = 4. .2 I .o 1.8 Rule 23 57 I 273 142 77 44 26 Rule 3a 104 17 11 5 44 25 17 13 I0 8 85 8 6 7 6 . we plot not the value of y .. Onesided ‘rests d 0. the variance o f y is and V(y.6 1. Against this.) = dcnotcd by cr2.__ 2. In the special case of Shewhart charts. For this reason. 13.1.1 values of y .6 0.ARITHMETICAVEMGE CHARTS 207 Table 13. the ordinate for f = 5 is . notably in econometrics and in enginecring process control.4 0. The ith sample is the sample drawn at t .5. y. t h e rncan of that and the previous k . we change our notation to make it more general in its application. through the rest of the chapter. ‘l‘hc value of k used in the chart is the choice of the engineer: k = 4 and k = 5 are commonly used. EWMA) chart uses all the prcvious observations. along the horizontal axis. 1 l k . ‘I’he cumulative sum control chart uses all the observations that have been made with the same weight.7. but z. The arithmetic moving average uses the average of the k latest observations with equal weights. the observed response is denoted by y l . is measured along the horizontal axis. I. Sample number.8 1 . The movingaverage charts use the earlier data to different extents.2 0. and neglects observations made before then.0 0. but with weights that decrease as the points grow older. Average Hun Lengths to Detect a Shift d ( u h 5 ) in the Mean.4 I . Kule 2 employs only the latest point and its immediate predecessor. or time. . = i I . ui. The geometric moving average (or exponentially wcightcd moving average. We plot the sarnplc numher as discrete timc.
1.1. Arithmetic moving svcrage. this is the first average that contains a point with the “new” mean.10.50 + 10.28 + 10. the smaller changes are easier to detect with higher values of k ..10.+ X2 + X3 + X4 . The shift in the mean that took place after the thirtieth sample is now more apparent than it was with thc traditional xbar chart. There are 47 averages on the chart. but they are slower to detect smaller changes.7.1525. The twentyseventh average is (&.208 FURTHER TOPICS ON CONTROL CHAKTS Figure 13. .7.1 shows a movingaveragc chart with k =: 4 for the data of figure 9. but at a  0 10 20 30 Samplc number 40 Figure 13.8.70 + 9. For low valucs of k . beginning with 2.13 2 3( 0 442 p%? . + XZ8 + X2y + XjJ /4.13 4 4 = 10. these charts detect large changes in the mean quickly. On the other hand. . the outer control lines are pulled in to 30. Because we are now plotting averages of four samples.
say.8. and are at. z I = C ( ~ . (13. In Cusum charts. . Two parameters are involvcd in the test procedure: the lead time. i = . The engincer uses a Vmask. d. these charts may accidentally introduce some specious trends. 1976. Hence. The mask is placed on the chart with the V opening backwards (>). p. which is a piecc o f cardboard or plastic with a large Vshaped notch cut in it. Page (1961).. The apex of thc V is at the same level ( z value) as the latest point on the chart with a lead distance. 2 . The values ofthese parameters can be calculated as one would with a sequential test procedure. . and the angle. . we plot along the vertical axis the sum of all the deviations of the sample averagc from some predetermined value since an arbitrary starting time. i x .1) where plris sometimes called the target valuc. The horizontal line joining those two points bisects the angle of the V. The lead time is a function of the values of the alpha and bcta risks that arc acceptable. z . using previous data on the performance of the plant.CUMULATIVE SUM CONTROL CHARTS 209 pricc. 13. . . ) . ) .p l . ) . Cusum charts. are symmetric about the target value. . which is symmetric about the horizontal line. they are commonly called by the shorter name. I’hc enginccr should then choose the angle 6 so that the . corresponding to the center line on the Shewhart chart. The similarity to the twosided sequential probability ratio tcst with its Vshaped critical region will be apparent. they are established by cutandtry methods. it is usually the value of the process mean under steady state. which will happen if X has been rising or falling sharply. The procedure for using these charts amounts cssentially to performing a sequcntiai probability ratio tcst backwards. Thus. t 6. Often. . The choice o f the angle depends on how large a deviation from the target value the engineer is prepared to tolerate. between the edges of the V. 26. I . CUMULATIVE SZJMCONTROL CHARTS Cumulative sum control charts were introduced by E. It has also been shown by Nelson (1983b) that in their effort to smooth out actual time trends in the data. d. S. the apex of the Vmask is placed at ( d + t . the two critical values. Lucas has written several papers on the use of Cusum charts including parabolic masks (1973. The damping effect o using high k values slows the detcction of the f larger shifts. The use of the Vmask.~. The latest point to bc plotted on the chart is ( f . the two values that are used to decide that thcre has been a shift in the mean. implies that the engineer is carrying out a twosided test. The process is declared to be out of control if the cumulative sum disappears beneath the mask (crosses the edge of the V). 1982).8. p o .
28 1.70 9.30 1. we recall from section 12.1.442)2 (0.5. and j3 = 0. = 10.96 12 10.06 0.65 1s 10.88 2 10. The V in the Cusum chart is facing backwards.15 9.60 8 1. otherwise.30 0.9 that the outer lines of the acceptance region in the twosided sequential test are where k.35 1. = (1 .43 0. The chart. Suppose an engineer makes a cumulative sum chart o f the data in table 9..88 3 10.05. tan 8 = 612 will be taken.18 9 10 10.10.27 9.8. which were the process values calculated earlier in section 9.6.lO. is made with the following parameters: p.25) and places it on the chart with the apex at a distance d to the right o f the fifteenth point. In calculating the lead distance.210 FURTHER 'rorics ON CONTROL CHARTS edges of the V have slopes corresponding to L S / 2 .8.23 2.60 4 5 10.23 The lead distance. = arctan(0.25 Cusum C (X .01 11 11.5.1.88 7 10. tan 8 will be modified accordingly.79 6 1. and so we take For a cumulative sum chart that is equivalent to a onesided test.25 1.j 3 ) / ( n / 2 ) The lines intersect at .58 0.40 10.1and wishes to check for a shift in the mean after 15 points. 'The .60.51 13 11.76 3.So)2 in(36) = 5. If the scales on the two coordinate axes are the same. The engineer takes a mask with half angle 6.13 0. which is shown as figure 13.8. Example 13.13 and crL= 0. we can follow the argument of section 12.442.66 14 11. d. is calculated as d= 2( 0. The data points and the Cusums are as follows: Sample number 1 Sample average 9. and 6 = 0.l3)0.08 0. 0 = 0.
251 . y . .66> 1. Since z I Z=0.33.92.3 Sample number Figure 13. (13.251 . $)(1.92 .92 = 1. . whcn t = 13. the arm of thc mask is thc line y = y . When the mask i\ used at the nth point. Cuniulative sum chart.40 .08.33.1.1. the arm of the mask intcrsccts the graph between I = 12 and f = 13. 3. = 0. The fifteenth point has z = 3. The engineer concludcs that there has indeed hccn a shift in the mean. 0.251 .3 3.25.8. 0. it is = 0. When the Vmask is positioned with its apcx a distancc d ahead o f the fifteenth point. 4 n . its arm will intcrscct the Cusurn graph if there is a value of 2. and so the equation of the lower arm of the mask is the line that passes through the point (15. below 0. lower arm of the V intersects the plot of the Cusums between samples 12 and 13.CUMULATIVE SUhl CONTROL CHARTS 211 4.nFor this example.251 .1.1.00 + 5..7 27 .60. i. When I = 12.2) n + V .08 and z l l = 1. we have 0.92= 1.I.23) with slope 0.7 1.251 d).e.23...8.51< 1.1.
3 0.35 15 Cusum 4 43.00 16 10 5 44.  6d 2 (13.35 13 42.60 .1.10 3 42.25 1.15 0. z . = z. Table 13.25t.25 43.7 0.9.95 0.1. and ~ z I 2= 2.9 0.5 with Q = 0.70 18 43. .65 6.52.0 .40.3 I ..9.45 3.85 0.15 2 43. = 2(2.45 8.25 17 45 .05 2. The mask cuts the graph if at some previous time t. .7 0.15 0.10.20 45.2) Exumple 13.  !) = 2 (X. Data for Example 13.85 II 44.38) = y .1.45 1.9.1 Observation Deviation 1 42. I5 8. = 4.15 6 42. < Z .1 are daily production rates (conversion percentages) of product in a chemical plant.1.9.1.55 3.75 43. .7 Observation Deviation Cusutn Observation Deviation Cusum 7 8 44.0 and standard deviation w = 1.45 0.0 0.65 0.8.1) the arm of the Vmask becomes the horizontal line z = Z .40.9.7 0. The data in table 13.49 < .1.4 0. A HORIZONTAL VMASK FURTHER TOPICS ON CONTROL CHARTS If in the example o f section 13.10 12 44. The plant manager wishes to be able to detect an increase in production ratc of 0. In this example.25 3.55 42.9.7 0.7 0.9. Z.05 and p = 0.2 0.4 1.1.10.92 .4 2.40 14 43. The general form of equation (13.1) is z.9 2.45 3.60 9 I .30 44.4 0.7 1.212 13. The plant has been running at an average ratc p = 43.95 0. (133. we let z.55 0.0.
In thcsc charts. we plot the ordinate 2. Roberts (19%). (13.2) becomes z. The choice of r lies with the engineer.10 5 r 5 0. ) 18) = z. . Values in the rangc 0.99 The plant manager will report the shift in the rate after the sixteenth run.6.. 1.30 (13.43. W.1) against time.GEOMETRIC MOVING AVERAGES 213 Thc plant manager calculates deviations z.1.r)z. movingavcragc charts were introduccd by S.21 ( . = C ( x .25). 1 5In( .9. o r exponentially weighted. I.10.(1 .25). = z . GEOMETRIC MOVING AVERAGES Geometric.10.10. I Sample n u m b Figure 13. =: ry. Geometric average ( r 2 0. Equation (13. 0 13.10.2) are usual.
. the predicted valuc o f t h e first observation . points for r = 0.11. and may be estimated from the data by choosing the value of r that minimizes C s: Table 13. is the deviation. denote the response that is observcd from thc rth samplc. and 0. He lets v .= 9. EWMA AND PREDICTION J. S. r measures t h e depth of memory that the process itself has. like thc terms in a geomctric series.1) where 2.. one notices that this chart detects the shift in the mean faster than the Shewhart chart. = rY. The first obscrvation is 23. instcad of z. we get j . + r( 1 .2s for the data o f figure 9. + 4 (13.4) 13. Substituting for 6 . y.3) This is the same as equation (13. +. we notc that thc variance of z. and j . = ry. in terms of the responses: z.r)f.1 shows a geometric chart with r = 0.2 is good enough for most purposes. and 0.j . + (1 .3) To compute the control lines for the EWMA chart.11.1. or time series..6. 0.11. * (13. .214 FURTHER TOPICS ON CONTROL CHARTS The weights given to earlier observations drop as I incrcascs. ( 13.1) may be transformed to give z. Equation (13. . Again. of the predicted response at time I from the observed value.2.10. The engineer who wishes for a value with more decimal places can use quadratic interpolation with the values 0. is given by Thc sum o f this serics for large t is (13.1.10.8.1.11. + r( 1  r)’y. ‘The lowcst valuc of !: Of occurs at r = 0.2.10. 0. hence the name. The sum of squares increases as r becomes larger.1) with j .1).10. in equation (13. .3. The value 0.3. Huntcr (1986) approaches the EWMA from the point of view o f prediction in a stochastic process. ‘I’hesc charts are usually preferred to the arithnietic movingaverage charts.r)y. In this formulation.. instcad of z. .11.2.10. .1 shows these calculations for a set o f 20 . + . Figure 13. and he estimates the response at time r + 1 by the prediction formula j l + .
9 24.39 0.54 0.20 0.15 .06 0. 14 0.81 23.08 4. I 24.0 23.oo 0.32 24. ns 24. EXERCISES 13.70 23.91 23.04 24.0.45 1 . This is n o t true for the arithmetic .9 24.28 0.n5 0.79 33.62 23.03 24.94 24.0.EXERCISES 215 Table 13.32 0. 23.w 0.5 24.18 24.12 0.35 0.15 24.3 24.00 24.64 0.7 23.38 1.35 24.12 0.4 23.48 0.76 23.85 23.75 23.22 0.68  is the target valuc.1.02 24.18 0.M 0. (b) an arithmetic moving avcrage using four samples.22 24.80 23.(M 0.8 23.27 O.40 0. The changc of emphasis is important.71 23.74 0.  23.7 24.1 24.30 0. o n the same piece of paper.75 23.03 24.80 23.67 23.1.45 0.14 3.76 23.0.86 23.81 23.45 0.16 .1 23.92 24.90 24. For the data set in exercise 9.41 1 07 0.3 24.20 0.1 1 0.40 0. 14 0.12 3. It turns our attention to the use o f the quality control chart as ii tool for process control.80 23.20 n.74 23.24 3.81 0.14 0.80 23.56 0.78 23.27 I: 2.1 24.40 0.78 23.14 0.t)s 0.34 0.16 24.71 23.76 23.78 0.43 0.70 . We are seeing a drift upwards in both the observed and predicted values.23 24.63 23.79 23.3 24.66 23. KO 23.40 0.33 0. KWMA E ’ o r h s t Cornnutations for Estimating r 23.6 23.08 0. Note that in the arithmetic mean chart with span 3 and the EWMA chart the last point on the chart is close lo thc upper control line and the graph is heading upwards.8.18 24. Hunter (1986) suggests that we plot both y.31 0. The reader is rcferred to his article for his insights into the use o f EWMA for control purposes.1 24.1 make the following charts: (a) an arithmetic average using three samples in each average. (c) a geometric (EWMA) chart with r = 0.00 23. and j .31 n.68 23.24 0.4 23.12 0.84 n .56 23.5 24.11.16 24.84 23.3 23.70 23.29 0.18 24.21 24.70 23.39 0.76 23.75 23.07 0.69 23.20 0.25.73 23.78 23.03 0.
4.4. We wish to detect a shift in the mean of as much as 1 .871 0.855 0.5.905 0.806 0.863 0.3.3. and continue with the 16th. 13.8. etc. In this example p = 27.855 0.892 0. 13.910 0.2 as if they were single observations and make an I chart.786 0.5.78.909 0. assuming that you wish to detect a shift of 0.819 0.9.0 in the data of exercise 9.841 0.792 0. The process mean was 0.862 0.3.797 0..803 0.837 0. The following data set consists of 40 sample averages. until you spot a significant shift. Make the same three charts for the data in exercise 9. How many samples will you need before you detect a shift? assume that you wish to detect a shift of 0. Make the same three charts for the data in exercise 9.827 0.78.865 0.808 0.842 0.85 and u = 0.825.828 0. 13.7. Start testing at the 15th sample.846 0. 13. 13. Make the same three charts for the data in exercise 9.882 0.797 0. and also without a mask. Make a Cusum chart to test for a shift with a = 5% and = 10%.6. the standard deviation of the sample averages was 0.754 0.6 for the data in exercise 9. Repeat exercise 13.3.871 0. 13. assuming that p = 39.03. .216 FURTHER TOPICS ON CONTROL CHARTS averages with span 4.4.800 0.827 0. Repeat exercise 13.850 0.85 with u = 0.831 0.868 0. Treat the data of exercise 13.0 from the target value of 64. Note also that the one sample that falls above the UCL in the xbar chart does not cause an outside point on the average charts. using a Vmask.838 0.878 Make the same three charts as in exercise 13.6 for the data in exercise 9.834 0.837 0.801 0. 13.811 0.850 0.8 from a target value of 39.841 0.8 from an original mean of 13.2. 0.5.1.
. R. our data will be a set of n points. Similarly. Wood (1980).1. The reader who wishes to go further will find numerous books on the topic. We can only give an introduction to thc topic here. but not by some algcbraic function. Draper and H. The gcneral topic of fitting lines and planes to data by thc method of least squares is called regression. In both these examples. or else. Prominent among them are Applied Regression Analysis by N. We do have a considerable number of arts students who are weak quantitatively when measured against the average for graduate students. and science students whose verbal skills are low compared t o the graduate student average.. INTRODUCTION ln the previous chapters. Each observation was a single measurcmcntthe length of a rod. In this chapter. we turn to multivariate data: bivariate data in this chapter. tall men are heavier than short men and heavy men are taller than light men. the connection is not so clear. In the next two chapters. thin bcanpole. By and large. x. his scores in the Scholastic Aptitude Test (SAI') or the Graduate Record Examination (GRE) for English (verbal) and mathematics (quantitative). John Copyright 0 1990 by John Wiley & Sons. there are exceptionsthe short. the better students tend to have high marks o n both parts of the test. M. and his weight. Daniel and F. each one of which consists of observations on a pair of variables. when w e look at the whole range of students. fat man and the long. We may have taken n male undergraduate students and for the ith student. When the range of students is truncated and consists only of those who are accepted to study for master's or doctoral degrees. y. perhaps. and more than two variables in the next. the breaking strength of a strand of fiber. we have dealt primarily with univariate data. Inc CHAPTER FOURTEEN Bivariate Data: Fitting Straight Lines 14.Statistical Methods in Engineering and Quality Assurance Peter W. but this is not true over a narrow range. measured his height. the variables are obviously connected. Smith (1981) and Fitting Equations to Data: Computer Analysis of Muftifactor Dafa by C. That does not mean that they necessarily rank below the average 217 . Of course.
of X and Y is a scalefree version o f the covariance. or corrclation coefficient. The covariance is then estimated by ( 14.E ( X ) ] [ . Often.E( Y ) ] }* Y When the covariance is estimated from a data set. If there is a strong tendency for Y to increase as X increases. or cigarshaped.2. If the scale o f X is changed from meters to feet.2. if we are told that a student is six feet tall. of the j7) data. t o predict his weight. In this chapter and the next.t for E ( X ) and 7 for E( Y).2. the covariance is multiplied by (3. we substitute .15. Y will tend to decrease as X increases. and so the covariance is positive. cloud of points running from the southwest to the northeast.205) = 7. we use in the illustrations edited printouts from the Minitab softwarc package. The data will tcnd to fall in the northwest and southeast quadrants where the cross products are negative.for the man in the street. the cross product is positive.A'IION COEFFICIENT In section 4.23. The value of the covariance depends upon the scale of the variables. the data will tend t o fall in the southwest and northeast quadrants. If we take our sample of undergraduates and niakc a sciitter plot of height against weight. and from kilograms to pounds. A more practical use would be to predict the yield of a chemical process given the pcrccntage of impurity in the feedstock. It is defined by (14. The reader should check with the local computer center to find out what is wailable. if Y is the yicld of a product and X is the percentage of impurity.1) Suppose that we draw coordinatc axes through the centroid (i. 14. with the aim o f being able to use it to predict future values of Y given X. just that they rank below the average for the subgroup that we have chosen. we introduced the concept of the covariance of two random variables: Cov(X. and the covariance will be negative. as with height and weight. Our basic problem in this chapter is to fit a straight line to bivariate data. They vary only slightly in detail and in the availability of optional features.2) . it looks as if there is a straight line around which the points fall. THE CORHE1. The corrclation. In each of these quadrants. Every standard software package has subroutincs for fitting lines and planes. We would like to be able.28)(2. Y) = E { [ X . we will get an elliptical. On the other hand.
3) r has the same sign as the covariance.2.1. Y) 14. If the line is horizontal. CORRELATION AND CAUSALITY One should avoid the trap of confusing correlation with ciiusality.COKKELtZTION AND CAUSALIIY 219 It is estimated from a samplc by r . then when Y > 0.0.I simple algebraic exercise to show that if W = u + hX and u = c + d Y . They arc both positively correlated with tinic. we are as likely to have X >0 iis to have X < 0.1 . over the past half century. It does not follow that if X and Y are highly correlated. The population of Russia and the output o f cups and saucers in the United States have both increased. and. It does not follow that if r is close to unity.3. thc variables are independent. by and large. 4. it is .9626 and r = 0. Suppose that X takes thc values I . with one . r? = 0. If the points lie on a straight line. then X must cause Y.205. the relation between X and Y is close to linear. hence.28 itnd the standard deviation of Y by 2. the11 ' corr( U. On the other hand. if X and Yare independent. 0 5 r 5 1.98 Nor docs it follow that if r = 0. The covariance of X and Y is zero. The correlation coefficient is invariant under changes of scale and changes of origin. It can be shown algebraicidly that r must lie in the interval . The multipliers cancel out in the correlation coefficient. the covariance was multiplied by (3.1. the correlation is zero. U ) = corr(X. We have alreiitly seen this in example 4. where (14. and their corrclatiori is x r o also. the sign of r k i n g the same a s the sign of the slope of the line. but the standard deviation of X was multiplied by 3. the correlation is 5 1 .205). More generally. or vice versa.3). 2. 3. One must he careful of converse arguments here. I'hcn wc have i d X Y 1 1 2 4 3 9 4 16 5 25 Applying equation (14.28)(2.2.O. I n the case where we switched from meters and kilograms to fect and pounds. and 5. and that Y = X'.10.
” The name regression has stuck to the technique and is generally used by statisticians to describe fitting lines and planes to data. The numbers of both men and women have increased. perhaps. The weights o students who arc 70 inches tall vary about a value E ( Y I X = 70). for Y. FITTING A STRAIGHT LINE Suppose that we have a set of n bivariate data points. we may. University enrollment in the IJnited States has increased steadily since the end of World War 11. Let the ith data point be ( x . There are numerous examples of these nonsense. but we could use thc enrollment figures for inen to estimate the female enrollment. and if X is much cheaper to obtain than Y. (14. in particular of the “law of universal regression. ) .4. We propose to fit a straight line to thc data by the method of least squares. The noted English biomctrician Sir Francis Galton fitted a straight line to the plot of the hcights of sons versus the heights of their fathers in his studies of inheritance. = P.3) We now make our first assumption about the behavior of e l . ) . . Our model implies that thc expectations E ( Y ( X ) lie o n a straight linc.1) The cocfficients (Y and /3 are unknown constants to be estimated from thc data. 1 . the values of x . represents the random noise in the system.4. Y. but we cannot rationally argue that the one caused the other. This line is called the regression line of Y o n X . or surrogate. or vice versa. with the primary idea of prcdicting the value of Y given the value of X. It would not be appropriate here to speculate on whether the increase of women caused an increase in male enrollment. = CY + P x . if X and Y are two chemical measurements that are highly correlated. + e.4. The model is y . If we take several observations at the same value of X. feel comfortable about using the cheaper X as a proxy. = u + px. + e. = E ( Y I X = x . the observed f values of Y will vary. that is a random variable with zero expectation. (14. We can write 1 .4. correlations. y . are taken as given. Then p.2) and then Y . . * (14.220 RIVARIATE DATA: FITTING STRAIGHT LINES another. t. or spurious. 14. namely. On a more practical note.
and we reject the hypothesis that the slope is zero. It is not error in the sense that someone has made a mistake. is comfortably larger than 2. The standard deviation of Y refers to the standard deviation of the error terms.13. X .12. Y 110 70 124 130 IS. Then thc coefficients in the equation are repeated and their t values are given.08 3. The calculation of those r values is discussed in section 14. IhO X 3.36 3.32 3. if our model is wrong. the t value for the constant term is small. X. e. ‘Tofit the line on Minitab.11. and we do that later. 11. The printout begins with a statement of the fitted equation.61 2. telling the program to regress the Y variable that was in C1 against one predictor variable.5. 14. Y. On the other hand. The data are the number of calories. 96 7.1 appeared in the Austin Aniericun Stutesniun newspaper in the fall of 1978. 9.1. 8. For the moment.0.1. e. 68 70 97 134 140 X 2.89 3. 5. 12. and the pcrcentage of alcohol. Beer Data Y 1.81 3. 96 4.14. for the regular and light types of beer brewed by eight different producers.40 3.43 2.79 3. This suggests that it might be appropriate to fit a line through the origin. 140 16. The first eight beers were classified by the brewers as light beers and the last eight as regular. The t value for the slope of the line. 129 X Y 2. 140 13. 134 10. we entered the data in columns 1 and 2 of their work sheet and gave thc command REGRESS C1 1 C2.76 3. we assume that our model is correct.” We are not going to discuss that situation yet. per 12ounce can.5. The derivation of the square of the correlation coefficient and the table of sums of squares that follows it arc in section 14.5. 5.22 3. 6. If our straightline model is correct. as it would be if we were fitting a straight line to points that lie on o r about a parabola. A plot of the data is shown in figure 143. represents random variation.2. It is discussed in scction 14.27 3. will be a combination of both noise and “badness of fit.52 3.5. 14.89 .in C2. An cdited version of the printout is given in table 14. The details of the printout are discussed as this chapter proceeds. Table 14.90 3. AN EXAMPLE The data of table 14.17 3.A N EXAMPLE 221 It is an unfortunate accident of history that we usually refer to ei as the error.
and the t value for the slope falls to a nonsignificant 1.00 Y PRED.INES TIE REGRESSION EQUATION IS Y = 23. The point with the R rating.5 4315. 27.2 (BACK SOL. DEV.00 70. numbers 5 and 6 (number 2 conies very close to earning an X rating).5 308.14 THE ST.UTION FOR SPECIAL POINTS) ROW 5X hX 13R X C2 2. For the whole group o f 16 beers.41.654 R = 0. There are two o f them. thcy stand a long way from the main group.384 S?.222 .65 0.71 70.8 ss MS = SS/DF 81 38. .66 0.08% alcohol: number 13 had 140 calories. DEV.I _ BlVAKlATE DATA: FITnNG STRAIGHT I. OF Y ABOlJT REGRESSION LINE IS S = 17. The X ratings denote beers that have especially large influence on the fit.27 3. 2.86 35 * 83 2.14 and 14. 3%.054 TRATIO= COEF/S. say. those three beers are below specification. has a bad fit to the data.2.3 COLUMN  Table 14. the correlation coefficient between calories and alcohol content is r=0.18 9. 14 d.32 2.15).56.f.30 .00 140.12 The printout cnds with a listing. beer 13.4 X X C2 COEFFICIENT 23.25 8. 0. RSQUARED = 0. with somc dctails. the correlation coefficient falls to 0.RES.5. When the three lightest beers are excluded.04 4.D.2 12453.65 104.71 0. they pull the line down.81.47.8087 ANALYSIS OF VARIANCE DUE TO REGRESSION RESIDUAL TOTAL DF 1 14 15 8138. The fitted line predicts only 104 calories for beer with 3. They are very weak beers.17 ST. Y RESIDUAL ST.30 41. One argument would be that we wish to confine our investigation to beers o f a certain minimum strength.86 5. Printout of Minitab Regression for the Beer Data (Fdited) + 41. It seems to have too many calories for that amount of alcohol. 9. PKED. DEV. Y VALUE 72. o f three “special” points (see sections 14. One might argue that the three wcak beers should be excluded.08 Cl 70. OF COEF.
2). (14.6. take for & and residuals: = y. . .6.THE ML'IJIOD OF LEAST SQUARES 223 Alcohol Figure 14. The method of least squares is to the values that minimize the sum of the squares of the s. .6.. (14.5. j . Substituting those estimi3tes in equation (14.4.2) T h i s difference is called the ith residual. . = & + px. I ) The difference between this estimated value and the observed value is the deviation . = 2 ef .3) .6. THE METHOD OF LEAST SQUARES Supposc that we estimate a and p hy & and b. Regression of calories on alcohol 14.j . we have an estimate of p.1. (14.
6.6.p c x: .a.6.11) j = 7+ p ( x . gives two equations p. c xjy. squares.e.X) . (14.6) are called the normal equations.6. .6.6.7) by C x . i. If we multiply equation (14. and subtract it from equation (14.x c = PS. The quantity S is called thc sum of squares for error. (14. in turn.8) We rewrite this equation as s. ) = 0 .7) and obtain &=jBX. This says that (2.6). .6) and dividing equation (14. (14. or the residual sum of . the fitted line passes through the centroid of the data.12) .6. = ( yi . ( y ..c$ .63) and (14.5) by n .6.)' . (14.6. = c 4.4) Differentiating S . The prediction line can now be written as (14. . XI) (14.f i x . XI& (14. which simplify to C yI=nci++Sxx.& .65) (14.6. with respect to 4 and to . .PX. and equating the 2 c 2 (yi ..6.6. derivatives to zero.7 )= p c (x: . we have s.. =0 c x ..6..2).6.7) c so that x . we have u) q= 6 + px .10) We then substitute this value in equation (14. lies on the fitted line. ( y . Equations (14.224 BIVARIATE DATA: FITI'lNG STRAlGHT LINE.9) (14. Substituting the estimates & and fi in equation (14.px.
= 84.0 1. and 6= v = 10. j = 114.658. y ..2 +3.0 2.9 25.0 2.1.1 19.3844. X = 3.0 YY + 8.0.4 X = 4.y. wc also see that the multiplier on the right of the equation is s. c (x. Sx. Example 14.0 5. S. = 4.0 0.0.7.X? .658/4.0 7. = y .1 1.1 .752 = 41.0x. = 196. so that 22.  x.8. .752. and S.7. 6 14.7 31.1 + 3. . & = 23.389.0 21.6 +9. .0 = 10. .X)(Y.0 X 13.2 27.0 3.0 1.8.niy x c y . in the algebraic identity. .Y ) . Then = 196.y. We can show in the same way that and so the term on the left of equation (14.9) is s.Y c x.6. I . = 28. + w +nij Z: x .nxy .1 +3.o +2.0.o 6.0.297.n i j . (14.7.0 3..12. . ) (14. .y) = = = x c x. A SIMPLE NUMERICAL EXAMPLE This examplc has been constructed to illustrate the ideas of the previous section. = c = (x.0 +5.2) 14.3 16.  X ) ( y .A SIMPLE NUMERICAL EXAMPLE 225 For the beer data. and S..0 4.8 6.1) If wc set x . . AN ALGEBRAIC IDENTITY c (x. j = 22. The prediction line is fi = 3.5 P xx  3.875.1 0.
(ii) .1. (i) This assumption is cnough t o makc an unbiased estimate of the slope of the line.0 4.) = 0 for h # i .1 22.9 25.8). = 0.42. .8.8.7 31.0 5. . U 14. The obscrvations are uncorrelated: E(e.1 19.12) since f?.)= 0 .4 The prcdictcd valucs and thc residuals are shown in table 14.2) So there are two restrictions on thc residuals.1) Also.6.0 6.1 19.0 0.1 28.2 27. and s o  X) 3 (14.P ( X .8.0 13.1 0.226 'I'ahle 14. ) I 2.0 .1 16. THE ERROR IERMS So far we have only made one assumption about the nature of the error t e m s: E(e. = y .1 0. I 31.5 _ 0.2 i0.2 1.1 d t 0.f.8. have been lost because two pirrametcrs. I 25.0 21.0 3. have bcen cstimatcd from the data.j . Note that the sum of the residuals is zero. = ( y .0. from equation (14. One can as well say that 2 d.7 ). =0 .9. X j .4 +0. 7. The residual sun1 of squares is S. and thc sum of squares o f the rcsiduals has n .e.3 16. That follows from equation (14.2 degrees of freedom. I BJVARIATE DATA: FITTlNG STRAIGHT LINES I  X Y t 13. ~2 and p. Three more assumptions are ucually added. (14.
(14. The last assumption is that the errors are normally distributed. we develop the properties o f the estimate. and (iii). i. we write a.. it is a linear combination of the observations. The cstimatc.. We now show that a is an unbiased estimate of the slope. o f the slope o f the line. fi. is a linear estimate of the slopc.'THE ESTIMATE OF THE SLOPE 227 The observations have the same variance: V( y .10. 14.10. The variance o f fi is computed with assumptions (i). THE ESTIMATE OF THE SLOPE In this section. (ii). ) = E ( ~ . Even if the observations . y. This implies. For convenience.1) This is intuitively rcasonablc. Note that in showing that is an unbiased estimate o f the slope. the more precise is the estimate o f the slopc. we have only used the first assumption about the error terms. that and that the observations are independent.e. /? is a linear combination o f normal random variables and is itself normally distributed. in conjunction with the other three assumptions. (iii) Assumption (iii) is called the assumption of homoscedasticity. IJnJcr the normality assumption. The wider the spread in the values of x. Z ) = {r 2 for ail i .
The residuals reflect only the noisc in the system. P.2)u’ .” It is 4315 with 14 d. tTESTS If the errors are norm.2 d.)= ( n . 17..illy distributed.f. the sum of squares of the residuals appears on the second line of the section headed by “Analysis of Variance. so that is an unbiased estimator of u2. the quotient S e / u zhas a chisquare distribution with ti . the residuals provide an estimate of the error variance. of the slope is a normal random variable.12.12.2 d. and the locus of E ( Y ( X ) is a straight line.1) and (14. The statistic I= = P (14.12.2) .1) has a f distribution with n . appears earlier in the table as the estimate of thc standard deviation. 14. We have seen in equations (14. P . which leads to the estimate 4315/14=308 for the variance. E ( Y ) = E ( Y ) at each point. In the printout for the beer data. With homoscedasticity (homogeneous variance).8. 14.P fi (14..228 BIVARIATE DATA: F I l I I N G STRAIGHT LINES are not normal. It follows that t = P .f. ESTlMATlNG THE VARIANCE If our original model is correct. The square root of that quantity. We have already seen that the estimate. that E(S.6.11.8. This t statistic can be used to compute confidence intervals for the “true” slope.2) that there are two linear constraints on the residuals. after some algebra. It leads to the tests that are derived in the next few sections. The reader should note that this is the first time that we have made use of the normality assumption. It can be shown. the central limit theorem justifies us in acting as if wcrc approximately normally distributed.f.
61= 5. to the estimate of the variance. to arguing that the cvidence is not strong enough to reject that null hypothesis. If the “true” line is indeed horizontal. We recall from equation (14. Similar calculations lead to the tvalue for thc other coefficient.X) . the model becomes E( Y) = constant . worth of noise.1 d. This is the Ftest.2 d.0 in absolute value as significant.13. Fitting the predictor. The effectiveness of the reduction. .13. It is called the tvalue for fi. Sy.. + s.3) We now look at the sum of squares for regression more closely. (14. or.. Smaller valucs of I would lead to accepting the hypothesis that the true coefficient is zero. which suggests refitting a line through the origin. . . corresponding to the slope. The tvalue for ci is only 0..13.e.s.86.& X I . for S. to the error variance.2) The total sum of squares. X. which means that X is of no use as a predictor of Y.f. 14. (. about the horizontal line is S = 2 ( y . reduces the scatter to S.14.y)’.6. SUMS OF SQUARES Another approach to testing thc significance of the slope of an estimated line is through the analysis of variance. The . at least. That leaves n .SUMS OF SQUARES 229 can be uscd t o test the hypothesis that the “true” slope is zero. =s . = 9 4. p. (14. can be measured by comparing S ..13.1) s. describes this division of S into the two components: s. The reduction is obtained at the cost of 1 d. the apparent reduction should just be equivalent to 1 d. and the constant is estimated by The scatter. . ci. . A common rule of thumb is to regard values of r that cxceed 2. In the beer data. i.f.14. or sum of squares of residuals.12) that 9.f. : amount of the reduction is called the sum of squares for regression and is denoted by v.3844 = 17. therefore. The analysis of variance . has n . = s. which confirms that the line of best fit is not horizontal.f. the 1value for the slopc is 41. Wc argue that if fi = 0.
= ( Y . The analysis of variance also leads to the correlation coefficient. described earlier. is to compute S.2114 = 308. = Y .453. which is explained by adding X as a predictor.6) and so the Ftest is equivalent to the ctest that was described earlier. . and S R and subtract. p2ss. = 12.2) d. 7 S (14. . s .8 .2psxyt.7) This gives a physical interpretation of the correlation cocfficient i n terms of the fit o f the linear model.PS. . n .654. the fractional reduction is scalefree: ( 14.5. For the beer data.13. the easiest way to compute S.5 = 4315. = (41.13.5) Under the null hypothesis that P = 0.8 = 0. . S.. whence r 2 = 8138..4) Cornputationally.230 BlVARlATE DATA: FlTTlNG S l M I G H T LINES so that 2.(F. is F .3844)(196.13. This quantity depends on the scale of Y. The F statistic.V and ).F. = = s .S R.2 and s2 = 4315.55112. (1413.453. However.658)= 8138.f. But we can also write (14. .V ) s. Fhas the Fdistribution with (1. We have derived the sum of squares for regression as the reduction in scatter.8138. . S. .2.
and it is flagged by t h e letter K . They earned these ratings because they were a long way from the centroid o f the tiara base. Such action can sometimes be justified. the two weakest beers received X ratings to indicate that they had a largc influence on the slope of the prediction line. Thc actual criterion of how far is a long way is a matter of choice. In the particular cxample of the beer data. in retrospect.0 for the standardized residual indicates a point whose tit is questionable. back solution for the “special” In the last part o f the printoutthe pointsthere is a column for the standard deviation of 9. obtained from equation (14. This kind o f action has to be taken on a casetocase basis. POINTS In the beer data. 14.1) that V(9 . each residual is divided by the estimate of its standard deviation. they were considered to be outside the limits within which the scientists wished to work in future. The quotient is essentially a [value for that residual.The last column contains the standardized residuals.10. Then they might be justified in using a prcdiction line based only on the stronger beers.14.14. Beer 2 missed an X rating by a whisker.X)‘V( p) (14. In thc current version of the Minitab package. and only with due consideration. STANDARDIZED RESIDUALS It follows from equation (14. . This computer run was made with an earlier version of Minitab. In that situation. the authors have changed the cutoff value and those two beers no longer earn that rating! One should be careful not to fall into thc trap of automatically dropping from the data any points that have X ratings.14.V(y ) + ( x . INFI. A value greater than 2. when Y is predicted by several X variables.1) with s substituted for the unknown u.). on the argument that they must surely be cxtremc. IS. In the beer data. In this column. ). it is not always easy to spot the influential points with only a visual examination of the data.INFLUENTIAL POINTS 231 14. ii case could be made for setting those points aside if. . point 13 earns an R rating. Its choice of cutoff point gave X ratings to beers 5 and 6.1) A similar formula ciin be derived for V(t?. A similar rating can be used in multiple regression.UENTIAI.
86. at X = x . obtained from equation (14. A confidence interval for the "true" value. so that the estimate of the variance has actually fallen from 308 to 302. constrained to pass through the origin.3) This can be appreciably larger than V ( j o ) . .1). Notice the change in the analysis of variance table. y .2 to 4540. It is important to remember that there will be noisc in our observation too. C y : . V(j .1) This is the formula that is used to calculate the entries in the column headed "ST. but there has been an increase in thc number of d.2) where t* is the tabulated value o f t with n .Lo + 1*s* .f. This is done in Minitab by two commands. It takes its minimum value when xl.and the confidence interval is corrcspondingly wider. 14. we niay wish to compare the observed value with our estimatc. Y" in the printout. X. ) increases as x g moves away from the center. can be obtained in the usual way by j " . y o .of the data basc.16. from 14 to 15.232 BIVARIATE DATA: FIITING STRAIGHT LINES 14. and S* is the standard deviation of PI. ..6. Clearly.16. If we now take an observation.16.16.2 d. (14. N o reference is made to the correlation . the tvalue for the constant term is 0. The sum o f squares of rcsiduals has increased from 4315. The printout appears in table 14. first REGRESS cl I c2:.16. LINES THROUGH THE ORIGIN In the printout for the beer data. DEV. it has not been adjusted by taking deviations about the mcan value. when x = xn.t*S* I 5 j" /.17. . I ..f. = X. followed by NOCONSTANT. pl.7. CONFIDENCE INTERVALS FOR A PREDICTED VALUE Suppose that we predict the value of Y when X = x 0 . and so we fit a new line. The variance of the difference is (14. PRED. We cannot reject the hypothesis that the constant term is zero.. The variance of the predicted value is (14. The sum of squares for Y is now the raw sum of squares.17.
THE REGRESSION EQUATION IS Y = +34.7 coefficient. DEV.4 302. bccause there is x (JL (14. and so we . TRATIO= OF COEF.17. x. The least squares estimate of the slope is obtained from the single normal equation.6 X1 COLUMN NOCONSTANT XI c 2 COEFFICIENT 34.2) We have already mentioned that the residuals (when there is a constant term) are constrained to sum to zero. Minimizing the residual sum of squares leads to (14. Now we ask if it is an improvement over the line y = 0. NOCONSTANT.588 ST.90 ANALYSIS OF VARIANCE ss DUE TO DF REGRESSION 1 219053. and that V ( f i ) = ./S. Beer Data (Line through the Origin) HEGRESS cl I c2.17. 14. RESIDUAL PLOTS The residuals are no longer constrained to add up to zero.1.17.1) It is easily shown that fi is an unbiased estimate of the slope. It is the loss of that constraint that adds one more degree of freedom to the error term. 1.18.286 26.RESIDUAL PLOI'S 233 Table 14.6 TOTAL 16 223594. We are no longer asking whether including X as a predictor is an improvement over the line y = 9.0 MS = SS/DF 2 19053.D.4 RESIDUAL 15 4540. That is the only constraint. COEF. no normal equation that corresponds to a constant term.
) One can notice the phenomena that have just been mentioned.rps there really is a straight line passing through. all t h e points but one. In that case. Before plotting on the computer was as convenient as it is now. perhaps.1. If the residuals are alternately plus and minus. on H straight line. that one point is an outlier somewhere above the line. and perhaps something else. Roughly half the residuals should be positive and half negative. or close to. Suppose that the data are entered in ascending order of X . Do the residuals seem to get bigger as X increases? Do they get bigger as Y increases? Does this mean that. (A plot of the residuals for thc beer data is shown as figure 13. Some like to make a normal plot of the residuals and see whether they lie. we may have fitted a straight line to ii sine curve! If there is a string of positive residuals followed by a string of negatives and again a string of positives. . perh. it is easy to plot the residuals against X.234 BIVAKIATE DATA: FlTTlNO STRAIGHT LINES should expect them to be more or less randomly and normally distributed. If you have one large positivc residual and the others are small and negative. we were wrong to think that the variance was constant:' Maybe the variance is a function of Y. the ohscrvation with the smallest value of X being number 1 and s o on. these precautions of checking the randomness of the plus and minus signs were standard practice. A glance at the list of residuals should show a random ordering of plus and minus signs.18. we may have fitted a straight line to a parabola or a hypcrbola (see the example in the next scction). Nowadays. more o r less.
19. and that there is sonic underlying linear connection between the two variablesnot necessarily a functional relationship. we have assumed that our model is corrcct.2. If one extends the line far enough. shows that the end points lie above the fitted line and the points in the middle lic below the linea distinct pattern.1 Wcck ~ Dcf. A company sets up a new process for producing parts. A plot of the residuals against X. The reader may wish to refer t o a book that specializcs in regression. Example 14. 27 1 tl 10 Week 4 Def. A plot of the data. The points appear to lie on a fairly smooth curve that looks like a rectangular hypcrbola. one might bc tcmptcd to think that the straight line is the "correct" model. OTHER MODELS So far.OTHER MODELS 235 one should be using weighted regression or a transformation. Does this mean that fitting a straight line to this data is uselcss? Not neccssitrily. The management is delighted to see that the number o f defectives decreases.2 shows a summary version of the printout whcn a straightlinc model is fitted to the data. The straightline model gives a reasonable approximation to reality a s long as one stays within limited ranges of the X variable.19. 41 21 IS 8 6 Week 6 10 Dei. reveals this to be incorrect. 33 18 3 7 11 8 15 23 12 16 30 14 Y 19 7 S 20 24 6 4 17 21 25 3 22 26 14 18 11 10 4 3 .19. In the 1970s. some economists plotted Tahle 14.19. SO Week 5 9 13 Def. There is no question that Y decreases as X increases. they start to take a sample of 1000 parts each week and to record the number of defectives.1. The data sct is given in tiible 14. we fit a straight linc where a hypcrbola is more appropriate. With s o large a value of r as 0. A large value o the correlation coefficient tcnds to f give credence to that idea. After two weeks. In the following example.1. The consideration of that complication is beyond the scope of this book.19. Table 14.19.875. 'There is enormous potential for trouble if one tries to extrapolateto use the equation to predict values of Y outside the range of thc data. 14. 65 . which is better than even the most enthusiastic quality engineer would predict.1. tigure 14.19. figure 14. such as Draper and Smith (1981). it will cross the X axis and will predict a negative number of dcfectivcs.
2.00 X.004 R the price per barrel of crude oil against time. Regression of defectivcs on week. It is also very. 1 /X.36 . and.1. + 2 1 6 4 x) 1 Weck Figure 14. squarc roots.236 3. fitting a straight line with 1 /Xas the predictor instead of and X.19.875 = 8. Predictor Constant X S Coef 46. linear extrapolation.22 0. reciprocals. etc. . Extrapolation. in particular.94 with s = 1.9Y3.48 tratio 12.236 Table 14.948 and r = 0. very dangerous.792 8. We can fit a hyperbola to the data by introducing a new variable.2 BlVARlATE DATA: FI’ITING STRAIGHT LINES The regression equation is Y = 46.3586 2. That straight line extrapolation forecasted that oil would cost $80 a barrel by 1990. some of us are finding that hard to believe now.001 74 R’ = 0. Engineers are used to converting curves to straight lines by using special paper in which X and/or Y are replaced by their logarithms.19. is very easyyou just keep drawing thc line. The new “line” is Y = 3.766 Stdev 0.
however.2. Multiplying both sides of the equation by x gives the rectangular hyperbola: xy + 3. make it linear by a transformation. by which we mean that the righthand side is not a linear function of the two unknown parameters. We can.94x = 216. This is obviously a good fit to the data.20.20. None of the usual assumptions has been violated by using l / X instead of X as the predictor of Y. this model is not a linear model. and d = /3. but it does enable us to fit the model.TRANSFORMATIONS 237 d s 3 Week Figure 14. Residuals plot for defectives data. We fit the straight iinc and then .1) As it stands. We take natural logarithms o f both sides of the equation and get Z=c+dX (14.19. a and p.2) where Z = Iii(Y). TRANSFORMATIONS Another model that might be appropriate for a data sct that looks like this is the exponcntiai model: Y = (Yepx (14.20. ‘This procedure does some violence to our assumptions about the error term. c = In(a).81 0 14.
. b. T h e new regression equation is In(Y) = 4. + 14. that minimizes the residual sum of squares in thc loglog world is the same as the pair that minimizes the sum in the untransforrned world.20. and the reader is again referred to Draper and Smith (1981). . Fit a straight line. have C x = 200. .0.0. We can also get the equivalent of loglog paper and fit the model y = NX h by taking logarithms of both sides and fitting the line w=c+bz where it’ = In( Y ) . 3 to the following data: 5 X= Y= 1 2 4 6 7 8 9 1 0 13 17 19 12 14 29 34 27 36 33 Plot Y iigilinst X and plot the residuals against the predicted values. and z = ln(X). Fit the line I y = Lr p x to the data and tcst the hypothesis /3 = 1. we.n + px.0 against thc alternativc /3 # 1. EXERCISES 14. 9 . which corresponds to The tit that we have made is equivalent to plotting the data on semilog papcr and drawing the best line through thc points.9 ) = 87. Test the hypothesis p = 2 against the alternative p #2. 2: ( x 2)’ = 92. y = 810. a = exp(c). C ( y . ) . Compute the correlations between y.1).2..X( y . For a set of 20 observations. 2. It does not follow that the pair of coefficients.124x. j . y)’ = 104. Note that thc corrclation between the predicted values and the residuals is zero. c = I n ( . and 2. Thc problem of finding the best fit in thew nonlinear problems without using linearizing transformations is full of mathematical complexities.238 BIVARIATE DATA: FIITING STKAIGll’l’ LINES take exponents to return to equation (14. The results that wc get with these transformations are only approxirnations to the best fit.28 . y . and 2: (x .1.
The residual plot shows a clear parabolic pattern. Since the data in exercise 14. This illustrates how extrapolation from a model that seems to be plausible can be a trap. m.EXERCISES 239 14. The table shows the miles per gallon for the ten most economical cars in the 1990 EPA fuel economy survey.4. . You have fitted a straight linc to a parabola! Estimate from your fitted line the values of Y when X = 13. there is an estimate of pure error between thosc duplicatcs.4 has repeated observations at city = 46 and 38. highway 58 52 50 50 46 43 40 38 38 37 50 49 44 42 39 43 Obtirin a least squares prediction line to predict the highway mileage from the city mileage: highway = Po + f3. This can be used to check your modcl. city 53 49 46 46 m. PUKE. X = 1 3 2 3 3 5 4 5 6 7 8 91011 59 75 93 12 Y= 9 15 23 33 45 113 14. 14.. i. the line is not an implausible model at first glance.g.915. Make plots of Y against X and of the residuals against the predicted values. Test that hypothesis.5. Test that hypothesis. and 15.g. You will see that for x > 4. as reported in thc Ausrin American Sturesmun.3x + 5.(city) It could he argued that the line should pass through the origin.p. Fit a straight line to the following data.3. Note that R 2 = 0.p. 14.e. and compare them with the actual values from thc quadratic. Verify that for each point y = s’ . which is very impressive. Readers who are using Minitab should use thc commands REGRESS Cl I C2. to test whether the straightline modcl is adequate.
60 15. which is not trivial.531. Y (300K) 9020 8620 9670 9830 11200 7880 7.45 1.50 22. i. 10. In exercise 14.50 7.2. in the model.10 22.20 16 17 18 13 14 15 19 12 10. It can also be In shown that corr(2.20 3.70 17.80 18.80 2.70 15.. reasonably drop it. 11.50 21. 3.50 23 24 22.78 4. 9.35 23. Fit a straight line to predict the mobility at 300K from the mobility at 77K. 1 X = 0.50 13. 12. 6. when tlie line is not forccd to pass through the origin.10 6 7 8 9 10 3. The measurements in mcters for 32 trees follow.20 1.10 15. Suppose that you felt comfortable in arguing that for the data in exercise 14.240 RIVAHIATE DATA: E’III’ING STRAIGHT LINES 14.40 25 23. a. therefore. ) = 0.10 9.35 9.00 22 20 21 19.30 19.90 Y = 2. Fit it line Y = a + p x t o the data and comment about your results.20 3.40 3. 5.00 2.8. Two scientists wish to develop a formula for predicting the height of an aspen tree ( Y ) standing in a forest from its diameter at breast height. y. y.50 23.60 22.) = G.70 5.80 3. j .50 16.80 22. 2. exercise 14. engineers recorded the van der Pauw measurements of mobility at both 77K and 300K.9. Fit that model and compare its fit to that of tlie first model. 8.40 3.80 17.40 11.20 9.50 23.50 .10 15. rather than y.I . A plot of the data will explain why the twelfth sample has large influence and also a large residual. 14.80 27 2 8 25.40 15. 4. X (77K) 3270 3960 3920 3860 3330 2330 Y (300K) 3760 10460 9970 81 10 4890 2200 390 3940 3980 3930 3980 3770 14. Is the fit markedly better? 14. the twelfth sample was far below the range o f interest and that you might.40 15.60 2.50 19.20 27.. X (77K) 1.35 5.) = 0.30 5.6.6. Prove that this is always true when there is it constant term.40 11 2 3 4 5 1. This a reason for some statisticians to prefer to plot i against 3. In a study of the growth of InP. Interpret this. you showed that corr(2.718 and so corr(6. It is suggested that they should use v3 rather than x as the predictor. r’ = 0.e.00 13.50 26 25. Do that and fit a new line.90 16.7.00 22.10 22.10 23.
40 22.EXERCISES 241 29 30.0. Show that p is an unbiased estimate of the slope. ? when X Z O .3 93. The data follow with F1 as Y and F2 as X.2 87.1 7 8 98. 14.5 25 91.4 101.7 98. 14.4 18 11 86.0 81.5) 85. .8 95. D.9 30 31 32 28 29 27 93.8 23 24 21 22 20 85.4 99.7 93. In an octane blending study that will be mentioned again in the next chapter.3 07. Express the regression line in terms of the new units and prove that the correlation between price and length is unchanged.4 94. Suppose that the price had been computed in Swiss francs (1 S.3 1 w .4 84. Another estimate of thc sbpe o f a line through thc origin is / = X/y. Fit the model y = p x .4 91. X = 105.3 99.9 10 105.0 85. where y is the cost in U. Fr.10 32.10. C y = 60.5 .3 91. Test thc : hypothesis p = 2.0 87.5 106.8 19 14 15 13 12 83.6 97.83~ .80 23.8 96.7 103.10.0 90.4 88.52 +0.0 against the alternative p f 2.7 87. Calculate its variance and show that it is greater than the variance of the least squares estimate. C xy = 200.2 87.6 16 83. R. and 2 y’ = 801. =$0.8 85.50 32 35. C x 2 = 91.1 86.2 101. dollars.62) per meter.8 89.5 97.2 93.7 94. 2 96. and x is the length in feet.9 98.1 86.12.2 90.2 99.9 84.1 91.7 9 94.4 17 1 2 3 4 5 6 99. 14. Suppose that for a set of 12 observations.7 84.3 92. Fit a regression line to enable an engineer to predict the Fl number from the F2 number.11.20 23.0 95.5 26 90. A purchasing agent comparcs the prices of a certain kind of wire and fits a regression line y = 0.50 30 31 32.S.40 23. 83.1 92.4 M6 Y = 1 .50 14. Snee (1981) determined the octane ratings of 32 blends of gasoline by two standard methodsmotor (F1) and rcsearch (F2).0 88.7 89. C x = 21.7 97.7 (90.3 99. Do the calculations for the data of exercisc 14.13.5 92.
stands out as influential. 22 different asphalts were compared. Asphalt 4 is seen to be very influential. Do you get a better fit to the data? 14. Suppose that you drop it. is dclctcd? X = 22 Y = 35 1 2 21 35 13 23 31 3 25 20 14 24 31 4 23 32 15 22 36 5 29 27 16 29 6 26 7 25 28 18 24 31 X 27 31 19 24 29 9 1 0 1 1 25 21 24 30 36 39 20 27 27 21 24 27 22 34 23 12 26 33 27 26 17 29 32 14.14. docs that improve the fit? What justification would you have for dropping one. and it is rcasonable to set them aside. Drop them from thc data set and rerun the regression. X i s the percentage o f resin. Y is the percentage of asphaltene in each asphalt. 57 172 112 147 76 154 145 100 93 14. too. U ) . Does the fit improve when point 4 is dropped? When you fit the data without number 4. It is suggested that number 22 is somewhat different from the others. They have by far the highest octanes because of a high aromatic content. Is this suggestion plausible'? Is therc is :I marked change in the prediction line it number 22 is deleted? When that is done.X is the softening point in dcgrecs F.16. Fit a straight line to predict the percent asphaltene from the percent resin. Fit a straight line to predict Y from X. number 1 1 has a large residual.242 RIVARIAlE DATA: FIT17NG STRAIGHI' LINES Blends 1 and 10 stand out from the others. Y is the pcnetra tion at 115°F. Is this plausible? Is therc a marked change in fit when number 1 1 .15. too. of thosc asphalts in practice? Y= X= 77 158 106 150 149 136 119 142 111 153 156 49 194 153 145 = . number 5. Make a plot of the data and explain this. or both. the old asphalt. Prove that if W = a f hX and U corr(X? Y). then corr( W. These are measurements on I 1 samples of asphalt. In a survey of roofing asphalts. c + dY.
we express thc equation of the last line in terms o f the original variables. As in chapter fourteen.3) /2.1. the term e denotes the random error. M. we progress from the straight line. There is no requirement that the X variables be iridepcndent of each other in any broad 243 . even if no errors crept into the calculation. Inc CHAPTERFIFTEEN Multiple Regression 15. x was called independent and y dependent. x became the dependent variable and y the independent variable. This goes back to the terminology of elementary algebra. In that procedure. We use two methods. to the more complicated problem of handling several predictorsmultiple regression. Bcfore the advent of electronic computers. When one took the equation y = 2x + 3 and substitutcd a value for x to obtain y . finally. Later. the labels changed. or fitting Y to a single predictor variable. John Copyright 0 1990 by John Wiley & Sons. INTRODUCTION In this chapter.Statistical Methods in Engineering and Quality Assurance Peter W. but nowadays. At first. the sheer weight of the calculations involved prevented this technique from being used very much. We tit the model to the set o f sevcn points shown in table 15. Sometimes Y is called the dcpendcnt variable and the X variables are called the independent variables. multiple regression is one of the most commonly used statistical tools.and then fit a line to the adjusted values. We start with a manufactured example to illustratc the underlying theory.1. we make a twostage fit using straight lines. The second approach is the method o f least squares. one has to invcrt a ( p + I ) X ( p + 1 ) matrix. and Y for X. it is subject to the same conditions. we use matrices to develop the least squares method for the general case with more than two predictor variables. In order to handle a problem with p predictor variables. when the equation was transformed t o x = ( y .. we begin by adjusting X . inverting a 5 X 5 matrix on a desk calculator took several hours.2. for X . Thirty years ago.
where y” is the fitted value. Our procedure accommodates the polynomial equation Y = P” + P . 2.81Y(X. In this procedure. calculated in this example. It is left to the reader to show that we would have arrived at the same equation if we had started by “adjusting” Y and X.4. + 0.0 .2.x”. .0 = 1. which means that we cannot have x I being a linear function of x2.0 +0 . and It also handles the trigonometric model Y = Po + p1sin( wr) + p2 cos( w f ) by setting sin(wt) = XIand cos(wt) = X. X. 3 0 ~ ~ ) +0 .2J = (5.y. we chose to “adjust” Y i d XI for X. X 2 = X2. We do require that X.244 MULTIPLE REGRESSION sense.and then to fit a line to the adjusted values.0 + 0 . 3. The procedure for fitting Y to two predictor variables.and X.35~~) . we could start by adjusting for X. X. 8 1 9 ( ~. involves fitting three lines: 1.. for example.0 .and X.and then adjust for X.. u = p u . Substituting y . More is said about this later.7 for u and x I model: for u in the last line gives the final fitted }’ = y’ + 0. The three lines... 15. 3 0. for X. be lincarly indepcndent. 8 1 9 ~+ 0 . 1 0 4 ~ ~ .. in terms of X 2 with residual u = x1 . we cannot have. Y in terms of X2 with residual u = y . X + PJ2 by setting X = XI. If there were three predictor variables. instead..819~ .This is a process that won becotnes tedious. = 4.724 + 0 .1. ~ ~ u = 0.= 2X2 + 6. X. are = 5.2. THE METHOD OF ADJUSTING VARIABLES This is an oldfashioned method that has only pedagogical value in this age o f computers. . The data for an example are shown in tablc 15. 3 5 ~ ~ . .
2 + P I 2 x..9 0. PI.9 9.30 11. s. + P 2 c X.IX. = & J = Po = X./ 3 2 ~ . THE METHOD OF LEAST SQUARES We denote the estimates of &.Bo .2 3 X.4 9. and set the derivatives equal to zero. + + I40&. c c c Y.. 7&+ 708.with respect to each in turn. and we choose the estimates so as to minimize the sum of squares of the residuals: s.95 0. c + a.70 2.and &.4 10. and P2 by b. We differentiate S.95 10.x.65 13.2. We subtract 10 x ( I ) from (2) and 20 x ( I ) from (3) to eliminate obtain so.and .3 9.z + P? c 4 2 . 3. .35 0. Adjusting Variables 15 12 14 13 13 9 8 14 12 11 10 23 18 19 23 22 17 18 13.7 10. + 1412j$ .1 2.1 15. The estimated value of Y when XI = x t I and X 2 = x I 2 is so.1.9 3. c + x. 1694 = 140& + 14128.4 3.70 10.j .3.70 3.6 1. the equations are (1) 84= (2) (3) 874 = 708.6 2.lY X12Y 8" c x. . = c ( Y .70 2. respectively.2 . .95 1.1 0..s l x r l . as before. For the data in our example.6 9.30 12. + 2t)40& . 7406.05 1.THE METHOD OF LEAST SQUARES 245 Table 15.05 9 7 7 11.05 11.1 + s. This givcs us a set of three equations that are called the normal equations: p. c x2. = w+ Blxrl + P2xt2 The corresponding residual is t! = y i . 2 > ' * The summation is made over all n observations.
‘The variance.68 4 with s = 1. it would be reasonable to conclude that p2 could be zero.29 0. __I n3 10.819/0.50 . and the residuals are shown in tablc 15.75 1. = o .03 0.1 Y 1s 14 13 13 9 8 j 12 15. the new least squares equation is thc straight linc: b. = 0. changed when x2 was dropped. u L.71  d 12. Whcn this is done.5% increase over the sum of squares for the modcl with two predictor variables. and for &. is estimated by $2 .2715 = 3. 2 ti?: = 10. N . y = 3.59 13. t .__ = 2.I0. : We note that Z . The coefficicnts Po and p. The sum of squares for the residuals about this line is 11.1.S.43 12. = 0.31 10.2715)’.43 1.75 9.23 .59 .02.23 0.3. The estimated values. Also. j . We will see later that in this example.10. wliich is only a 3. In this case. The addition of s2 to the model with x.3.69 2.72 = .1.01 (it should be identically zero but for roundoff : error). Thus. which is estimated by (0.72.. and to drop nl from the prediction equation for y. .870~. for t . Table 15.97 9.64..246 MULTIPLE REGRESSION wlience so that the fitted equation is indeed the same as the equation that was obtained earlier by the first method. alone docs not improve the fit significant1y .
otherwise. . The vector o f residuals is 2 = Y .4.3) The estimation procedure calls for minimizing the residual sum o f squares. which is always unity. that they are independent normal variables with zero expectations and the same variance. for the ith data point.. We wish to obtain a vector. we write this model as: Y = X/3 + e . The matrix X is called the design matrix.Y = Y .4. the element x. . sines or cosines. x 2 .4. (15. We wish to fit the linear model with the usual assumptions about the error terms. in which there are p predictors. equating the derivatives to zero.MORE THAN 'TWO PKEDICI'ORS 247 15.. from which we obtain the estimates fi. and p + 1 columns. numbered 0 through p .xp I . As we said before. and is associated with the constant term. In matrix notation. by the method of least squares. j3 is a vector that consists of the p t 1 coefficients. It has N rows. p. in the ith row and the jth column is the valueAofX. this corresponds to a dummy variable. namely.. j3. it matters little to us whether the x variables are squares or cubes. x . This model is called a linear model hecausc it is a linear expression in the unknown coefficients. e l . and e is the corresponding vector o f errors. . It can be shown that differentiating S with respect to each of the .2) where Y is thc Ndimensional vector of the observations. lcads to the normal equations X'Y = (XfX)P . of estimates of the coefficients. Corresponding to the coefficients. (15. one for each data point. .4) ..x. X. MORE THAN I W O PREDICTORS We now turn to the general case. The iero (leftmost) column of X is a column of ones. Let %' = X/3 be the vector of estimated values of the dependent variable. the important point is that the model is a linear expression in the unknown betas. each of which contains p + I coordinates (variables). Suppose that there are N data points. and (15..4.. y . .. m2.
5) The reader should confirm equation (15.6.13. The mathematical justification for the use of the method of least squares is found in the GaussMarkov theorem. This states that of all possible linear estimates of the unknown coefficients. This implies that the X variables must be linearly independent. y 2 .3) 15.5. that are Unbiased. estimated by the unbiased estimator is s = 2 sc Np1' (15.4) for the example in section 15. . which means that they must not have a linear relationship among themselves. They are sometimes called the BLUE estimates (best linear unbiased estimates).2) This estimate has N . 15.5. .1 degrees of freedom. . y. fit. in order for its inverse to exist. Each of the estimates. .1) In particular. 'cr2 =vu2.. y . PI.5.3.5. is a linear combination of the observations. if 0) is the j t h diagonal element of V. we have (15. the variance of bI is uIIu2.4. Their covariance matrix is COV(8) = (X'X).4. (15.p . (15. For PI. . It is obviously necessary that the matrix X'X should not be singular. the least squares estimates are the (unique) estimates with the smallest variances.2 are as follows: . E( /?) = p. MORE ABOUT THE EXAMPLE 'The X'X matrix and its inverse for the example in section 15. . An example of this problem is given in section 15.248 MULTIPLE REGRESSION 6 =(x'x)'xlY. Such estimates are called linear estimates. It is used to compute tvalues for the coefficients. u2. PROPERTIES OF THE ESTIMATES The lcast squarcs estimates of the coefficients are unbiased. The error viiriancc.
0 2 7 5 ~ ~ 15.0. X1 is the depth in meters at which the sample was taken. so that.993 X4 .0121 X1 .72.1 contains such a set of data. There are 37 samples (rows) and six measurements on each (columns).1. as mentioncd earlier. which is (84. The regression equation is Y = 102 . = u22 = 280/10.dev. V( b. The remaining columns contain thc percentages of the components.1.10). The first column is the observation number and plays no further part in the discussion. 101.7.1. against five predictor variables that are in columns 2 through 6.816 X3 . .005373 tratio 41.7. 2. 280 Multiplying X’X by the vector X’Y.0.26 . The next section consists of an edited version of the printout. we get the vector of coefficients (1.~ 6 This command means regress the Y variable. as in table 15.462 0.192 = 0. 874. 0. A sample of the composition o f the earth is taken at various depths and analyzed. The data was read into the Minitab worksheet in columns 1 through 7. Much of the analysis consists of noting the percentages of various geologic components that are present in the sample.0 I 2 137 St . We wish to obtain a linear prediction equation to predict the amount of spherulites from the other variables.THE PRINTOUT and 107856 1120 1120 280 4760 84 4760 841.7. A GEOLOGICAL EXAMPLE A geologist sinks several dccp holes into the ground. which is in c7. Table 15.720 0. We also see from the matrix X’X‘ that u . 15.8.)= V ( B 2 )= 0 . 1694)’. 0.03 XS Predictor Constant XI Coef.82.31 2.58 X2 .0. THE PRINTOUT REGKESS ~7 5 ~ 2 .0275.
X4 .6 7.5 0.9 314.8 339.8 1.1 x3 0.1 0.4 2.0 0.0 0.1.8 291 .1 1.6 0.1 1.7 2.1 1.6 326.3 1.3 0.6 3.2 0.4 85.0 1.0 2.3 1.3 153.6 83.6 185.4 5.6 0.1 230.7 220.0 0.6 1.1 92.9 3.7 1.1 Y 89.1 1.8 1.0 0.2 0.0 0.96 granophyrc.6 x5 4.8 4.4 0.7 359.3 1.6 4.9 1.0 0.6 2.4 81.7 0.8 86.2 3.0 93.7 88.5 9.9 277.8 3.8 1.3 9.3 474.7 81.5 3.6 1.7 0.8 303.8 254.8 1. Y = % spherulitcs.1 89.1 0.2 1.9 83.1 83.2 9.7 0.7 1.6 XI = depth in meters.0 0.1 0. X3 = % lithic fragments.0 0.0 0.5 1.2 0.9 84.7 0.0 91.8 226. Analysis of Geological Samples x1 1 2 3 4 5 6 7 8 9 10 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 11 x2 0.4 4.8 1 .8 2.8 0.2 5.8 258.9 0.7 86.6 94.Table 15.0 2.0 0.2 90.2 7.5 0.8 0.3 483.2 86.1 0.4 319.6 1.9 0.6 89.1 84.5 272.4 0.5 10.6 10.6 94.3 0.0 90.1 0.5 1. .0 1.6 283.1 1.2 2.1 5.5 0.3 1.2 1.3 1.0 1.S 1.9 10.1 193.7 308.0 2.1 3.2 2.8 1.0 310. X2 5% phenocrysts.0 86.8 1.5 0.9 0.6 90.7 1.8 4.1 155.3 0.7 2.4 1.2 7.0 193.4 0.6 188.3 83.2 0.3 0.8 1.7 6.5 285.3 77.8 4.0 0.1 212.6 266.0 1.6 289.3 0.2 91.3 1.0 5.5 6.5 0.5 0.0 1.0 83.0 6.2 13.7.2 5.9 6.1 1.4 7.3 1.0 80.4 0.8 13.6 6.7 x4 2.6 91.0 3.8 6.3 350.8 82.6 93.0 0.7 432.9 6.4 87.1 376.9 1. X5 = 92) arnygdulcs.o 291 .0 7.0 0.4 323.2 84.9 2.1 405.7 254.2 0.8 445.6 4.6 4.9 1.4 6.4 0.0 88.8 90.3 11.1 10.9 89.
498 5.8 161 0. . In this example.THE PHINIOU'I' 25 1 x2 X3 x4 xs .when the percentages of the other components increase.has 37 . . in thc numerator .491 170.2544 0. The column headed MS contains the mean squarcs. that tells us nothing new because we have already established by the fvalues that several o f the coefficients are not zero.240 ss MS 94. and n . Analysis of Variance SOURCE Regression Error TOTAL DF 5 31 36 472.f.ON4 1.f.f. The mean square for error is s2.K 0.498 = 17.735 = H 2 . It has 5 d .1 d.2781 1. The error sum o f squares has 36 .1. .5 31 d.1 = 36 d.347 R2 = 0.735 R = 0.p .16 5. The ratio 472.240 = 0. the value of F is clearly significant.9935 . The F test for the hypothesis that 5 uses for the test statistic the ratio of the mean square for regression to the mean square for error. equal to the number of predictors.X6. f .21 7.j ) 2 . However. This is reasonable.82 3.4911643. dropping X2 from the model.1. at least for X2 . The analysis of variance table shows how the total scatter o f the observations has been divided.15 3.71 s = 2. The sum of squares for regression is obtained by subtraction.508 This is compared to t h e tabulated value o l t' with p d.0 in ribsolute value. we make a second run.578 0.86 We note that the coefficients of all the predictors are negative.1270 0. All except the coefficient of X2 have rvalucs that exceed 2. They are the sums of squares divided by their degrees of freedom.508 The total sum o f squares.f.373 0. Later. F = 94. in t h e denominator. the amount of spherulite decreases.749 643. C ( y .86 is called the coefficient of multiple correlation.
120 Residual 4.1271 0. One should be wary of discarding an observation just becausc i t has a large residual. There is a trivial increase in s and a trivial decrease in R2.959 .657 1.9866 R' = 0.OOO 85.005385 0. If we regard the standardized residual as a fvalue with 2.5. is not signilkant.dev.dev.01 X4 .8556 .087 X5 Predictor Constant x1 x3 x4 x5 Coef.0 corresponding roughly to one chance in 20. there may have been an error in the analyis.0. REGRESS ~7 4 ~2 ~ 4 .94R The residual for observation 30 is about 10% of the observed value.700 88.1.959 327 83. 2.137 0. . MULI'IPLE REGRESSION THE SECOND RUN Since the rvalue for the coefficient of X .869 St .198 8.17 3.30 5.02 643. 2.0.29K 3.300 88.Fit 1.0069 0.359 R = 0.22 178. The geologist should check this obscrvation if possible.2768 tratio 49.Resid.169 St .56 All the tvalues are now significant.85 Analysis of Variance SOURCE DF Kegression 4 Error 32 TOTAL 36 465. we should not be surprised if about 5% of t h e residuals earn R ratings.40R 2.92 3. we test and accept the hypothesis that Pr = 0. More formally. we can conclude that X .252 15.9. or even to repeat it if possible.498 483 80.2533 0.~ 6 The regression equation is Y = 100 .1. Three points have large standardized residuals: Unusual Obs.856 X3 .0.723 St.041 0. 100. There are small changes in the remaining coefficients (but not d l in the same direction) and in their tvalues.120 0.24 ss MS 116.56 s = 2. is not needed as a predictor.011673 0.0117 Xl .04 2. The presence of a large residual should prompt an engineer to question an observation and t o check it.38 7. 10 23 30 Observations XI Y Fit 154 81.
0.99 R = 0.862. The regression equation is Y = 853. we let X. but that could be a mistake.93 + 0. and we are interested in changcs in the yield of a few percent.002128 tratio 6. 853.043 0.34 s = 1. A plot of the data is shown in figure 15. 127.116 0. The X variable is the reactor temperature. X = 65. and carry out the usual multipleregression procedure with two predictor variables.2 . We substitute 255 for X in the ' fitted equation. What we have just done is theoretically correct.10. X Predictor Constant X2 Coef. = X .9 230 71.dev. The first is the size of the numbers involved. which means that we have to carry several places of decimals in our estimate of the coefficient p2.10. However.995 The significant tvalue for thc quadratic term shows that it was a good idea to fit the quadratic and not to stop with the straight line.X.270 and R 2= 0. Y is the yield of tetramer.1 16X .300 R2= 0. but there are problems. FITTING POLYNOMIALS The data shown in table 15.3 7. ' To fit a quadratic model.O1341(2X2. Suppose that we wish to predict the yield at a temperature of 255 degrees.10. yield ( Y ) 220 58.6 240 78. It is clcar that fitting a straight line explains a large fraction of the scatter in the data and one might be tempted to stop there.1 come from a chemical engineering process for polymerization. there is a tendency to overdo the polymerization and to producc pentamer instead of the desired product. tetramer.FI'I1'ING POLYNOMIALS 253 15. l'his becomes even more of Table 15. At lower temperatures.70 6.1.0 250 81. = X.4 1.0 270 85.6 260 86. increasing temperature increases the yield.025.5094x with s = 4. thc regression line is Y = 47.013482 St. Polymer Yields Temperature (X) Ave.10. When the model is fitted to the data. at higher temperatures.82 6.1.3 + 7.
5 . +3. and so we elect t to work with the more convenient numbers. we use Z . 245)/ S = . . while “independent” in the sense that neither is a Iineur function of the other. Instead of using as the predictor variable X = temp. the denominator 5 was chosen to make the values o f 2. small integers. The more serious problem lies in the fact that when the values of X are positivc integers. Regression of yield on temperature.. . 245 is the average of thc X values. thc estimates arc highly correlated and. . are highly correlated. . = (temp. .254 MULI’IPLE REGRESSION Temperature Figure 15. . a problem when the X variable is time and is recorded as X = 1982.8 / 3 . Even though nowadays we can do the inversion on PCs. i. and +S. 1984.. . + I . Thirty years ago this would have been a problem because of the inability of the small “mainframe” computers to actually invert the matrix and solve the normal equations. This causes thc matrix X‘X to be badly conditioned. In this transformation. . Then Z. Subtracting this from 2: gives 40/3. and Zf will have zero corrclation.. they will be orthogonal. X and X’. .1 . The problem can be mitigated by changing the coordinates. The average value of Z is 3513. 1983. .1. which is merely a matter of convcnicnce.e. It is more convenient computationally to modify 2:. of dubious value. 3. therefore.10.
A WARNING ABOUT EXl'RAWL.ATION
255
Z , and 2, are a pair of first and seconddegree orthogonal polynomials.
Using thc orthogonal polynomials has two consequences; the calculations are easier, and we can assess the contributions of the linear and quadratic terms independently. The tit to this quadratic model is
Y = 76.8833 + 2.54712,

0.X9813Z2
with s = 1.300 and K = 0.99 as before. ' We can substitute back for X in terms of Z , and Z 2 to obtain the earlier equation. However, if we only want to predict the yields for particular temperirturcs, we can more easily work with Z, and Z2.The predictcd yield for 235 degrees is obtained by substituting Z, =  2 and Z,= 5 to obtain 9 = 74.4.
15.1 1. A WARNING ABOUT EXTRAPOLATION
Table 15.11.1 shows t h e predicted values of the yields at three temperatures by the two models. The t w o estimates differ by 2.6 at 235 degrees, but we note that the observed yield for 230 degrees was 71.6 and for 2M degrees 78.0; both estimates fall between those two observations. A similar remark can be made about the estimates at 245 degrees. However, at 280 degrees, the estimates are far apart; they differ by 12.6. As the temperature increases. the linear model keeps climbing at a steady rate of 0.51 per degree. and will continue t o do so. The quadratic model, on the other hand, is beginning to losc stcam; t h e estimates are starting to drop. Here lies the trap of extrapolation. For some sets of data with relatively small variance, the linear and quadratic models give predictions o f Y, within the ranges of the independent variables for which data were taken, that are rather close to one another; the engineer might not get into serious difficulty by using either model. However, the engineer who extrapolates is asking for trouble. It is then that the linear and curved models diverge faster and faster. Clnfortunately, the simple firstdegree model is a vcry appealing model. Sooner or later, somebody, at a safe distance from the plant, is going
Table 15.11.1. Predicted Yields
Tcmperaturc
Linear
Quadratic
235 71.8 74.4
245 76.9
80.8
280 94.7 82.1
256
MULTIPLE REGRESSION
to notice that for the reported data, yield increases at a rate of five points per ten degrees, and will surely draw a straight line and expect you to match that line outside the range!
15.12. TESTING FOR GOODNESS OF FIT
In the previous sections, we fitted our polynomials to the means of observations at six temperatures. There were actually four observations at each temperature, and we can use this data to illustrate a method of testing the adequacy of the fitted model. The residual sum of squares is the sum of the squares of the deviations, C ( y ,  3,)’. If the model is correct, this deviation is due entirely to noise, both the noise in the observation y , itself and the combination of noises in the observations that went into computing j , . If the model is incorrect, the deviation also contains a component because E( y,) is not the same as E( j , ) . As we square and sum thc: deviations, these discrepancies from the inadequate model accumulate. The residual sum of squares thus consists of two components: noise and a lack of fit. In the usual situation. we cannot separate the two. However, thcrc are two cases in which we can examine the residual sum of squares and obtain a test for “goodness of fit,” or, more appropriatcly, lack of fit. The first case appears in the next section in an example on octane blending. With octane measurements, experience with the testing procedure has given the engineer a good estimate of u2, both from previous work and from the work of others. The engineer can compare the mean square for that is residuals, s’, to the prior value, u2.If sz is considerably larger than a’, evidence of a lack of fit in the model. In the case at hand, we can use the variation between the duplicate observations to obtain what is called an estimate of “pure error,’’ the true ‘ . value of a The actual data arc shown in table 15.12.1; some minor adjustments have been made to eliminate roundoff problems. The differences between the four observations at 220 degrees have nothing to d o with whether we chose the right regression model. They are due entirely to noise. Their sum of squares, C ( y ,  >;)*, is an estimate of
Table 15.12.1. Polymer Yields Temperature Yields
220
60.2 59.2 58.9 57.3
230
72.6 70.3 72.3 71.2
240
77.1 77.5 77.5 79.9
250
260
86.4 85.5
85.6
270 84.4 86.6 84.1 85.7
80.7 80.7 81.9 83.1
86.5
SINGULAR MATRICES
257
3a’ with 3 t1.f. Pooling this with the similar estimates from the other temperatures provides a sum of squares for “pure error” with 18 d.f. When a straight line is fitted for yield against temperature using all 24 observations, we obtain the same fitted line as before when we used the means. The sum of squares for error is 313.2 with 22 d.f.; the sum of squares for “pure error” is 21.44 with 18 d.f. The difference, 313.2  21.4 = 291.8, f measures error and a lack o fit with 22  18 = 4 d.f. We divide each sum of s uares by its d.f. If there is no lack o f fit, we should have two estimates of 9 We u. compare the ratio of these estimates to the F statistic. In the present example, wc have
F = 291.814
21.444118
= 61.2
to be compared to F*(4, 18)=2.93. In the general case, if thcre are 4 d.f. for pure error, we denote the sums of squares for error and for pure error by S, iind S,,<, respectively; the test statistic is
with n  1  p

q and y d.f.
15.13. SINGULAR MATRICES We have already mentioned that the matrix X‘X must be nonsingular (a singular matrix is one that does not have an inverse). This meam that the columns of X must be linearly independent. The following example illustrates how such a singularity can occur in practice. Suppose we are mixing three component gasolines to make blcnds, and that the volume fractions of the components in any blend are denoted by x , , x ? , and x3; the sum x , + x , + x , is identically cqual to unity. If we try to predict sonic response, such as the octane number, by the equation
the X matrix will have four columns, but its rank will be three. The sum of columns 1 through 3 is a column of unit elements, which is identical with column 0. In this case, we will not be able to find a unique solution to the normal equations. If P = (&, PI, &, &)’ is a solution, s o is the vector (p,,+ 31, p1  t , &,  t , PI  t)’ for any viilue of 1. There is clearly one variable too many. The solution to this difficulty is to drop one of the columns from the model.
258
MUI.TIPI,F. REGRESSION
IS.14. OCTANE BLENDING
The following data are taken from an octane blending study reported by K. D. Snee (1981). T h e complete study involved six components. We confine ourselves to the blends involving only three components: alkylate (ALK), light straight run (LSR), and light catcracked (LCC). The response recorded is the research octane number of the blend in the presence of an additivc. The data are given in table 15.14.1. The usual approach, recommended by Scheffk (1958) is to drop the constant term from thc model and fit thc new model:
Minitab achieves this by adding the subcomniand NOCONSTANr. The data Y, X , , X,. and X, are read into columns I , 2, 3, and 4, respectively. The variables are denoted by Y = OCT, X , = ALK, X , = LSK, and X, = IXC.
REGRESS c l 3 ~ 2  ~ 4 ; NOCONSTANT'.
The rcgression equation is
OCT = 100 ALK
Predictor Noconstant ALK LSR I,CC
s = 0.9608
t
84.1 LSK 4 101 LCC St.dev. tratio
136.15 122.33 146.95
Coef. 106.054 84.0741 100.876
0.770 0.6873 0.686
Table 15.14.1. Octane Blends
106.6 83.3 00.7
1.00
94.1 103.1 03.6 97.4 88.8 97.4
0.00 0.00 0.50 0St) 0.00 0.00 0.00 0.33
0.00 I .w 0.00 0.50 0.00 0 *so 0.25 0.75 0.33
0.00 0.ou
0.00 0.50 0.50 0.75 0.2s 0.34
1.MI
THE QlJADRATlC BLENDING MODEL
259
This formulation is thc usual linear blending model. The individual cocfficients are estimates of the octanes of the pure components. For the prediction of the octane of future blends, we can act as if the octane number of the light straight run is 84.1. Wc could, instead, have dropped one of the gasoline columns and fitted, for example,
Y = P,, + PI + P3 .
With this model, the fitted equation is
ocr = m . 9 + 5.2
ALK  16.8 LSR .
This model amounts to using thc light catcracked as a base fuel; its octane is estimated by the constant term. Thc other coefficients represent the differences between the component and LCC. Thus, the estimated octane of pure alkylate is 100.9 + 5.2 = 106.1, and of light straight run 100.9  16.8 = 84.1.
15.15. THE QUADRATIC BLENDING MODEL
The tit of the linear blending model to the data in table 15.14. I had s = 0.96. The standard deviation of an octane determination is about threctenths of a number. Clearly. there is strong evidence that the linear model is not adequate, and that there is some curvature, or syncrgism. The engineer should try a quadratic model. The complctc quadratic model. under the restriction that X x, = 1.00, is
A term such as
is redundant, because XT can be written as
its inclusion would only lead to mother singular matrix. Let us continue with the octane blending example. Thc fitted quadratic equation is
Y = 107.0 A I X + 83.2 ISR + 99.8 LCC;
with s  0.1955. This is clearly a more satisfactory lit.
 2.84X1X2 0.050X1X,3 + + 8.73X2X3
260
MULI'IPLE REGRESSION
EXERCISES
15.1. The data set used in exercise 14.4 contained :I third column. This was
a composite milcage, obtained by combining the city and highway mi lcages : comp. =
M.P.G. CITY
+ p1 (city) + &(highway)
M.P.G.
HIGHWAY
M.P.C. COMP.
53 49 46 46 46 43 40 38 38 37
58
50 50 49 44 42 39 43
50
52
55 45 47 47 47 45 42 39 38 40
Use least squarcs to calculate the coefficients pu. p,. and pz. In retrospect, the composite value for the second car is clearly wrong, probably a misprint; why'? Drop that car and recalculate the valucs.
15.2. The plot of the data in exercise 14.6 shows that the appropriate
model is ii parabola rathcr than a straight line. Fit a parabolic model
y = P" + p,x t p2x2
to thc data.
15.3. Thc valucs o f x and x z in exercise 15.2 are highly correlated. What is
their correlation? 3700 is a good approximation to parabolic modcl:
X.
Fit the
Notice that the Ivalue for the quadratic term is the same as it was in exercise 15.2, and that the fitted curve and s2 are, of course, the same. Why do the rvalues for the firstdegree terrn diffcr in exercises 15.2 and 15.3'?
15.4. An engincer wishes to predict the mean annual runoff of rainwater.
He has observations from thirteen regions for the following vari
ables:
EXERCISES
261
Y: mcan annual runoff in inches X,: mean annual preciptiation in inches X,: area of watershed (square miles)
X,:average land slope (%)
x,: stream frequency
No.
(squarc miles)'
Y
17.38 14.62 15.48 14.72 18.37 17.01 18.20 18.95 13.94 18.64 17.25 17.48 13.16
XI
44.37 44.09 41.2s 45.50 46.09 49.12 44.03 48.71 44.43 47.72 48.38 49.00 47.03
x 2
x 3
x 4
1 2 3 4 5 6 7 8 9 10 11 12 13
2.21 2.53 5.63 1.55 5.15 2.14 5.34 7.47 2.10 3.89 0.67 0.85 1.72
50 7 19 6 16 26 7 11 5 18 21 23 5
1.36 2.37 2.31 3.87 3.30 1.87 0.94 1.20 4.76 3.08 2.99 3.53 2.33
Fit thc linear model
Is it reasonable to regard point 7 as a special case? Assuming that it is, refit the model, omitting that point.
15.5. Coking coals werc investigated hy Perch and Hridgcwater (Iron and Steel Engineering, vol. 57 [ 19801, pp. 4750). They observed X,= percent fixed carbon, X , percent ash, X, percent sulfur. and Y = coking heat in BTU per pound. Fit the prediction model
The first coal seems to differ from the others, and to earn an X rating from Minitab. How does it differ? Omit the first coal from the data and rerun the regression. Do you get a better fit? In what way?
x,
I. 2. 3.
83.8 11.2 78.9 5.1 76.1 5.3
x,
x 3
Y
625 680 680 14. 15. 16.
x, x,
55.6 6.6 54.7 6.5 53.4 4.5
0.61 0.60 1.65
0.90
1.54
x,
1.10
y 715 705 730
6 9. You will notice that X.36 7.1 59.3 27. 5. 72.2 30.91 7.0 25. 12.3 56. (A continuation o f excrcise 15.9 26.74 l.8 6. until all the predictors left in the model have significant rvalues. ‘I’hcsc are measurenieiits of properties of twelve samples of asphalt.1 33.08 8. 25.1 1. What is your final prediction equation‘? 15.76 1.8 70. x. 18. 24.1 73.6 1.: flash point ( O F ) y 77 149 111 49 57 76 106 119 93 153 112 100 IS8 136 153 194 172 153 150 142 155 145 147 145 x.9 2. x.8 57. 10. is the predictor with the highest correlation. 23. 21 18 20 16 14 12 17 10 18 A ’ 4 18 14 12 35 38 34 26 27 26 38 26 32 41 29 32 28.85 0. Continue this process.6.8 55.9 61.8 26. called backwards climination.6 27.:penetration Y: pcnctration at 115°F (OF) X 2 : penetration at 32°F at 77°F X.8 1.41 6.3 1.4 56. 9.03 6.0 1.07 725 710 710 700 715 710 740 730 730 15.71 1.262 MULTIPLE REGRESSION 4.3 33. 705 685 680 700 720 70s 730 60.6 0.2 11.3 8. 20.0 27. predictors with Y. 8.8 10. 21.4 0.) Find the correlations of each of the We want to prcdict the penetration at 115°F from the other variables.93 710 710 685 17.8 61.0 70.99 6.6 1.1 5. 11.4 31.7 63. Yet the best subset of four predictors does not .6 68.6 8.l: percentage of insoluble npentane X.9 30. 22.9 61. The measurements arc X.5 6. 7.8 635 590 550 625 595 625 625 SW 5.6.: softening point X.0 73.2 7.W 0.75 0.0 59. 13.02 0.50 550 s45 595 Fit Y to all five predictors.06 1.8 61.4 60. 6. 19. Then eliminate the predictor that has the sniallcst rvalue and fit again.70 0.7.4 9.2 8.8 0.31 7.90 6.
The method used in section 15.2 and thc method of least squares give the same fitted equation for the data in table 15..76.8..5). Find R Zand sz when Y is fitted to X . (A continuation of exercise 15.9.10.EXERCISES 263 contain X.2. Drop that point from the data set and repcat the backwards elimination procedure used in exercise 15. X . alone and compare them to the eorrcsponding values for the best subset of four. you might have decided that A’.. . 15. Provc that they give the same equation for any linear model with two predictor variables. Prove that the least squares estimates are unbiased (scclion 15. 15.0 and j = 124.point 3 had a large residual: y = 111.) When you fitted Y to and X.. In the previous exercise.was of no use as a predictor. This peculiar behavior does sometimes happen. X. . I .7.6. X. 15.
the same variety might be planted in each plot. The two different catalysts compared in chapter eight can be 264 . such as different fertilizers or different methods of cultivation. and test certain hypothescs using Ftests and [tests. That procedure included obtaining an cstimatc of the variance frotn the experimental data. we extend those ideas to the situation in which we want to compare more than two populations in thc same experiment. This is done by the analysis of variance.Statistical Methods in Engineering and Quality Assurance Peter W. A. The essence of the procedure is that we take the sum of squares of the deviations of all the observations in an experiment about the overall. thcsc are the data. We then subdivide the total sum of squares into components. which correspond t o different sources of variation. INTRODUCTION In chapter eight.2. A fuller treatment of this subject can be found in John (1971). The basic idea of the chapter was that the engineer took two samples. mcan. or grand. Fisher about 60 years ago. At harvest time. THE ANALYSIS OF VARIANCE The analysis of variance was introduced by R. The analysis of variance was originally used in agronomic expcrimcnts. and then different treafinents applied.1. we considered experiments to compare two populations. 16. John Copyright 0 1990 by John Wiley & Sons. usually one variety to a plot. Inc CHAPTER SIXTEEN The Analysis of Variance 16. one from each population. The reader will notice the very strong connections between the analysis of variance and multiple regression through the underlying principle of least squares. In those plots. M. It is convenient to continue to use the word treatments in this chapter as a general expression. and so the jargon of agronomy continues in the field. different varieties of some crop are sown. Altcrnativcly. and compared the sample means by the twosample ttest. An agronomist plants crops in a field divided into small plots. this is callcd the total sum of squares. In this chapter. the yields of each plot are recorded.
are the (unknown) treatment mcans that we wish to comparc.25 L) E 43.4 41.0 43.2 45. denote the average of the observations on the ith treatment and j denote the average of all N observations. observations on thc ith with expectrcatmcnt provide a sum of squares of deviations.)2.6 42. the el.1 45.3 41.1 H E ONEWAY LAYOUT 265 called two treatments. .1 42. We can call y. observations with a total of N = I n.2) Table 16.3.1 shows a set of data for an experiment with six observations on each of five treatments.5 45.. (16. The n.8 4h.. we have v. Let 7.5 42.5 43. The mathematical model for the data is Y 11 = PI + el] . THE ONEWAY LAYOUT Oneway layout is jargon for the simplest experiment in which we compare f diffcrcnt treatments.3 43.9 32.3 41. observations.3. 16. usually denoting a “run” made in a chemical or other plant under a certain set of operating conditions.0 42.1 d. ‘The agricultural word “plot” is often replaced by the word ritn.7 45.1) The parameters F. Table 16.3 43.97 C 44. Latcr in the chapter. terms are thc random noise with the usual assumption that they are statistically independent normal variables with the same (also unknown) variance c2.3.8 42. Suppose that the engineer has samples of observations from each o f t treatments and that the sample from the ith treatment contains n .1 A 45.I)c2arid n .]. the response observed on the jth run with the ith treatment.9 46.1 43.f. (16. ted value (n.97 Averages 42. Summing these terms. I t is now common t o replace the niimc analysis of variance by the acronym ANOVA.1 44.8 44.1 41.8 45.72 . The jth observation on the ith treatment is denoted : by y. the six temperatures in the example in chapter fifteen can be regarded as six treatments.3 42.9 41.1 44.1 42. more complicated experimentstwoway layouts and threeway layoutsare introduced. C ( y.6 42.3.9 43. .60 B 42.3.
the expected valuc o f t h e numerator of F is v ’ ) . + s.3.C. = C. with n . . .1 d.S. These sums of squares are combined in the analysis of variance table The mean square for treatments is M..266 THE ANALYSIS OF VARIANCE We now write S.4) E(S. We conipute C = G 2 / N .= N j . = and that its expectation is c n. We compare the valuc 18.f. y i . Suppose that there are n. S.333/ 0. The F statistic tests the hypothesis that all the treatment means.682 = 18. = s. d. The sum of squares. It is easier to proceed as follows. _S 2  S .. Table 16.f. is called the sum of squares for trcatrnents.I . p. Lct G he the grand total of the N obscrvations: G = C 7 . N .( I*.2.1) = 49. ANOVA Table SOURCE treat. s’ = S .380 ss  _ _ l _ _ 12.08. = S .3. which is 2.76 at a ~ 0 . .333 0. with a little algebra.682 MS  l _ ~ _ _ F 18. .2. error TOTAL DF 25 29 4 4Y.3.t ) = 17. = C ( T : / n . Then wc compute the sums of squares: ‘J’he total sum of squares = S.3) The term S.. (16.1) s2 ’ has ( t . SS for crror = S = S.. ) .t d. that s.08 .f. (16.3..330 17. / ( t . has t .3.y)’. It can be shown. .  $ y + ( t . SS between treatments = S.050 66.1)cr’ . and : s. C.25) 16.t ) d. l ( t .f. / ( N .682. 0 5 and reject thc null hypothesis. The formulas that were givcn heforc are not the formulas that arc used in practice. Their ratio is the F statistic: F = 12.5) where $ is the average of the p. which is called the correction for the iiiean. .333. = 2 ( y . observations on the ith trcatment and that their total is 7.08 lo the tabulated value of F for (4. The F statistic.jy . arc equal (in which case.) = c n.( y. M.33/4 = 12. (16.05/25 =: 0.C.
.98.5.. and only if. as thcy are often called.06. With s2 = 0. .4.72 E 4?.682 and n = 6. Some software packages compute these difcrenccs automatically or on request. are equal. if where [* is the tabulitted value of t with probability a / 2 in each tail and N . such a s the example of the previous scction.f. It is called Fisher’s least significant difference (LSD).4. in increasing order. or. we declare that ph # p.4. the LSD is 0. THE ANALYSIS OF COVARIANCE We can sometimes sharpen our ability to distinguish between treatments by taking into account and adjusting for covariates. p. if.4. Sample Means B 41.25 A 45.4.1 lists the treatment averages.2) The term on the right side of equation (16.1. Applying this I S D to the avcragcs in table 16.(Y7 C 44. In the example.THE ANALYSIS OF COVARIANCE 267 Table 16. Eyuivalcntly.1. 16.1) In the particular case. we conclude that we cannot conclude that p4 # ps or that p4 # &.4. (16. equation (16. WHICH TREATMENTS DIFFER:’ We have just reached the conclusion that it is not true that all the population means.f. Other formulas for significant differences have been derived by Tukey.97 D 42. and Duncan. j .4. Newman and Keuls. wherc thc sample sizes are all equal to n.2) applies to all pairs of treatments. This does not iniply that they all differ from each other.1) reduces to (16. We would reject the hypothesis ph = p.60 16. we have for a = 5% with 25 d. I * = 2.t d. Details are given in the manuals.4. Somc statisticians argue that Fisher’s LSD is too liberal and that it declares too many differences to be significant. Table 16. .
1 52. The data are shown in table 16.7 52.17 2. unfortunately.81 SS MS F 2. Wc illustrate the analysis o f covariance with thc following example. Nevertheless.19 2.5.3 c x 3.57 3. We add a term to our earlier model in cquation (163.268 Table 16.42 3. In animal nutrition experiments where some young animals are fed ration A for a few months and others are fed ration B.5.6 51. The plot o f Y versus X i s a straight line (except for random noise) for each treatment and.5 53.7 51.4 50.45 3. The adjustment for the covariate is linear.38 3. is the value of the covariate for the ijth observation. which ignores the levels of the impurity.22 Y Sl.16 53.1 2. and that. Example 26.2.1.1.9 52. I) where x. Three reagents for a chemical process were compared in a pilot plant.48 83. the yields with treatment A are lower than the yields with C. Eight runs were made with each reagent (treatment). 9 (16.1 53. by and large.2 49.4 51.40 Y 54. the oneway ANOVA in table 16. Yields nnd Impurities THE ANALYSIS OF VARIANCE Y A 54.7 51. one might take the initial weight of each animal as a covariate.43 2.12 .41 2.1..2.6 54.5. contained an impurity. Figure 16.1 47.76 51. and the weight gains are compared.31 2. I is a plot o f Y versus X.48 3. The model implies two things.1 51. It is clear that the yield decreases as thc amount of impurity increases.5. This could be called a firstorder adjustmcnt.48 3. A common situation occurs when each run is made on a separate batch of raw material and we may want to “adjust” Y for the amount of some impurity in the batch. furthermore.1) and write ))I] = PI + P X l l + el.72 3.3 50.78 2. Y denotes the yield and X the percentage of the impurity in the raw material for each run.23 3.79 3.6 54.86 3. Oneway ANOVA Ignoring Covariatc SOURCE Treatments Error TOTAL DF 2 21 23 16.5.16 concomitant variables.73 2.29 2. The raw material.9’) 2.3 B X 2. the slope of the line is the same ( p ) for every treatment.6 56. docs not Table 16.5.Y6 8.9 2.5.0 55.9 X 2.99 100.9 50.17 3.
If. Table 16.437 .6512 = 10.400 . = 14. C = 53.0 Y 52.THE ANALYSIS OF COVARIANCE 269 56. c .43 and 3. however. In the next section. and s2 = 0. 108.32 13. The two smallest yields for C are 51.3 are obtained.77 10.3.6512.5.5.02 100. B = 52. but those valucs of Y occur with relatively high values of X.32. S.8331 0. I I L 1 give a significant value of F. : After adjusting for the covariate. which is much lower.5.16. we adjust for the amount of impurity o n each run by the analysis of covariance. 3.6512 4.32/2)/0. we show how the figures in table 16.3 shows the printout from an analysis of covariance program. Analysis of Covariance SOURCE Covariate ‘I’reatmcnts Error TOTAL DF 1 2 20 23 ADJ S S 70.83 14. Table 16. wc can establish that there are differences between treatments. which is significant.B A A C B C B  B A A C C A C C BC B A6 A B 49.81 MS 70.5 .0  A 1 .3832 7.99. and we are able to conclude that the three catalysts are not all equal.162 . a value that is exceeded by three runs with A and by five runs with B. and wc cannot concludc on thc basis of that analysis that the trcatments differ.. which is somewhat lower than heforc.1585.6. The F ratio increased to (14.99 F . ‘The treatment averages were A = 5 1.namely.
10. y = h. X3 = 1 for C . L e t X I denote the percentage of the impurity present in a run.6.8799 St. The following model fits thrcc parallel lines to the data.34. = p3 .7445 1.263 0.8070 I<’ = 0. = p1.86. x . table 163. X. ‘Thc reduction in S.92 .2) Thc common slope of the three lines is b.6. DF SOU KCE Regression 3 Error 20 TOTAL 23 87. (16. h. + b .94 F P 0.88 X3. wc refer back to the onewzry analysis of variance.dev.13. THE ANALYSIS OF COVARIANCE AND REGRESSION The rationale for the analysis of covariance table 165.43 1.651 44.3. We first fit a single straight line through all 24 observations with the model y . + b. sce the accompanying table 16.3454 0. one for each treatment.66 s = 0. on the other hand.3.744 X2 + 1. Table 16. = . = p .871 Analysis of Variance.3 can be seen if we approach the problem from the point of view of regression. .02. the intercept for the line for the first treatment.024 100. h. .789 13. 61 919 3. + h 2 X 2 + h3X3+ e .048 0.5. This is the (adjusted) bum of squares for treatments in table 16. (16. If.x. that occurs when we include the covariate is The equation is Y = 61.p . With this model.1. XI =Impurity.813 ss MS 29.6. the sum of squares for error when we ignored the covariate was 83. Rwression x1 Predictor Constant x2 x3 Coef.270 THE ANALYSIS 0 VARIANCE 1 16.000 .4038 tratio 59. 1. and h.342.85 4. S. = 1 if the observation is made with catalyst B and zero othcrwise.60 X I + 0.b. X2 = 1 for B.07 .4035 0. = 1 if the observation is made with catalyst C and zero otherwise.6. The reduction in S.6.6.1. drops to 13. We next introduce two indicator variables: X .2).p l .achieved by considering three separate parallel lines rather than just one line is 27. equation (16.6023 0.32.02= 14..2.I) The sum of squares for error about this line is 27.
but that C gives significantly higher yields than A . The analysis of variance tablc for all 24 observations is shown in table 16. from section 15.)/7(J = 2. This quantity (allowing for some roundoff error) appears in table 16.6 84. and Chcvrolet.3 240 77.801.02 = 70. Wc recall.5 85. make of car with levels Ford.5.4 85.3 72. = 4(quaJratic contrast)' .2.9 57. We conclude that catalysts A and H d o not differ significantly. we introduced an exiimple in which a quadratic polynomial was fitted to the yields of a polymer plant. we cannot conclude that 6 .12. we can take coniponents with single dcgrccs o f freedom for the linear and quadratic trends.10.62 . that has six levels. Since 1.00.QUAN'FITA I'IVB FACTORS 27 1 83.pl f 0 . 16.. with C r: = 70.1 for convenience.5 79.1.5 77.1 77. temperature.1 2(A1 270 84. or that cc.9 250 80. The term quantitative is used for a factor whose levels are expressed as quantities.7.84. Y = 75.6 86. Table 16. The corresponding sum of squares is s.13.2 71.6 70. which is repeated as table 16. the linear polynomial.7.Y 83. We notice that the rvalues for the two indicator variables arc 1.7.3 230 72. We can also establish by similar calculations that C' givcs significantly higher yields than B. that the valucs taken by Z. 3.7. 1. A similar argument defines the quadratic contrast as 2: ( z l .3 iis the (adjusted) sum of squares for the covariiite. c 2: the factor 4 in the numerator is thc number of observations that were made at each level of temperature.7 60.2 86.5 and ) the corresponding sum of squares as s . as opposed to a qualitative factor whose levels are not numerical. The coefficient in the regression equation was C (z1.M = 271. = 4(1inear contrast)* = 1816.12. 1._ 22. 3.44 84 _ _ I z.8s < 2.1 85. QUANTITATIVE FACTORS In section IS. We look itt it again as an examplc of oneway analysis of variance with a factor.2 58.7 81. f 0. Polymer Yields I I Tcmpernturc Yields 220 3.g.7 80. The data were shown in table 15.4 86. and 5.5 . Plymouth.5471. were 5. e.86 .66 for C'. Out of the sum o f squares between temperatures.? .j.85 for B and 4. I ..
76/1.3. 18) = 6. A = C sly. Chevrolet.44 2129. Mercedes.76 1. Single degrees of freedom for orthogonal contrasts can also be isolated with qualitative factors. The contrast compares the American cars as a group t o the forcign cars.y h s compares the luxury cars to the ordinary cars. and R = C h . are orthogonal if C a. .9 . . F(3. y . Temp.44 20.272 Table 16.44 6. .2.b.7.62 271.44 2129. Quad.62 271. Supposc that we have as the levels of a qualitative factor six makes of car: Cadillac. = 0.68. Plymouth. Residual TOTAL DF 1 3 18 23 1 1816.79 ss MS 1816.3.79 ss MS F 421. Two contrasts. Analvsis of Variance SOURCE Lin.35 21.19 We can now amend the analysis of variance table in table 16. Cubic etc. Its sum of squares is The orthogonal contrast B = 2y. Analysis of Variance THE ANALYSIS OF VARIANCE SOURCE Temperature Residual TOTAL DF 5 23 18 2108.y2 .7. with n observations on each make.7.19 F 5.68 .19 = 5. The F statistic. indicates that the model with only the quadratic trend is not entirely adequate and that there is still some unexplained curvature. Its sum of squares is s z  nB2 12 = nB2 c bf' Table 16.7 3 + 2y. Temp.67 354.29 21. The reader will notice that this is the same as the pure error test for goodness of fit that was used in the previous chapter to test the adequacy of the simpler firstdegree model.01 1. and Toyota. Mazda.
06 43.27 11.32 47. The fact that in this cxample the two factors each have the same number of levels is irrelevant.49.8.55 9.76 9.97 174.88 3 12.8.11 49. Example 16.f.8.8. temperature arid pressure.99 40.Y.54 12. If we were to make a oneway analysis for the temperatures.95 10. F * ( 3 . The engineer can arrange the data in a rectangular array. ordinary for American cars) 2j4 v5 . at four levels each.1.59 9. THE TWOWAY LAYOUT Suppose that a chemical engineer wants to investigate two factors.17 13.13 2 3 4 1 Sum 8.THE TWOWAY LAYOUl 273 Indeed. Similarly.28 44. ordinary for imported cars) The two new contrasts arc orthogonal.j 3 and (luxury vs. A .60 4 Sum 40. A Twoway Layout Temperalure Pressure 1 2 9.52 with 3 d.1. The analysis is illustrated in the following example.08 11. The data set for an experiment like that which has just been described is shown in table 16. . The corresponding Fstatistic is 2. with the columns corrcsponding to the four levels o f temperature and the rows to the levels of pressure.03 11. We ciin analyze the data as if there were two oneway layoutsone for rows and one for columns. . both to onc another and to the first contrast. B can bc divided into two contrasts: 2 j . I . We could just as well have five temperatures and six pressures.96 11.10 8. we would get S. 1 2 ) = 3.89 11.V.f.= 8. and S.84 F(3.12)= = 1 . 7 0 .1.. ignoring the fact that the ohscrvations were made at different pressures. and makes one run with each combination of temperature and prcssuret6 runs in all. This is called a twoway layout.CN . (luxury vs.44 12. = 20.67 which is not significant. a oneway analysis for Table 16.47 41.48 40. 16.06 with 12 d.
8.1 d. In the general case. ORTHOGONALITY 0 The reason that we were able to take out the sums o f squares for each of the two factors i n the prior example is that the experiment was balanced. for the sum of squares for A . . and similar contrasts for the pressures.580 MS 2. the two factors are said to be orthogonal to one another.200 the pressures. quadratic temperatures.I d. . for H. The F statistics arc compared to the tabulated valuc F * ( 3 .” and vice versa.67 when we looked at temperature alone. also does not give a significant F value. contrasts for linear temperatures. there will be a . To compute the sums of squares for A and for B.4 to see which temperatures diffcr. 9 ) = 3. and ( a . or we can.936 28. table 16. we let 7. ignoring tcmpcratures. Then the sums of squares for A and B are The sum o f squarcs for error is obtained by subtraction as s. wc ciin take out of the total sum of squares both the sum of squares for tcmpcraturcs and t h e sum of squares for pressures and obtain the following ANOVA table.. Twoway ANOVA ‘THE ANALYSIS 0 :VARIANCE 1 SOURCE temp. denote the sum of the observations at the jth level of B. Notice that the value of s2 has droppcd to 0. Howcvcr. denote the sum of thc observations at thc ith level o f A and T. This balance has the technical name o f orthogonality. and cubic temperature. if appropriate.180 9.39 5.f.8.518 15. and we now conclude that there are differences both between temperatures and between pressures. = s. for the sum o f squares for error.s.f.548 from its earlier value of 1.s.l)(6 .8.f.042 0. We can go ahcad and apply the methods of section 16.274 Table 16.E. 16. respectively..548 F 5.9. A and H . whcre thcre are two factors. prcss. h .86. Each temperature a p p ~ i r c d exactly once with each pressure.2.1) d.126 4. apply the methods of section 16. with a and b levels.2. the pressures “canccl out. and s o when two temperatures are comparcd. error TOTAL I>F 3 3 9 15 SS 8. Orthogonality is a very important property in .7 and take out singlc d.
. For F . there is a total of nine cells that can be filled freely.0. average. .061 0. The complete set of residuals is shown in table 16. but the fourth is fixcd because the column total must be zero. The sum of the squared residuals is indeed 4.055 0.9. given the first three entries in each row.1.883.520.868 . . and c. We can now see why the sun1 of squares for error has nine d. = hi iF. or grand. the first three entries are random.625 0.1.95 . and the residual in that cell is 8.570 .570.f.1) whcrc m represents an overall mean. Wc can go further and calculate an estimated value for the observation in row i and column j as j I . + t. The sum of the rr and the sum of the c. Thus. (16. we subtract the grand average from the avcrage of the observations in the first row: similarly. In each of the first three columns.  so that j . In our examplc.277 0. . 0.520. the entries in the other sewn cells are forced by the conditions that the rows and t h e columns sum to zero.13/ 16 = 10. .883 .ORTHOGONALITY 275 Table 16.257 0.082 0.484 an experimental design.9. and we will see more o f i t in the ncxt thrce chapters. the fourth is fixed. is the extra contribution for being i n the ith row. The sums of the residuals in each row and in each column are identically zero. are both identically zero.325 0. m is estimated by the overall.040 0. tel. Similarly. = 10. 174. Residurls 0. r . is the extra contribution for being in the jth column.2.587 0. as was shown in table 16.9.160 0.0.485 0.495 = 9.1.9. . = rn + r r + c.048 0. We can use the following model for the observation in the ith row and the jth column: y.701 1.m3 0.8.936.
'I'hc technical term used for the strips is blocks. he would call it a split lot.11. and so on.0 47. Filling Weights of Cans 68.10. We can return to our example and rcplace the word pressure by the word block and go through exactly the same calculations. Instead of taking 2n units o f raw material and assigning n to each treatment. we noted that the precision of the experiment can bc improved by pairing. might take n batches of raw material. Our engineer. who wants to comparc 1 treatments.0 51.0 49. and the batches are the blocks. I I for variety 8 .0 52. Then the agronomist would take IZ plots and sow variety A in thcm.0 57.0 50. of plots available. we add to the mean first a row term.11. 16.0 36.0 65.0 34.0 45.1.0 61. and. and the cxpcriment is called a randomized complete block cxgeriment. The name carries over to enginecring applications.0 87. we should.0 56. RANDOMIZED COMPLETE BLOCKS In our discussion in chapter eight of the twosample ttest.276 THE ANALYSIS OF VARIANCE 16.0 48.0 32. Table 16.0 70.0 75. One experimental design would be to choose the n plots for A at random from the total number.0 50.0 52. vrt. we trciiL the blocks as if they were the levels of a second factor.0 88.0 70. and assign one of the portions t o each treatment. divide each batch into t equal portions. it is important that each treatment appear the sume number of times in each block in order t o achieve orthogonalityto make treatments orthogonal to blocks. if possible. Again.0 52. take n units. similarly for each of the other varieties. In the analysis. But often there is considerable variation in fertility over the whole field. split each unit in half.0 40.0 45. An agronomist w h o is carrying out an experiment to comparc u varieties of barley would divide a field into plots. A semiconductor manufacturer could takc a boat that holds 32 wafers in an oven and load it with 16 wafers that have received one treatment and 16 that have rcceived another. INTERACTION The model. and run onehalf with one treatment and the second half with the other.0 52.0 55. equation (16.9. then a column term.0 . that we have used so far is a simple additive model.0 73.0 65. The portions of raw material play the role of the agronomist's plots. and so the ficld might be divided into strips of u plots each that are fairly homogeneous and each variety plantcd in one plot chosen at random from each strip. I ) .0 84.
should not show any pattern.11.13 8. The analysis of variance table is table 16.12 MS F 5.13 8.10 6. which is discussed in the next scction.93 7. because its column contains the four largest residuals.88 tinally.1. ANOVA Table SOURCE times heads error TOTAL DF 4 5 20 29 SS 2601.20 7. we have a test for the existence of interaction.3.70 4.20 23.67 5. When there is only one observation in each cell.37 2602.3. Snee (1973) reported an investigation of the heads of a multihcad machine for tilling cans of a soft drink. The data shown in table 16.50 2.11.27 0.47 6428. It can be argued. There is no significant difference between heads.11.47 8. They took samples of cans filled by six heads of the machine every hour for five hours.70 4.1.28 245.1 are the coded weights of the amount of drink that the head put into the can. pattern is called interaction. with the benefit of hindsight.90 7. This implies that the improvement in yield by increasing the temperature from level 1 to level 2 is the same. we see that head 5 exhibits some very erratic behavior.13 26.2. But when we look at the residuals. no matter what the pressure. Perhaps so. Ott and R. The residuals in the table.90 4.30 7.11.07 9. or nonadditivity.21.20 0. The sum of squares for error measures both interaction and error.2. Example 26. D. and is correspondingly inflated. It is also rather restrictive because it implies that the amount added to each observation because it is in the first column is the same in every row. that wc could have spotted this peculiar behavior by looking carefully at table 16. Table 16.50 17.30 0.13 1225. Residuals 4. The fiiilurc of thc data to follow this simple.07 130.00 1. an error term. like the residuals in a regression.47 2.60 6. however. We can sometimes. but rigid.11.97 650. When there are several observations in each cell of our array.11. R.00 1.70 0 . but it is easier to spot this kind of anomaly from the table of residuals.INTERACTION 277 Table 16.33 4.10 0. spot interaction by examining the table of residuals.47 15.27 1. E.40 6.30 5. we have no such test. table 16.
12. Then the sums of squares are the following: S.00 70. i.12.e.2. in the ijth cell. SOURCE A B Error CrorAL AxB = S? . ..25 ss MS 21.1.12.75 42. The analysis of variance table is table 16. we can split the residual sum of squares into two components. 2.12.12.00 58. The data are shown in table 16. INTERACTION WITH SEVERAL OBSERVATIONS IN EACH CELL When there are several observations per cell. I . ' c br ' T' s.58 . This is done in the same way as in regression. we let T. Table 16. totals 80 82 86 84 81 78 I1 111 1V 85 82 83 Row sum 679 656 657 80 88 85 90 88 87 89 78 86 81 477 491 503 521 lYY2 16.278 Tablc 16.c'.s . Data THE ANALYSIS OF VARIANCE I 1.12. The sum o f squarcs for error is then obtained by subtraction as s. and two observations in each cell.13 F 3.29 330. wherc we divided the residual sum o f squares into two portions: pure error and goodness of fit.00 11. = 2 (2). .. = Tf (1) ar . dcnote the sum o f the observations at the i t h lcvcl of A and the jth level of B . A (rows) with three levels and B (columns) with four levels.00 43. Col.s.79 3.2. Example 16.1. Analysis of Variance 2 3 6 12 23 df 174. 83 81 7x 77 80 3. 'To compute the sums of squares for A and for B. Considcr an experiment with two factors.sa. one duc to interaction and the other to error.
.f. A .5 (row 3. together with (a . for A . the mean in the first row is at least three points higher than the mean in the last row.. we have a third subscript on the totals. column 4) should be looked at carefully.5 81 . there will he rrhr . h . table 16.f. and C at a. B .1) d.. perhaps. as hcfore.f.I for H .5 The value of the Fratio. we look at the cell means.29. 0 16.. for the A 8 interaction and the remaining ah(r . Cell Means 82 77.5 70 83. the last cell is appreciably higher than the others.. is significant. T. One suspects that the observations 86. which arc now denoted by T. 'The observation at the ith level of A .5 89 88 83.1. In the general case of two factors. and r observations in each cell. and the .13.3.5 85 81 79. and there is clear evidence of interaction. This will be divided by the ANOVA calculations into u .. h . and c levels each. kth level of C is denoted by Y .00. 83.. k . respectively. the analysis proceeds in the same way as for two factors except for some extra terms in the table.f. and c' are computed as before. A and B. and have a . thejth level of R . with a and 0 levels.5 86.279 Table 16.l ) ( b . 3.5 (row 3. and T . B. 12) = 3. and c . P ( 6 .1 d. We note that in every column save the third. for thc total sum of squares. for the sum of squares for error. 6. ~ The sums of squares for A .1) d.3. To see this more clearly. .1 d.f.12.1 .12. column 3) and. the interaction sums of squares are given by . THREE FACTORS With three factors. In the lirst two rows.Then Similarly.1 d.
with rn > 1.l)(c . these calculations were tedious.1) d. . A research engineer for an oil company wants to compare four gasolines. The residual sum of squares is then with abc(rn . pausing after each to decide wherc to go next. EXERCISE!! 16.1) d. In the formulas for the sums of squares. thc agronomist has to plant all the plots together in the spring and the plants grow side by side. The engineer makes runs. SaAcr which will have ( a l ) ( b .1. terms such as Tf. The large factorial experiment with several factors at four. Anything omitted has to wait until next year. SEVERAL FACTORS The methods presented in this chapter for the two.f. The cost per plot in the agricultural experiment is usually much cheaper than the cost per run for an engineer..14. 16. five. Years ago. The data are given in the following. then the agronomist has to wait until the end of the season t o harvcst the data. Do the gasolines differ significantly in their octane numbers? Can you recognize any of the gasolines as being different from the others? . the remainder is the sum of squares for error. but they point out a fundamental difference in experimental stratcgies in the two fields. The engincer makes five determinations of the octane number of each gasoline.lbr become Tf. After all these six terms are subtracted from the total sum of squares. they are easily done for us on a PC. and is therefore more able to conduct a series of small experiments. There are two reasons for this. Nowadays. wc turn to a group of factorial cxperimerits that are commonly used in industryexperiments with several factors at only two or three levelsand develop some of the strategy of scqucntial cxpcrimenta tion. Furthermore. generalities. . ‘These remarks are./brm.280 THE ANALYSIS OF VARIANCE If there are rn observations per cell. in sequence. In the next chapter. or six levels is more appropriate to agriculture than to engineering. of course. we can also take out a component for the threefactor interaction.f. one after another.and thrceway layout can be cxtended to several factors.
7 80.6 86.EXERCISES 28 1 81.5 88.5 71.0 V 87.3 80.6 86. the engineer wants to cornpiire live gasolines. The percentages of product from three sources of raw material follow.5 72.6 80.7 89.6 86.2 86.6 A B 80.9 I11 88.2. A 73 75 78 73 71 B 75 79 78 80 81 77 C 78 74 79 71 76 78 74 16. Do the sources differ in the amount of product? A= H= C 72.4.4 80.9 72.8 72.0 78.0 70. Unfortunately.2 .5 85. The figures shown are the percentages of raw material convcrtcd to product.9 79. the engineer was not ablc to makc the samc number of plant runs with cach of thc acids.1 86.7 80.9 86.8 86.6 87.6 87.1 72.8 80.0 79.5 71.2 85.3 $6. An engineer compared acid catalysts from three suppliers in a chcmi cal process at the pilot plant stage.0 71. 'I'here were seven samples from the first source and four from each of thc other two.9 86.1 IV $6.3 C 79.5 72.1 85.7 71.0 86.2 80.1 86. Is there a significant difference between cars? Which gasolines differ? Cars Gasolines A B c' L) E 88.7 71.6 80.2 69.4 80.6 16.4 80.7 1 I1 86.8 87. In another octane study.0 87.4 16.2 D 80.3.4 73.5 81.0 72.1 80.3 81. 'The cnginecr takes five cars arid makes one determination of the octane number (on the road) for each gasoline in each car. Carry out the analysis of variance calculations and find which acid differs from the others (if any).
134.6. B. They are tested at four operating tempcratures. Is there cause from this data to think that there are significant differences between positions on the wafer? Docs the thickness at the center differ from the average thickness on the edge? 16. 120 135.7 21. The response is a measure of the tensile strength of the material and a higher value is desirable. Positions (columns) 1. Two subcontractors are invited to make five samples of the rnatcrial prepared by the old method and five samples prepared by the new method.3 D 23. 122 128.6.2 24. Do the wafers differ significantly in thickness? (Wafers are rows. 126. 3. 129.5 22.5 I1 IV C 21.4 130. The components come from two different manufacturers. 2. Temperature Machine 1 111 A 21. In exercise 16. the rows correspond to four machines and the columns to live lcvcls of operating temperature. 124. and 4 arc 90 degrees apart o n the edge of the wafer. The measurements in column 5 are made at the center of the wafer. Five measurements were made on each wafcr.2 25.9 23. 133. 128 16.9 B 21. 137.3 24. 136.7.2 21. fabricated under different conditions.7 21. Three components are tested under each set of conditions. In a twoway layout. the columns represent different positions on the wafcr.9.7 16. Wc may assume that the . 127.6 23.8 22. 127.6 24.) 1736 1665 1623 1696 1711 1725 1657 1647 1725 1712 1726 1674 1685 1735 1742 1730 16Y3 1721 1735 1743 1659 1637 1738 1735 1701 16. 129 Supplier B Old New 125. Is there a difference between the methods? Is there a significant difference between suppliers'? Is there significant cvidcnce o f interaction? Method Supplier . wcrc compared for thickness. 127.7 26. A company wishes to consider changing the way in which one of the materials that it uses in its fabrication process is macic.5.282 THE ANALYSIS OF VARIANCE 16. Thc following data are the lifetimes of components.5 E 23. A .8.1 24. 128. 125. Five wafers. C.6 22. and two power ratings. Analyze the data.0 22.
c. Manufacturer A . 1848 2092 2328 2416 2664 1960 2896 3200 c. 3024 3360 3108 266U 2908 2212 3368 3148 4 n. Which are the important factors? Makc a twoway table of means to explain the H * C interaction. 1796 2088 2780 2528 4068 2188 2520 4176 c 2 3448 2548 3200 3236 3472 2392 2764 3064 B. 4 B 4 1252 2952 2802 2184 2544 2288 3076 2828 c. B2 Manufacturer A .EXERCISES 283 lifetimes are normally distributed. '14 2088 2452 2824 3172 2316 2124 1836 2412 . H.
i n chapter eighteen. We turn now morc specifically to applications o f experimental design to engineering problems. The factorial experiments are designed experiments. 20 observations. These experiments play a very important part in modcrn engineering experimental design. In this chapter. using factorial experiments or fractional factorials. each of which has exactly two Icvcls. S. The reader who wishes to go further will find fuller discussions in the books Stutistics for Experimenters by G . E. Hunter. K. and in chapter nineteen. Now they are widely used in 284 . His first example dealt with using fcrtilizers to help in growing potatoes. 1NTROL)UC'TION Chapter sixteen was a gcncrel survey of thc analysis of variance. Yiites (1937). John Copyright 0 1990 by John Wiley & Sons. Inc CHAPTER SEVENTEEN Design of Experiments: Factorial Experiments at Two Levels 17. They were first discussed in a systematic way by F. phosphorus. which wcrc applied at two levels: zero and four pounds per plot. M. one at each of five levels of A and itt each of four levels of H . W. His three factors were nitrogen. P. G. P. who was using them for experiments at the Rothamsted Agricultural Station near London.. especially in manufacturing engineering. Taguchi. M. applications to chemical enginccring were developed by George Box and others at Imperial Chemical Industries (ICI). and W. and Statistical Design und Analysis o f Experirizents by P. and potassium. factors with more than two levels. Hunter.Statistical Methods in Engineering and Quality Assurance Peter W. It ended with a discussion of factorial experiments. We only have space for an introduction to factorial experiments and their fractions. for example.1. Box. During and after World War IJ. J. is called a 2" factorial experiment. John and in articles in such journals as Technonzelric:s and the Journul of Quality Tec*hnology. The engineer carefully decides beforehand to make. That design gives a good experiment with the important property of orthogoniility. applications that have come to be associated with the Japanese engineer and statistician G . An expcriment with n factors. we consider factors at two levels. N .
1971). With four factors. and flow rate at 8 or 12 gallons per minute. the factors are independent of each other. Taguchi and his associates. for example.1 0 x3= . which increases the precision of the estimates. and with six factors. with 32 points.EXPERIMENTING ON THE VERTICES OF A CUBE 285 chemical engineering and in the semiconductor industry throughout the world. or balance. ' We discuss a onesixteenth replicate of a 2' factorial. with five factors. or a quarter replicate with only 16 points. This means that in the analysis. which were introduced by David Finney (1945). A chemical engineering example of a 2' factorial would be an experiment in a pilot plant with three factors: temperature at 200 or 300 dcgrees. pressure at SO0 or 550 pounds per square inch. or half replicate. the book by John. but they are mathematically equivalent. This reduction i n the number of points does not come without cost. In Japan.250 50 pres. . we can use averages. but they are bcyond the scope of this book. which we denote by 2 7 .12 that when we use fractions. In the chemical experiment. so that when we are estimating their effects. 32 runs. we can let the three factors be A (tempcraturc): B (pressure): C (flow rate): temp. they have been adapted successfully for many years by G . 17. we can run a wellplanned half fraction.4or 2'118.525 x2 = 25 ' rate . somctimes denoted by 2 I. factorial experiments have the important property of orthogonality. Each level of factor A appears thc same number of times with each level of factor U . 64 runs. it is necessary to assume that some of the interactions are zero. and it also makes the analysis much simpler. As we noted in the previous chapter. That many runs may tax the budget o f time or money too highly. the latter standing for seven factors at two levels each in eight points. His ideas arc now being applied extensively in the West. This important property holds for the 2" factorial and will also hold for the fractions that we will use. There are nonorthogonal fractions that are useful (sec. Orthogonality makes it easier for us to understand what is going on in the experiment. Each can bc reduced to experimenting on the vcrticcs o f a cube with coordinates k l . We will see in section 17. . 2 x. EXPERIMENTING ON THE VERTICES OF A CUBE The two experiments that were mentioned in the previous section come from vastly different fields.. so we turn to fractional factorials.2. = 5 . Instead of running all 64points in the factorial. the complete factorial calls for 2 X 2 % 2 x 2 = 16 runs.
With only two factors. respectively. a. Taguchi denotes the levels by 1 and 2. Many agronomists usc levels 0 and 1 rather than k l .= 52 * ub = 54 T h e (main) cffcct o f the factor A is thc averagc of the runs at high A minus the average o f thc runs at low A : . = (amount . and ahc. acd. h. ac. bcd. FXPERIMENTS AT TWO LEVELS In t h e potato experiment. and also the observation (or average of the observations) made under those conditions. and ah by c . In the threefactor experimcnt. The letter b denotes both thc point with high B and low A and low C. a . we add eight more points: d . The experiment has eight points: ( I ) . milltiplying by b to include the second factor. The point with A and bl at high levels and C at its low level is dcnotcd by ab. and uh = 54. and C correspond to N . we can let A . bd. In any case. 1’0 add C’. The point with A at its high level and the other two factors at their low levels is dcnotcd by u. V. the one point with all three factors at their low levels is dcnotcd by ( I ) . we are. The extension to more than three factors is obvious. AN EXAMPLE OF A 2’ EXPERIMENT Suppose ( I ) = 36.2)/2. we can arbitrarily call one of them the low level of the factor “ciitalyst” and the other thc high level. cd. 17. c . we can talk about the high and low levels of the factors. If a factor is qualitative. then. getting c. In a 2‘ factorial. h = 52. The point with all three factors at their high levels is abc. with x . U . we are experimenting at the vertices of an ndimensional hypercube. abd. h.3. we multiply ( l ) . ac. and K. as we will see shortly.286 DliSlGN OF EXPERIMENTS: FACTORIAL. The notation servcs double duty. ad. such as a choice hetwccn sodium hydroxide and potassium hydroxide as a catalyst. h. hc. ah. This formulation of the experiment was introduced by the scientists at ICI. and abc. a = 48. Yates gave a useful notation for denoting the various treiitincnt combinations. soinctimes called a treatment combination. experimenting at thc vertices of a unit square. With more than three factors. and abed. The points are usually written in standard order: first (1) and a . Each vertex of the cube reprcsents a set o f operating conditions. We can present thc data at the vertices of a square as follows: h . bc. ah.
. The interaction can be displayed graphically in an intcraction plot.(6 . and ah.( I ) = 48 . the interaction is (2 .0 = 7 . u. the two lines on the plot would be parallel.+ 5 4 2 36+4X . the effect of A is a6 . We usually write + and .= 2 36+52 2 51.52+ 36= 10. I'he difference is 2 .12)/? = . each of these effects has 1 d.36 = 12.f. Wc could just as well have compared the effect of changing the level of S at high A and low A . The All intcraction is u measure of the failure of the simple additive model. 0 .1.THE DESIGN M A l K I X 287 48+54 A=.0.1. X.0 .6 = 54 .5 . THE DESIGN MATRIX The design matrix.48 .44.. is an array that shows the lcvels o f the factors at the points of the design.instead o f + 1 iind 1. I'he main effect of B is similarly defined: 52 B=. Interaction plot 17. at low B. In this example. and obtained ( ~ bU ) . X has two columns corresponding to A and R and four rows corresponding to ( I ) .52 = 2.10.11. 2 Since there are two levels for each factor. .3. 6. Is the A effect the same at low H ias it is at high B'! Do we get the same improvement in yield with a ten degree iricrcase in temperature at high pressure as we do at low pressure'! At high B.3.4. figure 17.( l ) ]= 54 .12 . If there were no interaction. the effect o f A is N . 48 36 High b Low b Figure 17.
5. thc estimates of the effects are obtained by dividing the appropriate contrast by N/2 or 2"'. An important consequence of the orthogonality of thc design is that we are able to increase the precision of the estimates by using averagcs in this way.x* = 1) 9 so that AB corresponds to the product xIx2. which can be represented as the vertices of a cube. l h e r e are seven effects .288 DESIGN OF EXPERIMENTS: FAmORIAL EXPERIMENTS AT TWO LEVELS A x=  B + + . and AB contrasts. + 1 ) exactly one time. a b . + + + y= (1) U h ab The estimated main effects can be obtained by multiplying the vector Y/2 on the left by the transpose of X: A= c (XI= + 1 )  c (XI = 1) 2 and We note also that AB = uh + (1) . 17.. hc. B. 1). and ( + 1 . In the general 2" design. ac. ( + I . u . The numerators of the fractions that give the estimates are called the A . ( I ) . + l ) . c. Notice that every pair of columns contains (1. . THREE FACTORS With three factors there are eight points. This is orthogonality.l ) .h 2 c (X*X2 = +1) 2 c (x. a b c . We can add another column to the design matrix to correspond to the interaction term.u . (1. 6 .
and xzx3. and ABC = 4.28 The main effects are obtained by dividing thcse contrasts by 4. and RC.44 . temperature. B = 7. AC ( x .’ ( x 2 x 3 ) . B. A C = 20. Y. An explanation for the apparently large A C interaction can be found either by making an interaction plot or by making a 2 X 2 table of the totals for the four sets of levels o f A and C . Thc dcsign matrix and the data are given in table 175. we can make columns for x .1. x .and then proceed in the same way. and C. x .THREE FACTORS 289 o f interest.54 + 48 40 R +54 48 40 c +56 . This data comes from a chemical experiment. To calculate the contrasts for the interactions. A . pressure.84 54 . and C. The thrce factors were A . B. This is shown in table 17. cach corresponding to one of the seven d.and C’.2. ) . and C’ = 16. Exutnple 17. . x 3 . The response.48 + 84 42 4 92 +42 t92 +44 + 84 +42 +64 i92  t. x ? .5. B. ignoring B. acid strength. HC = 4.f. The A .’ contrasts are calculated by multiplying the Y vector by thc appropriate column o f the data matrix. Their estimates arc obtained by dividing their contrasts by 4. is the coded yield o f product.5.56 . and the thrce twofactor interactions: A R ( x . Table 17.. ) .1. There are thc thrcc main effects. The interactions are AB = 4/4 = 1.1 40 A +56 44 . They are A = 25.5. x .
17.6.5.54 .84 42 + 92 +4 +56 + 54 56 44 +84 42 I 92 + 80 .t  t + + 40 + + BC i 40 +48 .86) .94) = 110 .56 44 41 84 42 92 48 54 56 +40  All 54 48 AC t 40 48 t34 . although simple..290 DESIGN OF EXPERIMENTS: FACTORIAL EXPERIMENTS A 1 TWO LEVE1.S Table 17.10= 100.5. can be tedious and errorprone when the nunibcr of factors increases. THE REGRESSION MODEL 0 The calculations of section 17.(104 . Ih 84 +42 + 92 c A i Iigh Low Total Low 103 94 198 High 176 86 262 Total 280 180 460 The AC contrast is (176.2 + t . They can be niadc niore easily on a PC by using a regression program to fit the following model: .
the estimates of the main effects and of the twofactor interactions will be the same whether we include or exclude the higherorder interactions. 3 We could add anothcr term. corresponding to the threefactor interaction.5 xl + 0.00 ~ 2 x 3 2 3 2694. e.50 ~ 1 x t 10. This is illustrated in table 17. x z x to. Fitting the Regression Model for a Z3 Factorial M‘I‘H >set X I in cl D A T A > .3’ c4 ‘xlx3’ CS ‘xlx.1 + I +1 DATA > end MTB >set x3 in c3 D A T A > . It is important to note that because of the orthogonality of the design. equation (17.1 + 1 1 + I .2.00 2726. lable 17. which would force a perfect fit. which is an edited regrcssion printout from Minitah.1.1 . ~ .1).5 + 12.00 32.202 F .6. including the constant term.3’ c6 ‘~2x3’ MTB 3 regress c10 6 clc6 The regression equation is y = 57.I +1 +1 .3 ‘x.00 x3 P Analysis o f Variance SOURCE DF Regression 6 Error 1 TOTAL 7 SS 449.REGRESSION MODEL 29 1 The coefficients in the regression equation are multiplied by 2 to obtain the effects.l 1 1 .] .6.03 0.00 MS 14.1 + 1 .. That would makc a modcl with eight terms. z 3 x ..1 +1 DA’I’A > end M1’B >set x2 in c2 D A T A > .1 + 1 tl + I + I DATA > end MTB >let c4 = cl*c2 (This puts x1x2 in c4) MTH >let c5 = c l * c 3 (This puts xIx3 in cS) MTB >let c6 = c2*c3 (This puts x2x3 in c6) MTB >set y in c10 DATA > 40 48 54 56 44 84 42 92 DATA > end M T B >name ci ‘xl’ c2 ‘x2’ c. We usually make the working assumption that interactions involving three or more factors arc negligible a priori and omit them from the model.g. we will restrict the word interaction to mean only twofactor interactions unless we make a spccific statement to thc contrary. For the rcniainder of this book. A =2p. and eight points.1.0 ~ 1 x .00 + 3. ABC.00 32.6.50 x2 + 8.
the total of their sums o f squares can provide us with an error term. cxccpt for the total. I . 17. equivalently. and then column 4 from column 3. Take any pair o f columns. the entries. ac.c . we can rewrite the design matrix to incorporate columiis for the higher order terms. see tablc 17. Column 4 contains the contrasts in the standard order.6. are the sums: a + (1).5.7. c. by squaring the effect and multiplying by 2'12). With the model that includes all the interactions. together with the grand total.I . we divide every entry in column 3 by 4.( l ) . (In a 2" factorial.b . through column n + I). ub. followed by the differences: a . Column 2 is made up of sums and differences of pairs of observations. and enter them in column 1 of the work sheet. 1. . Column 3 is obtained from column 2 in t h e same way. ah t b .5.7. + 1. We then go through rcpetitivc cycles of adding and subtracting. and . YATES' ALGORITHM This is another way of calculating the contrasts that does not require a PC with a regression program. Again. a c + c. +1 twice. we illustrate the algorithm by applying it to the 23 example of section 17. In this section.2.. the operation is continued for n cycles. and abc+ bc. in order.6. ac . b. It is much faster than computing each contrast separately. including those with more than two factors. and ahc. see table 17. a .9.292 DESItiN OF EXPERIMENTS: FACTORIAL EXPERIMENTS AT TWO LEVELS Table 17. The first step in the algorithm is t o write the observations in the standard order: ( l ) .t I + + t  t The sum of squares for an effect is obtained by squaring the contrast and dividing by 2" (or.When we omit the higherorder interactions from the model.bc.2 A 1 R 2 AB 3 + C 4  AC 5 RC 6 ARC 7 f (1) I2  b (ih C (I('  c + +  + t   b + + bc a c + + + + + + + +    + t +   +  i + + . We mention this again in section 17. Yates (1937) presented an algorithm for calculating all the contrasts simultaneously. They contain + 1. ab . we notice the orthogonality. + 1 twice. as we did in section 17. 1 twice. hc. and ubc . In column 5 . 1 twice. which appears in the first place. Thc differences arc always taken in the same waythe second observation minus the first.
these are the effects. B.A N EXAMPLE WITH FOUR FACTORS 293 Table 17. the variance points each. these are the sums of squares for the effects. Yates' Algorithm for 2' 1' 40 48 54 56 44 84 42 92 2 88 110 3 198 4' 460 28 100 5' 57. and D. which is divided by 8. In thc general case. and 2'. The output from applying Yates' algorithm is shown in table 17.5 25 7 1 16 6' 26450 Mean 1250 A 98 El 2 AB 512 C 800 AC 32 BC 32 ABC 262 10 90 128 134 8 2 40 50 6 6 10 22 64 80 16 I6 4 20 4 4 ' Column 1 contains the data.~ unless the absolute value of its estimate is greater than twice its standard dcviation: This is sometimes called the minimum hurdle. whether the sample is preheated ( n o = low.1. He was investigating the homogeneity of mixing a vinyl compound used in nianufacturiiig tires.8. The. C. The response recorded is the torque o f the motor when the heads are running at a constant speed. It is also clcar that C is unimportant. 17.1.the weight of the sample. Church (1966) at the General Tire and Rubber Company. four factors are A . we can argue as follows. of an effect is C T ~ / ~ " We . the temperature of the vat. What about A and AB? If we know cr. With 16 points. AN EXAMPLE WITH FOUR FACTORS These data come from an experiment conducted by A. C. and that those two factors arc surely important. an effect is the difference between two averages of eight With 2" points. . It is clear that the main effects of B and D are very much bigger than the other effects.7. Column 6 is made by squaring the entries in column 4 and dividing them by 8.a'/8 = a2//4. Its variance is a'/8 +. 24.8. column 4 the contrasts. A sample of thc compound is placed in a vat and mixed with heads driven by a motor. An engineer who plans to conduct several factorial experiments can makc macros for 2". can arguc that an effect will not be significant . column 5 the estimated effects. and column 6 the bums of squares. we should divide the contrasts by 2"' and their squares by 2". yes = high). the speed of the motor.
75 438.0 .75 58. . This gives a pooled sum of squares 5625 + 900 + 2756 + 16. it can be estimated by taking duplicate observations at the points of the fraction or by taking several center points.5 10 320 EFFECT 3195.9.00 26. one could then declare as significant any effect whose sum of squares cxceeds 6.50 20. ‘The 5% valuc for F( 1. is midway between . This is what happens when we use a regression model that has terms only for the main effects and the twofactor intcractions.00 25.25 73.. we have a problem. and A B to be significant. ES’I’IMATIN<. i. and s o we can pool them to give us a sum of squares for error.8. BCD. Instead of restricting themselves to using the higher interactions for t h e error term. we would pool the sums of squares for ABC. as a working rule.50 765.e. whence s’ = 6387. THE VARIANCE If CT is unknown. therefore.00 76.1 and + 1.790 3510 9 0 I60 270 .75 15. In the absence of such duplicates.1.75 1 12.00 40832100 39006 770006 50625 1O 6U 4556 13806 5625 2340900 23256 2 1756 2500 2756 16256 6w 4 ss KANK 16 12 A 14 €3 13 AB 2 c 5 8 90 0 6 15 11 10 1 3 4 9 7 D 17. R. too. Yates’ Algorithm for the Church Data OBS (1) a b ab ac DATA 1460 C bc ahc d ad bd abd cd acd bcd abcd 1oMI I 065 2235 2070 1795 1730 2540 21w 1700 1670 1345 990 1075 1455 1330 CONTRAST 25560 .75 40. The following alternative procedure is sometimes suggested. must be due only to noise.937. A B D . ACD. We would thus declare 11.25 63.61 x 6387 = 42218.33. and come very near to declaring A to be significant. In the example in the previous section. points where each x.294 DESIGN OF EXPERIMENI’S: FACIORIAL EXPERIMENTS AT TWO LEVELS Table 17.5) is 6. interactions with three or more factors are negligible.61 . O O 98.470 300 6120 610 590 120 200 210 . Their sums o f squares. A traditional approach is to decide that. engineers might set some arbitrary ad hoc level of smallness and pool all the effccts that are less than that standard.75 37. and ASCD.256 + 6400 = 3 1. It is fraught with danger and should 1101 be uscd.
The bias makes the estimate too low. totaling 24.7) is 5. and ABCD. The list contains six effects: A . The 5% value of F( 1.000. A C .10. We introduce it when we construct the pooled sum of squares by loading it with the smallest terms that we can find. This has lately given way to normal plots.) Since then statisticians have turned more toward graphical procedures. A BD.NORMAL PLO’I’S 295 In this example.59 x 3477 = 19436. in which he plotted the absolute values of the effects. NORMAL PLOTS The procedure that was described in t h e first part of section 17.1 contrasts for the various effects are orthogonal to one another and are linear combinations of D * 1. That would give a pool consisting of C . ABC. 13. B . s. and B D ! The risk in this procedure is that we are using a biased estimate.2  ** * * 2 * 0. A B .f. see section 16. beginning with Cuthbert Dimicl’s (1959) halfnormal plot. ‘I’hc rationale of the normal plot is that the 2” . it led to the introduction of alternatives to Fisher’s least significant difference. ! I I I . (A similar phenomenon occurs in comparing all the treatments in the oneway analysis of variancc.4.9 was a standiird procedure until 30 years ago. and s ’ = 3477.IKK) is a fairly low value for a sum of squares in this group and pool all sums of squares that are less than 10.337 with 7 d. 17.59 and wc should declare any effect to be significant if its sum of squares exceeds 5. AD. o f the ’ variance in the denominator of the F statistics. the engineer might decide that 10.2 * * * t* . CD. ACD. when it became apparent that it led engineers t o declare too many false significant effects.0 * 8 * I 1.
1 1. giving it the values in column 3 of the matrix. Normal scores for the 15 contrasts are calculated by computer. The normal scores arc then plotted against the actual contrasts. b. we had eight observations and. a. Could we.r2 instead of S. the points ought to lie more or less on a straight line. instead. 17. use those four degrees of freedom for extra factors. The only other way to add a third column that is orthogonal to the first two is to use x. + P A + P3. We could go ahead and fit a maineffects model for three factors: Y= w. Supposc that we add a new factor. at a price. The three columns are orthogonal. FRACTIONS In the 2’ experiment that we described earlier. Figure 17. 17. THREE FACTORS IN FOUR RUNS The complete twofactor design has four runs with three degrees of freedom. but the points for the large effects will appear as outliers.10. If the effects were all zero. The complete 27 factorial calls for 2 X 2 X 2 X 2 x 2 x 2 X 2 = 128 runs and we will run only eighta onesixteenth fraction. including A and AB.XIX2. If there are a few large effects. as in chapter five. We added a third column to the design matrix to include the interaction.12. If we were to plot them on normal probability paper. + P I X .5 + e * The estimate of )(j3 would be half the C contrast: rtbc + c . seven degrees of freedom. lie on a straight line.1 independent normal random variables. the plotted points for the others will still lie on a line. wc would have 2“ .1 is a normal plot for the Church data.2% DESIGN OF EXPEKIMENI’S: I’ACI’OKIAL EXPERIMENTS A T TWO LEVELS normal variables. It is obvious that there are two extreme contrasts (two significant effects): D on the positive side in the upper right of the plot.. The other contrasts. They were introduced by Finney (1945). We get a 2 ” / / A a half replicateconsisting of the points c. The famous industrial statistician Cuthbert Daniel would call it a 2’‘ fraction. C. hence. o r write 27//8to denote a sevenfactor experiment in eight runs.u . and ahc.h 2 . We used three of those degrees of freedom for the main effects of factors and the other four for interactions. and thus accommodate seven factors in eight runs? The answer is yes. Such designs iire called fractional factorials. either positive or negative. each with variance 2”v2and zero expectation. and B on the negative side in the lower left.
A B with C D . we are better off than we were in the 25 1 case. we must be prepared to assume that it is due to the main effect A and not to the interaction BC.1 .13. and C. by setting x. = x.xj = C x2x3= 0. It is a cost that we may or may not wish to pay. we obtain the complementary half replicate: ( I ) . If we set x .1 A u + +  C D (1) ad hd ah cd ac hc ahcd + + . AC with HI). = . where main effects were aliased with twofactor interactions.e. or xlx2x3= . = x2x3. Taguchi calls them lattices.x2 = 0. However. the twofactor intcractions are aliased in pairs. and A D with BC. R is aliased with AC. Thus. B.is orthogonal to A . Since xz = 1 at each point. The main effect of C is conipletely confused with the AB interaction. Two factors in a design arc orthogonal if the products of their coordinates sum to zero. equivalently. FRACTIONS WITH EIGHT RUNS Suppose that we were to takc the 2’ factorial and iidd a fourth factor.FRACTIONS WITH EltiHT RUNS 297 What price have wc paid by doing this? We have set x j = x l x z . The design matrix for this half replicate is shown in table 17. The cost of using the fraction in order t o savc time and money is that the main effects arc aliased with twofactor interactions. uc. Similarly. The technical term for this is that C is nliused with A B . Notice that the D column is identical with the ABC column in table 17. i. D. 17.and so A is aliased with BC. Table 17. we can say that this fraction is defined by the relationship x. and bc. we are obtaining a reduction in the number of runs in our experiment at a priccwe cannot distinguish bctwcen main effects and interactions. I).1.13. The fourth factor.2.13. factors A and R arc orthogonal. If we see a largc A contrast.x .x2x3or.xzx3xJ= + l .6. A design like this in which all the factors are orthogonal to each other. x.. some say confounded. if 2: x. It follows that x.x2x3= 1. C x l x z = C x. is callcd an orthogonal array. and all four main effects are aliased with higherorder interactions that we arc prepared to re ard as negligible. ub. x 2 . In either case. In that sense.
for ncw factors. Wc have equatcd I1 t o A H C .X. or. E‘ to AC. We take the design matrix for the complete 2’ factorial and use all four interaction columns.X*X3X*X5X(. equivalently. = XjX0 = X4X7 .X. The main effect A is aliased with three interactions.14. = XIX2X3X4 = 1 .. = X2X4Xh= X?X3X5X6 = X.14. we can writc the four effccts in an alias chain: A Table 17.X7 . AC.S 17. SEVEN FACTORS IN EIGHT HUNS We said earlier that we could obtain an orthogonal array for seven Factors in eight runs.K4 = X. = XZX1X7 . E to Ah’. EXPERIMENTS AT TWO LEVE1. or A = BE .KIX&X7 = XIX3X5X7 = X2X4XSX7= XIX2X6X7 = XSX.1. are called the defining contrasts.’F 1.CF = DG. These products. AB. and G t o RC‘.+ + + + +  +  I I .+ + + + . we scc that we actually have a fraction defined by 1 = X. Alternatively. we have set XIX2X5 = XIX3Xh = X2X3X.X7 = X.X2XS = X3X4X5 = X1X3X(. = X.DG B A   E + C  F + G + D t 1 t  + +  + + + + t + t +  i . which define the fraction. ‘Taking all the products of thcsc defining contrasts.X.X. and A R C .298 DESIGN OF EXPERIMENTS: FACTORlhl. wc scc that X l = X2X.X.14. by the relationship I = ARCID= A B E = C D E = A C Y = BDF= BCEF= A D E F = RCG = AIJG = ACEG = RIIEG = ARFG = CDFG = EFG= AHCDEFG Multiplying through by xI and ignoring the terms with more than two factors.XZX3. That proccdure gives us the design in table 17. BC.1 + R E + C.X.
an experiment with four factors.1. _ +  beef ubcd .THE L ( 8 ) LATTICE 299 There are six similar chains: B+AE+CG+l>F C + A F t RG+ DE + AC. F = .2 shows both the design matrix for this lattice and also the interaction table.15.HF + CE E I. THE L(8) LATTICE If we multiply the columns for E .C'F . Suppose column 4 is chosen for C.BE . chooses four columns of the lattice for the factors.15. this says that thc A N interaction appears in column 3.A D .14. Taguchi changes the ordcr of the columns and writes 1 and 2 where we have written . any factor assigned to column 3 will be aliased with the AR interaction. It is just that some like to have the base point (1) in thcir dcsign. The alias chains are changed to A . and one can confirm from the interaction table. 1.A C . F. That leaves column 7 free for the fourth factor. we maintain orthogonality and get another fraction in the same family. we have L) = ABC. From a statistical point of view. E = .and +.. This time.1 ______ A B E C F G   + +  +  + +    + + +  D t  i +  +    c t t t  + i + + +  + (1) udef hdq rthjk Clij~ llceg + + + +  +  . ' and the BC interaction is in column 6.15. one fraction is as good as the othcr. and so that column should not be used for C'.DG. The entry in the interaction table in row (1) and column 2 is 3. and C = . The AC interaction is in column 5 .e. for example. 'I'ablc 27. that this Table 17. i. He calls the design his L ( 8 ) lattice. Supposc that column 1 is chosen for A and column 2 for B. and G in table 17. etc.A H + CD + FG D F G t AC+BL)+EG + AD + N C + E F 17.BC'. which is shown in table 17. The cnginccr who uses this lattice to design.15.1 by 1.
14.13. the following design results: Col.1.300 DESIGN OF EXPERIMENTS: FACI'ORIAL EXPERIMENTS AT TWO LEVELS Table 17. 1 (1) 2 3 4 6 7 5 7 6 1 (5) 6 ~~ ~ 7 6 5 4 3 2 ( 2 ) l (3) 5 4 7 2 3 (6) 4 5 (4) 3 2 1 (7) assignment of factors to columns givcs a design with main effects clear of interactions. I A 1 1 1 1 2 4 7 I) B 1 1 2 2 1 1 2 2 C 1 2 1 2 2 2 2 2 I 2 1 2 1 2 2 1 2 1 2 1 cd bd bc ad ac (1) ah ahcd This is the design that we obtained before in table 17. THE L(L6) LAT'I'ICE The method that wc used in section 17. can be extended . 17. No. 1 1 1 1 1 2 2 2 2 2 1 1 2 2 1 1 2 2 3 1 1 2 2 2 2 1 1 4 1 2 1 2 1 2 1 2 5 1 2 1 2 2 1 2 1 6 1 2 2 1 1 2 2 1 7 1 2 2 1 2 1 1 2 interaction Table for L ( 8 ) Col. If the engineer goes through with that assignment of factors. of adding extra factors by equating them to interactions in the basic factorial design.16. No.2. Seven Factors in 8 Runs Design Matrix for L(N) Col.15. No.
1 l 1 1 1 1 1 1 1 2 2 2 2 2 2 2 2 2 l 1 1 1 2 2 2 2 1 1 1 1 2 2 2 2 2 3 (2) 3 l 1 1 1 2 2 2 2 2 2 2 2 1 1 1 3 3 2 I 4 l 1 2 2 1 1 2 2 1 1 2 2 1 1 2 2 4 5 6 7 l 1 2 2 2 2 1 1 2 2 1 1 1 1 2 2 8 l 2 1 2 1 2 1 2 1 2 1 2 1 2 1 2 8 9 l 2 1 2 1 2 1 2 2 1 2 1 2 1 2 1 10 11 12 13 I4 l 2 1 2 2 1 2 1 1 2 1 2 2 1 2 1 l 2 1 2 2 1 2 1 2 1 2 1 1 2 1 2 l 2 2 1 1 2 2 1 1 2 2 1 1 2 2 1 l 2 2 1 1 2 2 1 2 1 1 2 2 1 1 2 l 2 2 1 2 1 1 2 1 2 2 1 2 1 1 2 15 l 1 2 2 1 1 2 2 2 2 1 1 2 2 1 1 l 1 2 2 2 2 1 1 1 1 2 2 2 2 1 1 6 l 2 2 1 2 1 1 2 2 1 1 2 1 2 2 1 Interucriori Table fur L(16) Col.16. SCREENING EXPERIMENTS When all seven columns of the L ( 8 ) lattice or all 15 columns of thc L(16) lattice are used for factors. and so on. 17. the design is said to be saturated. In designs that are saturated.SCREENING EXPERIMENTS 301 Design Matrix for Id( 16) Table 17. 31 factors in 32 runs. the main effects are all aliased with .1. 1 5 7 6 5 4 3 2 9 1 0 1 1 1213 1415 11 (I) 4 7 (3) 7 6 (4) 1 (5) 6 5 10 8 9 5 11 10 9 8 2 12 13 14 15 3 13 12 15 14 (6) 1 14 15 12 13 (7) 15 14 13 12 ( 8 ) l 2 3 (9)3 2 (10) 1 (11) 7 4 9 X 10 I I 12 14 15 I5 14 8 9 9 8 I0 11 11 10 4 5 5 4 6 7 7 6 (12) 1 (In) 13 15 12 13 10 11 8 14 13 9 6 7 2 3 (14) 4 5 12 11 10 9 8 7 6 5 4 3 2 1 t o generitte designs for 15 factors in 16 runs.16.17.1. No. The L(16) lattice is shown in table 17. Fifteen Factors in 16 Runs Col. or nearly saturated.No.
A R .1.17.8 1.302 DESIGN O F EXPERIMENI’S: FACTORIAL EXPERIMENTS AT TWO LEVELS several interactions.875 0.2 22.225 1.7 B AB C AC BC ABC The D effect is .6 23.15 35. and wants to spot those important factors to be investigated morc closely in a subsequent experiment. The engineer assumes that only a few of them will be important. wc have three choices.70 2.5 89.76 A 7 1.2 26. If we know what defining contrasts were used to obtain the fraction.075.76 1.6 84. In this case. and the G effect is + ABC‘ = 2.1. etc. As a consequence. (An example of a screening cxperimenl.6 6 Eff . E = .6 4.) An L ( 8 ) lattice is used in a screening experiment for seven factors. Exumpfe 17.A C .4 43.2 45.7 43.AC = 0. The lattice and the data are shown i n table 17.2 19. they are used mainly for screening experiments.4 13. In a screening cxperiment.275 4.15. 1 2 3 4 5 173.3 0.= .5 17.B C .2 8.425.8 3. and G = ABC.6 t + i  + + + t +  + + + + + +  + +  + +   . the engineer considers a relatively large number of factors.8 2. The sums of squares are not affected by the changes of sign.725. I.4 26.53 0.8 18.175 0.7 18. 075 2.5 5. It is obtained from the 2’ factorial in A . the Ed‘ cficct is .7 19. we should carry out the algorithm for 23 in ABC and then equate D to .17. and C by setting D = . To calculate the effects.6 24.9 4.2 1. H . The order of the columns and the assignment of factors to columns is slightly different from the design in table 17.4 7.1.8 23.1. The engineer argues that the few important factors will stick up through thc undergrowth and be detected._ +  + + +  +  +   G  Y 17. t h e ti effect is NC = .05 2.6 21.225.3 22. Table 17.425 ss 16.1.3 21.7 2.27.1 16.9 4.A l l = t4.7 11. H t + t __  c f n  E F  . 2.A R .31 11.725 I.2 24.9 4.2 40. we can carry out Yates’ algorithm. The Data A .
8 +23.9 0. 7 I = ABCD = RCEF = ADEF = A C E G = BDEG = ARFG = CDFG In this larger fraction. and the 16 points satisfy the dctining rclations 1 = x1x*x3x4 x2x3x5x.14.6 22.7 .28 A + 16.4 26.2.8 +23. and so on.5 1. Ll 17. XlX4XSXh= x.7 t 18.3 .6 21.14.19.2 24.7 .and G stand out as candidates for further investigation.6 +21.6 21.8 +23.2 . The threefactor defining contrasts. l‘able 17.8 E .17. because their defining contrasts have at least four letters.2 +11.17. Take each point in the first design and change the levels of every factor. Then cfg becomes ehcd. such as xIx2x5. The design in table 17.18.4 13 . the engineer can use the method of section 17.9 +322 +0.8 23. The combined fractions form a 16point design. Calculating the Effects A .6 4.17.4 22.88 0.5 or regression.2 22.7 + 18.2 24.17.7 D + 2.2 .19.18 23. D. FOLDOVER DLSIGNS The array in Table 17.17. the main effects are aliased only with higherorder interactions and are “clear” of twofactor interactions.17.0 t21.2 I.2 F G‘ . drop out.2 +24.08 + 10.2 +22.4 c ‘ n .72 .FOLDOVER DESIGNS 303 Table 17.17.7 +1.1.6 +4.2 +19.2 4 + 26. This procedure is known as folding over.4 t 26.6 ~ 26.0 +21.18.4 +26.1 in the following way.1 can also be obtained from table 17.2 + 10. as does x.6 .6 2. We see that A .x.2 shows the calculations for the method of section 17.6 +21.2 t 22.2 23.x2x3x4x.6.15.2 124. as in section 17.7 + 18.x.6 26.19.7 18. Such fractions are said to have resolution IV.1 Dividing by 4: t 2.2 +24.t.5.. 18.13.17.2 t 24.1 is also a resolution IV design.2 .xvx4x7 = = = X ~ X 4 X S X= XIXZX6X7 = x+4x6x7 .4 t 26.42 c ‘ If the defining contrasts are not known.19.2 t 22.17.8 +23.7 18.2 .8 23. 22.6 21 .1 has rcsolution 111 because . The design in table 17.
cd. Plackett and J.x2x3x4x5 1 = . each twofactor interaction is aliased only with a threefactor interaction. some main effects are aliased with twofactor interactions. HESOLUTION V FRACTIONS There arc also fractions with resolution V. Eleven Factors in 12 runs (Ptackelt and Burman. D to column 8. 17. 17. K. This fraction can bc obtaincd from the L(16) lattice by assigning A to column I. Whenever a resolution 111 fraction is folded over. in which we can estimate all the main effects and all the twofactor interactions clear of one another. uc. abde. ae. J. ce. B to column 2.20. bcde. acde.20. 1946) A 1 2 1 2 1 1 1 2 2 2 1 2 B 1 2 2 1 2 1 1 1 2 2 2 1 C 1 1 2 2 1 2 1 1 1 2 2 2 1) 1 2 1 2 2 1 2 1 1 1 2 2 E 1 2 2 1 2 2 1 2 1 1 2 I F 1 2 2 2 1 2 2 1 2 1 I 1 C 1 1 2 2 2 1 2 2 1 2 1 1 H 1 1 2 2 2 1 2 2 1 2 1 J 1 1 1 1 2 2 2 1 2 2 1 2 K 1 2 1 1 1 2 2 2 1 2 2 1 L 1 1 2 1 1 1 2 2 2 1 2 2 1 (1) ubdefk h tufgl ucdjkh bdeghj cefljk dfgjkl ueghkl nbfhjl czbcgjk bdhkl ocdrjl . P. FHACTIONS WITH 12 POINTS During World War 11. It consists of 16 points: ( l ) .1. ahcd. developed an orthogonal Table 17. A useful example is the half replicate o f the 25 factorial defined by x. In resolution I11 designs. de. ab. ubce. Burman (1046). ad. bc. two British statisticians who were engagcd in war work. he.304 DESIGN OF EXPERIMENTS: FACTORIAL EXPERIMENTS AT T'WO LEVELS some of its defining contrasts have only three Ictters. bd. C to column 4. and E to column 15. Each main effect is aliased only with a fourfactor interaction. the two parts combine to form a resolution IV fraction.19.
and what we have said about aliasing does not apply to it. acd. S9. 41.2. The fraction was constructed by taking a full factorial in A . Note that in some of them you will find that you reach different decisions about which factors are significant depending upon whether you rely upon normal plots or regression! 17. cefg.6 . 24. adeg. see table 17.2 . 42. 26. 44. ad. nb. a . bd.4 17.8. Which effects are important'? ( I ) .1. B . These 12point arrays should not be used when there are interactions. 27.1 . The following data come from a 2' ' fraction.0 .0.6. abc. 54. nc. C .EXERCISES 305 array for 11 factors in 12 runs. 40.2. Calculate thc contrasts for the main effects in this 2' factorial. abc. 43. bc.2.2. 49. 46. bc. hdfg. 29.0 . Which columns in the L(16) array correspond to the seven factors? (1) 95 dcfg 42 cfg 92 cdc 9 beg 90 bdf 65 beer 15 bcdg 3 . hcde. b . acdf.1. 64. 46. F = .8 .4. d .4. and setting E = .2 . These are the data from a Z3 factorial cxperiment.2. Calculate thc contrasts. Which effects are important? 17. The rcsponse mcasured was the percentage of defective stitches. Match each of the factors with a column in the L ( 8 ) array.3 : uhcg. 20.0 b. 68.C D .20.A R C f > .6 .0.6. 56. Do any of the factors seem to be important? ( l ) .6. B. and G = . EXERCISES These exercises involvc 2"k factorials. 49. and E. ubcf. 48.4. uhcd. 46.7 . c.7 . a . abd. 40. and C contrasts. 67.2. 47. 59.4. The derivation of this array is quitc different from the derivation of the other fractions that we have discussed. 23. 20. This fractional factorial for seven factors in 16 runs was used in investigating a stitch bond process. 43. 47. Calculate the A .0.B D .8.5. cd. 43. 17. bcd.3. uc. 26. 58.6 c . (l).4. ab.
17. he.8. 52. 46. hc. a.x2x3x4x5 1 and the experimental points became = .306 ag 94 DESIGN OF EXPERIMENTS: FACTORIAL EXPERIMENTS AT TWO LEVELS adef 4 acf 32 acdeg 5 ahe 95 abdfg 24 abcefg 6 abed 3 Calculatc the contrasts and identify the important factors. If you want to estimate A R and CD instead of AB and AC with an L ( 8 ) array can you do it? 17. 47.8. Show that you can also estimate the other five main effects if the other interactions are zero. Which columns correspond to the interactions? Show that the fraction (l).5. 46. NC. uhd.2. 5 1 . 56.4. 44.9.0 . Which points are in the fraction and which interactions are aliased with A? Which columns of thc L( 16) lattice give you this fraction? 17.7. 44. c. so that the fraction was the half replicate defined by 17. 46.6 46. obc. ( l ) . ub.9.6. 48.3. NC. assuming that thc only interactions that might exist are A B and AC. F = A C D . ahc.5. 56. Which interactions are aliased with A? with E'? 17.A B C D . (red. bc. Suppose that a fifth factor had been added to the example in table x. Was there a significant difference between thc two runnings? Which o f the factors A . An engineer makes a 16point fraction for seven factors by sctting E = .6 47. Find an assignment of factors to columns that achieves this aim.8. 7 . c . Find a 16point fraction of a Z6 experiment (find an assignnient of the six factors to the columns of the L(16) lattice) that will enable you to estimate A and the five interactions that contain A whether or not the othcr interactions arc zero. B . Bde.5 .2 . C is significant'! 17. ah.1.1 . cde. 56. h .2. 47.6. and G = AHD. An engineer carried out a 2j factorial and obtained the following set of data: ( l ) . abce meets the requirements. h. An engineer wishes to look at five factors in an L ( 8 ) array. ne. Thc experiment is repeated later and provided these data. a.6 . 47.
Which lines in the Yates table now correspond to the main effect of E and to the interactions that contain E'? Notice that y o u now havc a saturated fraction. ce. de. C is the temperature of the cold plate. ahce. He used two different photoresist materials. Considcr the smaller factorial that consists of the eight points at high A . Y2 is a measure o f the uniformity o f the film. I 11.7 54. Fish patties were made at the University of Florida from a fish called mullet.6 42. A is the ambient temperature. The wafer is first placed on a cold platc for 30 seconds to stabilize its tcmpcraturc.10.4 14. Suppose that we decide to operate at that level of A . First the patties were .12. is the average thickness of the films on four wafers at each set of conditions.0 ud bd abd cd UCd d 44. Will the plot that we made in thc text tic changed? How could you handle the analysis using regression? What could you use for an error term'! 17. You can makc ii normal plot in the usual way. He does not mention the levels o f the other factors. ae. B is the temperature of the photoresist. Arc any of the other factors significant'? Are any of your results shown in interactions? 17. McBcth (1989) investigates the effect of thrce temperatures on thc thickness of the film of resist that is put on a wafer.1 13.9 12.9 16. ucde.EXERCISES 307 (l). Then the photoresist material is applied.6 65. Each degree of freedom corresponds to either a main effect or an interaction.0 32.6 15. cd. 256 767 32s 805 389 737 440 942 30.0 hcd abcd 389 765 412 857 457 738 540 880 36. be.6 33. A .9 Answer the following questions for each of the two responses. He measured two responses: Y .8 5s. ad.4"Cand 235°C.8 50. ab.10 that the uniformity is greater at the higher level of the ambient temperature. It is clear in exercise 17.11. ahcd with the same data. uhde. bcde. l'hc levels of A were 20. hd. they constitute factor I). 'These data are taken from an example that is quoted by Cornell (1981). Which are the important factors? Is there any signifcant difference between the two materials'? 17. bc. l h r e c factors were involved. uc. KTI 820 (low) and KTI 825 (high).
67. ( l ) . acd. 73 . 47. 5 9 . 39 . 1 5 : ac. cde. 28 . . 17 . 61 . 5 5 .84. and E. 25 or 40 minutes. beg. h. ad. 38. 2. 6 3 .7 . uhcd.13 Which are the important factors and interactions? Explain. 67 . What interactions are aliaseci with the main effects of the factors‘? Show that A and F are significant factors. 66. ahcde.6 . ( I ) . ahc. 77. 120. 42. Which other interactions are aliased with A F ? Would you be content to attribute all that sum of squares to AF. he. 6 4 . 17. ude. a. Show that it is a foldover design. adcf. 36 .7 .Y . 6 2 . acf. The response recorded was a measurc of the hardness of thc texture of the patty. 61 .7 .0 .2 .7 . 53 . c. ubc. 140.65.0 . the tcmperaturc. 13 . The factors are A . 23 . ce. bed. 4. 59. hydrogen. .. beef. 65. C. ucde. cde. 16 . b. 375 or 425 degrees. Fit a model to the data using Z = ln(Y) as thc response. The data are ( I ) .4 .4 . 41.14. ahfg. 32. & . ae. ahe. Y is the coded rate constant as determined by a complex chemical equation.8 . steam.4 . 140 . for a given time (factor C). 68. B .6 . each at two levels. ahcdefg. cu’fg. hce. bd. I). be. 68.0 . c. 1. 36 . ahe. 111. The term involving A F is also significant. fib. ac.04. ab. bc. abcd.13. uhce. carbon dioxide. 2.13. carbon monoxide. cd. ahde. 3.32. Charles Hendrix of Union Carbide gives the following example of a fraction for seven factors in 16 runs. the dcsign used was a 2’ factorial. 29. ace. 17. Thcn they were cooked in an oven at either of two temperatures (factor B). 3. hdeg. 26 .308 DESIGN OF EXPEKIMENTS: FACTI’OKIAL EXPERIMENTS AT Two LEVELS deep fried (factor A ) for 25 or 40 seconds. 4. 30 . abd. A petroleum engineer investigated dynamic control of the production of synthesis gas From underground coal gasification and determined an accurate rate equation for new catalysts that are resistant to sulfur poisoning by developing a simple model using five factors. d . a . 29 .01 .86. 5 6 . bde. and not to the others? Make up a twoway table to explain that interaction. a c q . 81 : de. adg. bdf. e. 140 . 62 .6 . I I . 78. 86. hcdr. 1. 32 . 66.
17.1 . We cannot use Yates’ algorithm for this problem.6.3 . = X . + I . Include raw material as a fourth factor.6. Otto Dykstra (1959) gave an example of partial duplication of a factorial experiment. 18.2 . We can estimate from the 12 points all four main effects and all six interactions. and so on. 24. b. 1. ac. A D . 25. Are the interactions orthogonal to the main effects as they were in thc L ( 8 ) array:’ Can you estimatc the main effect o f A by subtracting the average at low A from the average at high A ? duplicated portion) are higher than the responses in the first row. 1 x. 21. X 2 . (l). be. 1971). c. Suppose that the last four points had actually becn run a wcek litter using a different batch of raw material. 27. abc. f> is estimated by the average of the two estimates from row 1 + row 3. bc. AC.2 . repeat your analysis. RD.3.1 . letting X . be the coordinate of A . (l).2. 23..EXERClSES 309 17.and CD are estimated from the first and third rows of the data table. The values of the predictor variables are XI . = XJ. ab. We can use multiple regression. . 20. and C’ are estimated from the last cight points. + I . 1.X. In the data table in exercise 17. 1. ab.15 the responses in the third row (the . first with only the terms for main effects. ac. and BC are estimated from only the first eight points.9.1.1 1 +I +I 1 +I . 1 +1 1 +1 1 +I +l 1 +1 I 1 +1 . .3 .16.. 1.8 . Make two fits.15. and from row 2 + row 3. ‘I’his design is a threequarter replicate (John. 24. D. 1. which takes the values: 1.1 +1 1 1 +l +1 1 1 +1 x 3 +I +1 I +1 1 1 t1 +I X. 26. A . 25. 29. 1. = X. + I . a. Notice that AB. +1. H. X We can add the interactions in the usual way by letting X. and then including the interactions. 25.
517 .55 . 475 : cdrf. Short times are dcsirablc. 490 . 5.50 . Which is it? What would you suggest as a viilue to replace it? 17. What are those conditions'? What would you estimate the delays to bc in thc two cases? 17. Calculate the contrasts for thc 11 factors and dccidc which are important. 5.19. acdfgh. in standard order: 138 124 145 153 149 157 148 154 358 166 150 159 152 158 151 150 One of the observations seems to be a bad observation. 399 . nefh. 4.33 . . bceh. cefhjk.1. (1).20.02 . 6. 7. 5. 5.87 . The following data come from an expcrimcnt using the Plackett and Burman (1946) L(12) lattice in table 17. 522 . 521 .310 DESIGN OF EXPERIMENTS: FACI'ORIAI. 7. nbcdefgh. 483 : rib&. which arc the sets of conditions that minimize and maximize the delay. bdfh. abcf. abgh. 5. 586 . 6. EXPEHIMENIS AT 'TWO LEVELS 17. In a simulation study an engineer measures the delay time of a dcvicc. 568 . 561 The engineer is interested in finding the best case and the worst case. cfgh. 506 . 4. abdefk. he/B. abfhjl.18. aceg.55 . 389 . 484 UL&.17.82 . and then folds he lattice over and makes eight more runs degh. 562 .44 . 5.08 . bcdhkl. bcefgl. A 2' experiment has the following data. The cngincer begins with seven factors i n eight runs using an L(8) lattice: ( l ) .48 . ncdh. abcgjk. bdeghj. dfgjkl. hciig.48 . 702 .00 : aeghkl. 452 . acdejl.
we can break the sum of squares into linear and quadriitic components. The analysis of variance table gives a subdivision into threc sums of squares: the main effects of A and B.1. but not in x f . and 3..x. Thcrc is. which may be a complete factorial or a fraction. for example. we turn to the 3" factorials. Others denote them by 1 . 1. ~ . x . M.Statistical Methods in Engineering and Quality Assurance Peter W. and the residual and/or A x 13 interaction with four d. The linear and quadratic contrasts for A arc given by 311 .. but because they have only two levels for each factor. In the case of factors at threc Icvcls.. We discussed in chapter sixtcen how the sum of squares for cach factor can be divided into components with single degrces of freedom. T . ctc. and T 2 . again. 1. and xf = 41 at every point. each with two degrees of freedom.x. The twofactor experiment is an ordinary twoway layout with two factors. they cannot handle quadratic models. John Copyright 0 1990 by John Wiley & Sons. model for the 2" contains terms in x. Suppose.x. respectively. n factors cach at three levels. Ysles (1937) used 0. The latter cannot be included because each factor appears at only two levels. a difference of opinion about notation. and that there are N points in t h c design. FACTORS WITH THREE LEVELS The 2" factorial designs are flexible and powerful. Some statisticians denote the three levels by . that we denote the sum of the observations at A = 0 . we will see an argument in favor of this approach when we consider fractionation.. Inc CHAPTER EIGHTEEN Design of Experiments: Factorial Experiments at Several Levels 18.f. Therefore. we need at least three levels of cach factor. . In order to incorporate quadratic terms in a model or to estimate operating conditions that optimize a response. 2 . each at three levels. . 0. and +. and 2 by 'ZU. The . and 2.
which becomes 1.T .. for example. With all the degrees of freedom used for the factors. 160.3. 1. ‘I’hc third and fourth factors are introduced by setting x3 = x. We can computc thc averagcs and the sums of squares for the individual factors in the usual way. The sums are then reduced mod(3). = x1 = 2. and D is the flow rate of the etching gas. and selectivity. The nine points in the fraction are shown in table 18.1 A R 0 1 1 1 C 0 1 D 0 2 2 1 1 2 0 0 1 0 0 0 1 1 I 2 0 2 1 2 0 2 2 2 2 0 2 2 0 1 . The engineers considcrcd thrce responses: etch rate in A/min.2cm. We use the notation 3*//9 to denote that it is a fraction for four factors. uniformity. and 550 mTorr. and thcir sums of squares by Lin A and 3(~in A)’ . 450. and 2OOsccm. 1987).312 DESlCiN OF EXPERIMENTS: PACTORIAI. and 325 watts. H is the pressure in the reaction chamber. when x. = 2 x .To. 125.F.3. x3 = 2 + 2 = 4. + x 2 .3.3. EXPERIMENTS AT SEVERAL LEVELS Lin A = T. in nine runs. C.2. They are as Table 18. We look at the first response.8.0.T. There are four factors: A is the power applied to the cathode. the other two are left as exercises for the reader. 0. This example is taken from an experiment to optimize a nitridc etch process on a single plasma etcher (Yin and Jillie. Example 18. The data are given in table 18. FRACTIONS: FOUR FACTORS IN NINE POINTS The 342factorial is a oncninth replicate of the complete design. there is no error term and F tests are not available. which means that the actual numbers are replaced by the remainder when they are divided by 3. 300. 500. and 1. . and Quad A = 2 T . + x2 and x. 2N ’ QuadA and (Quad A)2 2N 18.3. C is the gap between the anode and cathode.1. each at three levels. .1. etch rate. 275.
1.34 134. S.74 5. The totals at thc various levcls o f each factor.0 66. = 1787.0 37.5 54.0 65.5 59.5 57.62 893.55 .5 1.5 72.2..5 11.5 72.45 c Lin in8.66 MS F 36.ou 1138.0 25 77.2.81 39. The analysis of variance table is table 18.1 Temp ("C) H.. = 1871.O.48 53.4.6.0 M8.5 68.4.4 539.5 5555.5 B 574. The s u m of squares for the three main effects arc S.0 84.43 213.0 430.0 5 ~ 15 69.4.5 79. Analysis of Variance SOURCE A 0 C BC Error TOTAL AH AC DF 2 2 2 4 4 4 8 26 1871.0 50.12 2.4.0.5 78.. S.34 25.0 617. together with he contrar S and the sums of squares for the linear and quadratic components.06 1431.0 hX6.5 43.5 592.5 70 15 71.0 62.5 84. which is tungstic acid.5 75.314 DESIGN 01: EXPERIMENIS: FACTORIAL EXPEKJMEN'TS A1 SEVERAL LEVELS Table 18.0 60 5 26.3 157. (ii) thc concentration of hydrogen peroxide.0 55.0 77.11 8.79 222.o 248.30 1.79 35.96 SS Lin SS Quad Table 18.4 203 .5 5 64.0 89.0 107.6 889.0 278.5 15 83.18 498.5 25 79.5 676.5 105. (iii) the level of catalyst.2 1787.5 25 43. (%j W03(%) 0. are as follows: r =0 *=I x=2 Quad A SO3.5 606.2 ss 935.2. = 107.13 2. The data are in table 18. WO.1 n.5 67.0 590.5 thrce factors arc (i) the tcniperature.
but nothing to indicate any curvature in the response.5 1.5 163.0 ..14. and yet there is a significant A C interaction. Since there are nine observations at each level. Temperature vs.0 230.5 3.1.0 .18.7. 617. Table 18.0 .0 .5 43.(sum at x = 0) = (1. in both cases.. The Lin A Lin c' contrast represents a change in the linear trend in Lin C as we change levels of A .0 229. they are the following: linear: L = (sum at x = 2) . C . There are clearly quadratic trends in A and B . the linear component alone has an F value of 105. although not significant at a = 5%: exceeds the 10% value.C. the s u m of squares are L2/(2 X 9) and Q2/(6x 9). At 80 degrees. +2. For a factor with three levels.0 617. 8 ) 3.0.5 2. we should find.THKEE FACTORS IN 27 RUNS 315 The linear and quadratic contrasts for a factor were introduced in section 16.43 = 4.1 %?.0 59.5 80 206. .f.4. Table 18. There is thus some evidencc for a linear trend.2(sum at x = 1) + (sum at x = 2) = (.(96) 0.1 11.4. 0 . A L C Y .3.5 Sum Sum Lin C Quad c'  592.AQCL. the lowest level of acid is best.0 70 227.46. 1). if we fitted a parabola.59. I t i s the difference between the Lin C at high A iind Lin C at low A . there is no difference between the levels o tungstic iicid.13.o Temperature 60 140.0 188. 574. At 60 and 70 degrees.3.C.0 500. . for the AC interaction into single d.o _ _ 574. which is F*(l.18.0 = 77. and A. .12/ 25.f. which. totals for the levels of A and C . l.0 . 590. that the maximum response occurs soniewhere between x = 0 and x = 2. f The totals for the three levels are 0. However. Quad: Q = (sum at x = 0) . at A. It seems at first that. on the average.: A . Tungstic Acid Totals wo. How can we evaluate the significance of this switch in sign'! One way is to divide the four d. higher levels of acid are better. shows the interaction.5 107.0 .1.0 .5 1781.5 The spread in the totals for C is less than the spread for A or for B .5 Another way to say this is to note that the linear contrast for C is positive low temperature and negative at high temperature.5 11. We notice that the F test for the main effect of C is not significant at a = 5?h. 5 % .5 503.O%a. I ) .0 I9Y.5 686. In this case.
0.( . 0 18.0) = 36. Quad A Quad C = (14.0 . for example.09 5.0)(14. which is analogous to a resolution i V fraction. 3X2X2 The Quad A Lit1 C contrast is obtained by applying (1.5. which was done in thc following example. and is 59. 1) to the Lin C terms. Table 18.50 2 0 I 2 2 2 1 1 2 0 1 2 0 1 2 I 0 1 2 I 2 o . the corresponding sum o f squares is (77.0) = 22.18 6. + x2 + x3 . Their values are the following: LinA Quad C=(l. The sums of squares for the four components add up to the sum of squares for A C interaction.42 5.18.29 5.0)l/(3 x 2 x 6) = 36.75 A B C D n 1 2 2 2 1 2 0 1 1 1 1 2 0 1 2 0 2 n 0 1 0 1 2 2 1 2 2 2 2 2 2 2 2 2 2 n 0 0 o 1 Y 6.08 .09 4.64 5.(1.84 5.~2 s. The Lin A Quad C and Quad A Quad C terms arc obtained in similar fashion from the quadratic contrasts at the three levels of C.01 0. FOUR FACTORS IN 27 RUNS We can add a fourth factor to the complete 3’ factorial and still have a design that has main effects clear of interactions.5.0) + 2(3.63 6.0)= 13.5) .1 A B C 1 D 1 Y A B C 1 D 1 o n n o 0 0 o n 0 0 1 5.111 5.44 5.3.98 5.0.72 4.EVEI.85 I 2 n o 0 0 1 I 2 2 2 n I 2 2 I 2 5. We can. 7 . The lion’s share is accounted for by the linear by lincar component .82 6.0)’ = 494.0/(3 x 6 x 6) = 4.0.23 5. S39.5.31 5.316 DESIGN OF EXPERIMENTS: FAC~lOKlAL EXPERIMENTS AT SEVERAL 1.S Since there are three observations at each combination of levels of A and C.46 5.0 + 2(2.24 7. reduced (modulo 3).17 6. S S = ( 1 3 .45 5.5) . Its sum of squares is (36.2.83 5. 0 ) ’ / ( 3 ~ 2 ~ 6 ) = 4 .27 1 2 n o 1 1 I 1 1 o n 0 1 Y 0. set x4 = x .84 6. with SS 484.
783 5.421 5.623 5. B. and we are only interested in main effects.567 0. we ignore C and D and do a twoway ANOVA for A and H .826 5. and s o a contrast will be significant if its sum o f squares exceeds 4. The analysis of variance table is table 18.783 0.5.2.514 9.262 2. the best conditions are the lowest level : . we can niake two passes on a twofactor ANOVA program. None of the quadratic terms is significant.170 2. On the first pass.161 MS .686 6.028 0.Ooo 0.158 ss Lin 0. we ignore A and B and obtain the sums of squares for C and D.748 5. F*(l.41s’ = 0.497 5.567 0.773 1. on the second pass.089 0.781 1 2 6.038 6. of A . 18) = 4.770 0.525 5. The averages for the various levels of the factors are as follows: A C D B 0 5. the response was the propagation delay in nanoseconds in an integratedcircuit device.2. This gives us the sums of squares for A and for B and the total.FOUR FAnORS IN 27 RUNS 317 Table 18.5.164 To minimize the propagation delay.1.5.1. In this experiment. and D from the total. It is not necessary to have a computer program that handles fourfactor analyses of variance for this problem. C and D and the highest level of B.478 0.The sum of squares for error is then obtained by subtracting the sunis of squares for A . C .043 0. Analysis of Variance SOURCE A B C D Error DF 2 2 2 2 0.031 18 TOTAL 26 Example 28. The data are shown in table 18.541 5.896 9. all of the linear terms are.0286 Quad 0.158 0.1 14.748 s2 = 0.318 6.525 5.5.773 I .018 5.41. Because the factors arc orthogonal. The first pass gives the following table: SOURCE A E3 interaction error TOTAL DF 2 2 4 26 18 SS 0.
The design used in example 18. ?'he actual dcrivation of the array.1. B to column 2. This array is shown in table 18.6. C to column 5 . we obtained a Z7' fraction by starting with a 2' factorial and adding the four extra factors by equating them to interactions.5. A similar procedure enables us to derive an orthogonal array for 13 factors.6. The L(27)Lattice ~ ~~~ 1 0 0 2 0 0 0 I 1 3 4 5 0 7 8 0 9 0 1 2 I 2 0 2 0 1 1 2 0 2 0 1 0 1 2 2 0 1 0 1 1 1 2 1 3 0 1 2 1 2 0 2 0 1 2 0 1 0 1 2 1 2 0 I 2 0 0 1 2 2 0 1 1 0 0 0 1 1 0 0 0 0 0 0 0 I 1 1 1 I 1 1 1 1 2 2 2 7. The interactions in the threelevel Table 18. n 18.1 can be obtained from this lattice by assigning A t o column I . 2 2 2 2 2 1 2 2 2 0 0 0 1 I 1 2 2 2 0 0 0 I 1 1 2 2 2 1 2 2 2 1 1 1 2 2 2 0 0 0 2 2 2 0 0 0 1 1 1 0 0 0 1 1 1 2 2 2 2 2 2 0 0 0 1 I 1 t 1 1 2 2 2 0 0 0 0 1 0 1 0 2 0 1 2 0 1 2 0 2 0 1 I 2 1 1 0 1 2 CJ 1 2 0 1 2 0 1 2 0 o 1 2 0 1 2 0 1 2 2 0 I 2 1 2 (1 1 2 0 1 2 0 2 0 1 2 2 0 1 2 2 0 I 1 2 0 I 2 0 1 2 0 I 0 1 2 I 2 0 2 2 0 1 1 0 I 2 2 0 1 1 2 0 1 2 0 2 0 2 0 1 0 2 I 2 0 2 0 1 1 2 0 0 1 2 2 0 1 1 2 0 0 1 2 0 1 2 0 0 2 0 2 0 1 1 I 2 2 0 2 0 0 2 I 2 (J I 2 I 1 1 0 2 L O 2 1 1 1 2 0 2 1 2 0 0 1 2 0 1 2 2 0 1 2 0 1 0 1 1 2 0 0 1 0 1 2 2 2 0 1 .318 DESIGN OF EXPERIMENTS: FACTOPIAL EXPBKIMI!NTS A T SEVERAL LEVELS The interaction sum o f squares is spurious. with three levels each. and D to column 9. in 27 runs. but the sums of squares for A and B are what are needed together with the total.6. and so is the error term. is more complicated than it was in the twolevel case and will not be given here. THIRTEEN FACTORS IN 27 RUNS Earlier.1. which goes back to Fisher.
and four to levcl 2. THE L(36) LATTICE The last of the standard lattices that we present is the L(36) lattice. THE L( 18) LATTICE This is a lattice for 3'// 18. The eighteen points fall into triples. The 36 points in that array form 12 groups of three points each.7.7. at two levels by running the first three of the triples at the low level of H and the other three at the high lcvcl o f H . . We can add a thirteenth factor with three levels to the experimcnt by assigning four groups t o level 0 of the new factor. and thus obtained an orthogonal array for 23313//36.1.lattices form a complicated web.8. four to level I . The L(18) Lattice 0 C 0 1 1 1 2 2 2 0 0 0 J 2 0 1 2 0 1 0 0 0 l 2 1 2 0 2 0 l 2 2 0 1 1 2 0 7.1. l 2 1 2 0 1 2 0 0 I 2 2 0 1 2 0 l 0 l 2 2 0 1 0 1 0 l 2 0 1 2 2 0 0 0 0 0 0 0 0 0 0 2 0 1 1 2 1 0 1 7 0 1 2 0 1 1 2 2 2 1 2 0 I 2 1 2 1 2 0 1 I 2 0 0 2 0 1 1 2 1 2 0 2 0 1 1 2 0 1 1 2 0 1 I 1 1 2 0 2 0 1 1 1 1 1 1 1 18. The first 12 columns of the lattice form an orthogonal array for 3'*. 18. We can add an eighth factor. I € . we have also added two more factors at two levels.7.1. Table 18. and there are no interaction tables for them. We can say nothing about the interactions. An example of the use of this lattice is given in the next chapter. and this latticc can bc used for main effects only. See table 18. which is given in table 18. In the table.8.
2.320 0 1 2 0 1 2 DESlCN OF EXPERIMENTS: FACTORlAL EXPERIMENTS AT SEVERAL LEVELS Table 18. I . It is shown in table 18.8. 223'3//36 0 1 2 0 1 2 0 1 2 0 1 2 0 1 2 1 2 0 2 0 1 2 0 1 2 0 1 0 1 2 1 2 0 0 1 2 0 1 2 2 0 1 1 2 0 0 1 2 1 2 0 2 0 1 2 0 1 0 1 2 1 2 0 0 1 2 1 2 0 1 2 0 2 0 I 1 2 0 1 0 1 2 1 2 0 2 0 1 1 2 0 0 1 2 2 0 1 0 1 2 1 2 0 2 0 1 0 1 2 1 2 0 2 0 1 0 1 2 1 2 0 2 0 1 2 0 1 2 0 1 1 2 0 1 2 2 0 1 0 1 2 2 0 1 1 2 0 0 1 2 0 1 2 2 0 1 2 0 1 1 2 0 1 2 0 0 1 2 2 0 1 1 2 0 1 0 0 0 0 0 0 0 0 0 0 0 1 1 1 1 1 1 0 0 0 0 0 0 0 0 0 1 1 1 0 0 0 1 1 1 0 0 0 1 1 2 0 1 1 2 0 0 0 0 0 0 0 0 0 1 2 0 1 2 0 1 2 0 1 2 0 1 2 0 1 2 0 1 2 0 1 2 2 0 1 0 1 0 1 2 1 2 0 2 0 1 2 0 1 1 2 0 0 1 2 2 0 1 1 2 0 0 1 2 1 2 0 0 1 2 1 2 0 1 2 0 1 2 0 1 2 0 0 1 2 1 2 0 2 0 I 0 1 2 2 0 1 2 0 1 0 1 2 1 2 0 1 2 0 2 0 1 0 I 1 1 1 1 I 2 2 0 1 2 0 1 1 2 0 1 2 0 1 2 0 2 0 1 2 2 0 1 0 1 2 2 0 1 0 1 2 2 0 1 1 2 0 0 1 1 2 0 0 1 2 0 1 2 1 1 1 1 0 10 1 1 1 1 0 0 0 0 0 0 0 1 1 1 0 0 0 1 1 1 1 1 1 0 1 2 0 1 2 0 1 2 0 1 2 2 0 1 2 0 1 2 0 1 2 0 1 I 2 0 1 2 0 2 0 1 0 1 2 0 1 2 1 2 0 2 0 1 1 2 0 0 I 2 2 0 1 1 2 0 1 2 0 1 2 0 0 1 2 0 1 2 2 0 1 2 0 2 0 1 0 1 2 0 1 2 2 2 2 2 2 2 2 2 2 2 2 2 2 0 1 1 2 0 0 1 2 10 1 0 1 0 1 1 1 1 1 1 We could also start with the first 12 columns and add a Plackctt and 12) Burman design.8. The resulting array is an orthogonal design for 2 1 1 3 12//36. One row of the t( lattice is addcd to each group of three points. .
a BoxBehnkcn design.9. We can only . we arc focusing on fractional factorials that are orthogonal arrays. In this section. 1 o 2 I o I o 2 I n I i o I I o o 2 1 2 1 0 0 2 1 2 2 1 n 1 2 n 2 1 1 0 2 0 o 2 1 0 1 0 2 2 1 n i n i i o i i i o o O l O l l 0 l l l 0 0 0 1 0 I 1 0 I 1 1 0 0 o I 1 0 2 0 2 2 2 0 1 1 0 I 2 1 0 0 0 1 2 2 1 2 0 2 0 2 1 1 1 2 0 0 2 0 1 0 1 1 2 0 0 o 0 n 0 o o o 1 o 0 0 I 0 II I 0 1 1 0 1 1 1 I 1 0 I 1 1 1 1 n I 1 1 n o 0 0 I 2 1 n 2 0 1 1 2 1 2 0 0 0 2 1 0 1 2 2 o 2 n i 1 1 0 2 1 2 0 0 1 0 I 2 2 1 2 2 0 1 0 0 2 2 1 1 2 2 o n 1 2 1 I n o 2 2 n n i 1 2 0 2 2 i i o n I I o I I 0 I 0 1 0 1 1 0 I I 0 0 0 1 0 1 1 0 1 1 0 0 n n o o o 1 I o n i 1 1 n I I n I o I 1 ~ ~ 1 o 1 I I I i i ~ 1 0 2 1 1 1 0 0 2 1 2 0 2 I 0 7 2 2 I I 0 2 0 1 0 2 1 0 0 0 2 2 i n I 2 1 1 1 I o o 1 o I I n 1 1 1 0 0 0 1 O I 1 0 1 1 1 0 0 1 1 0 0 0 1 0 1 1 0 1 I I o o o I n i 1 n I 0 1 1 0 0 0 I 0 1 1 0 1 I 0 I 1 0 0 2 0 I 2 1 2 0 1 1 2 0 i o 1 2 n 2 n 1 2 2 n i 2 1 2 0 I 0 1 2 0 0 1 2 o 2 2 2 1 2 1 i o i o o 1 0 0 0 2 0 2 2 1 2 1 1 2 1 1 1 n 1 0 n 2 0 2 2 0 I I I n o n I 0 I 1 0 1 I 1 0 0 0 1 0 1 1 0 1 1 1 0 0 0 1 0 1 1 1 0 1 1 1 0 0 0 1 0 1 1 0 1 1 1 0 o 0'1 0 1 1 0 1 1 1 0 0 0 1 0 1 18. 2"3'*//36 n o o 0 0 0 0 1 I I I 2 2 2 2 1 i i i 2 2 2 2 n n n n 2 2 2 2 n n n 0 1 1 I i 1 1 1 1 1 1 1 1 1 1 1 1 2 2 2 2 2 2 2 2 2 2 2 2 0 0 0 0 c) 0 0 o o o o o o n n n n n n 0 n o n n o o n o 0 0 0 o n o o o o o o o o 0 1 1 0 1 1 1 0 0 0 1 0 ~ o o I 2 2 0 I 2 0 1 1 2 0 1 ) l n o 2 I 2 o 1 2 2 n 1 1 2 I 2 0 1 2 n n i 2 2 0 0 1 1 0 1 1 1 0 0 0 1 0 1 I 0 I I 1 o o n I n i 1 n 1 1 1 o o o 1 0 I 1 I o n n n i 1 0 1 1 i o ~ I 0 1 1 0 1 1 1 u 0 0 1 1 0 1 1 1 0 0 0 1 0 I ~ I I I O O O I O 0 1 2 0 2 1 0 2 2 1 0 1 I 2 0 1 0 2 1 0 0 2 1 2 2 0 1 2 1 n 2 I 1 n 2 n I 0 2 I 2 1 0 2 I 1 1 o 2 1 0 2 0 2 1 0 2 2 3. A NONOKTHOGONAI.8.2. FRACTION Throughout the last three chapters of this book.A NONORTHOGONAL FRACTION 321 Table 18. wc give an example of a nonorthogonal fraction.
parabolas. the response. . pulse.5 13.5 28. The ccntcr of Lhc system is the set of conditions that maximizes.9.27~.1.94xf. and the engineer fits a complete quadratic model: Y = P" + B. the ratio o f gascs uscd. in essence. The levels of the factors are taken as .. 'The design and the observed responses are shown in table 18. power.1 . 2 9 ~. we will look at oxide uniformity.67~: . By fitting thc .1.5 34.24 + a+: + P23X2X1 . or minimizes. y .0 45.~ . Thc BoxBchnkcn designs are.. This example comes from an etching study. The most commonly used dcsigns handle three or four factors.8~. 0.4 18. + P I + . ~1 ~ 0.1 0 +1 +1 1 0 0 1 0 1 1 0 +I 0 0 0 0 +I 41 22.3.1 1 0 0 0 t1 0 1 +1 1 +I +1 0 .4 38.1 .0 12. + .6 43.3 . + 8.x.1 25. which is called response surface methodology. B.0 28. x 3 .q + P Z X Z + P.1 37. Several responses are observed.1 Katio Pulse Power +1 Oxide Uniformity 21. 5 7 + ~ 0 . and hyperbolas).1 . 5 . The contours for given values of y can be a set of concentric quadratic surfaces.9. ~ ~ ~ ~ ~ ~ . and x 2 x 3 . Table 18. 5 0 . 3" factorials without thc vertices of the cube and with several center points. + P13XIXZ + PljXlX3 The reader who rcmcmbcrs analytic geometry will recall that this equation is a threedimensional gcncralization of conic sections (ellipses.24~5 4.0 0 0 The initial regression fit of the model to the data is y = 30. The three factors are A .1.~.0 0 1 +I 0 +I 0 0 1 0 I . and C'.0 25.2 12. + 1. Example 18.9.322 DESIGN OF EXPERIMENTS: FACTORIAL EXPERIMENTS AI SEVERAL LEVELS touch on the area represented by this example.2 ~ 5 7 The only significant terms at a = 10% are x 2 .
3.and so on. The model clearly indicates interaction between thc two rcmaining factors. the regression equation is y = : 27. 0 1 1 . 5 8 ~ ~ 0 . 18. 1 0 1 . They are the points of the L ( 9 ) array. usc Latin letters for the first square and Greek letters for the second. or a saddle surface.~ ~ ~ ~ 1 8 + For this set of data.4 . A Latin Squarc is an arrangement of p letters in a p x p square in such a way that each letter appears exactly once in each row and each column.55 50. ) .r2 and x . and. 2 0 2. the columns represent the levels of the second factor. It is an orthogonal array for three factors. and R instead of our usual A . We dcnote the factors by P. 2 2 1. following tradition. Thus. The points in the first row are at level 0 of P. The two squares are said to be orthogonal if. 0 2 2 . when we put one on top o thc f . The rows represent the levels of the first factor. the nine points are 0 0 0 . the gas ratio is not important. and C because we are using a. and c as the letters in the square.1.31. 2 1 0.55 24.1 1. and the model reduces to a quadratic in . which is a hyperbola. For p = 3. R . 1 2 0 . &.63 20. the points in the first column are at level 0 o f Q . table 18. The predicted responses at the four “corners. 5 8 .10. with the last factor omitted. 1 1 2 . and the letters the levels of the third factor. each at p levels.01 .GRAECOLATIN SQUARES 323 reduced model. wc have a square a h c b c u c a b This is a fraction for 3’ in nine points. LATIN SQUARES XI 1 +1 1 x 2 +I Estimate 12. h .” all of which are outside the arca in which the observations wcre taken. arc as follows: I 1 +1 +I 18. 4 6 ~ ~ . GRAECOLATIN SQUARES Suppose that we take two Latin squares of side p .
LEVELS other. 1 1 2 0 . each Latin letter appears exactly once with each Greek letter. 18.324 DESlGN OF EXPERIMENTS: FACTORIAL EXPERIMENTS AT SEVERAI. 7.3. 0 1 I 1. the columns to the second. Orthogonal Squares with p = 4 __ P a b c d i h a c i c d b a d c b a U Y G Y S 1 3 S Y a U P S P U v A D B C ' B C A D C H D A D A C B .2 is the orthogonal array for five factom. which is the lattice of table 18. 1 I .13. the Latin letters to the third. in 16 runs that we derive from the three squares. wc get the GraecoLatin square. 1 0 1 2 . . a a bIj by c a c/3 uy ba up cy We can now make an orthogonal array for four factors by letting the rows correspond to the levels of the first factor.. Algebraists have proved that one can always find a set of p .1 mutually orthogonal squares when p is a primc number (2. one cannot even find two orthogonal Latin squares of side six.).EVELS I 16 RUNS N Table 18. On the other hand. 0 2 2 2.13. 2 1 0 2 . 3.1. in turn. . Table 18. For example. and the Greek letters to the fourth. 2 2 1 0 . 'The square that we have just shown provides a fraction for 34 in nine points: 0 0 0 0.1 shows four mutually orthogonal Latin squares of side four. . 1 2 0 1 .13. each at four levels. with p = 3. 18.) or a power of a prime number (4 = 22. This. provides an orthogonal array for five or more factors. 5 .1.ATLl SQUARE3 If we can find three or more squares that are orthogonal to each other. we can superimpose them to form a hyperGraecoLatin square. and so no GraecoLatin square of side six exists. HYPERGRAECOI. 2 0 2 1 . each at p levels. We f will give an example o a square of side four that can bc used for a 4* factorial in 16 runs. 8 = 23. and of an array for 5h in 25 runs. . FIVE FACTORS AT FOUR 1. we can take the two squares u b c b c a c u b and f f P Y y CY p I j Y a Superimposing the squares. Table 18.13.12. .
28 15.16 24.08 ss 2. For those reasons.2.13 21.74 22.8) 23. Those who use all five factors argue that at the screening stage.14.86 3.98 23.13. FOURLEVEL FACTORS AND TWOLEVEL FACTORS A factor with four levels can be accommodated in a 2" factorial.10 21. they are associated with two of the .4. The price of accomniodating five factors is that all 15 d.68 0.15 26.03 MS It is obvious that D is a major factor and that A is also worthy of attention.43 22. The observations are written in the same order as table 18. Analysis of Variance SOURCE A D E c B DF 3 3 3 3 3 7.11 2.09 5.13. for the error term in the analysis of variance.f. The analysis of variance table is table 18.13.58 23. and that no formal test of significance is nccessary at this stage.37 0. and there is no error term for F tests.Cr) 22.3.54 23.3.13. There are three d.f. are used for the factors.10 Table 18.89 0. Data 20.27 22. so that they may be the targets of further investigation.87 21. This is it matter of choice. Table 18.29 1. 18. some statisticians suggest that this design should bc used with only three factors so that there remain six d.4.FOURLEVEL FACTORS AND TWOLEVEL FACTORS 325 Table 18. for the fourlevel factor.13. 4' in 16 Runs A B C D E 0 1 2 3 0 0 0 0 0 1 2 3 0 0 A B C D E 0 1 2 3 1 1 1 1 1 0 3 2 1 3 2 0 1 2 0 3 A B C D E 0 1 2 3 2 2 2 2 2 3 0 1 2 0 1 3 2 1 3 0 A B C D E 0 1 2 3 3 3 3 3 3 2 1 0 3 1 0 2 3 0 2 1 2 3 3 1 1 2 The data in the example are shown in table 18.f. the purpose of the experiment is merely to detect the one or two factors that clearly stand out as important.2.13.97 22.05 23.
C D : 13. ABCD . A C D . Exuniple M. I1 bccomcs the A contrast. but we can handle five fourlevel factors in 16 runs. A D . in the L( 16) latticc. Table 18.14. their interaction is in column 3.2. we assign A . 11. C. R .14. I 2 2 1 1 4 2 2 2 I 2 2 2 1 1 2 2 I 1 2 2 1 1 1 2 2 1 2 I I 2 1 . 2. respectively.14. a .1. and 8. respectively. BC. The combination 1 1 in the first two columns of L ( 8 ) (and. as shown in table 18. we write the levels of the fourlevel factor by 1. 1 in column 3) hecomcs level 0. 9. 12.14. Contrasts for a FourLevel Factor 1 2 +1 3 1 +1 1 I1 111 4 tl +1 +1 1 1 t l 1 1 1 twolcvel factors and their interaction.1. as we saw in the previous section. A B . 3. the remaining columns will represent: 3. We can follow chapter sixteen and associatc the three degrees of freedom with orthogonal contrasts. We take the L ( 8 ) array and assign A and B to columns I and 3.326 DFSIGN OF EXPERIMENTS: FACTORlAL EXPERlMENTS A 1 SEVERAL LEVELS Table 18. ab. ARC. B D . 2. The other columns are available for the twolevel factors.2.2 with the fourlevel factor in the first column. and 111 the AB contrast.14. conscqucntly. BCD.2.1. A C .2. [I Example IX. We cannot extcnd this idea to handle two factors at four levels in eight runs. 6 . 3.'. A B D . and D to columns 1. 2 1 becomes I. If we replace 1. If. The array is shown in table 18.14. The first examplc is an orthogonal maineffects array for five factors with four at two levels each and the fifth at four levels. 15. 7. 2. 4. 5 . contrast I beconies the R contrast. We demonstrate the rationale in the next paragraph ilnd follow that by two examples. and 4. 14. 1 2 beconies 2: and 2 2 beconies 3. To be consistent with the notation used in the L ( 8 ) lattice in table 17. t44' in Eight Runs I I 2 2 3 3 4 1 10. and 4 by (1) b.
A C D ) The array is shown in tablc 18.3. 12 ( C . 0 3 4 1 2 1 2 4 3 3 5 . C‘U) 5. It is left to thc readcr to till in the details of the dcrivation. S 1 3 2 4 1 3 5 3 4 1 2 3 4 5 2 3 5 4 1 3 3 3 3 3 1 3 2 5 4 1 7 1 8 1 9 20. 2 15. 5 3 2 4 3 2 2 2 5 4 3 3 5 1 4 5 5 5 4 5 4 1 3 2 . .13. 5 4 . . ‘4 2 . 3 1 5 5 2 4 2 3 1 4 4 2 5 3 21 2 2 2 3 2 4 25. . 5 3 . BD. 1. 8. 2. . 2 4 1 3 5 3 4 5 1 1 I 2 3 4 5 3 5 4 1 2 I 2 1 s 1 4 1 1 1 3 4 4 4 4 4 4 1 3 2 s 6 7 8 9 1 . 14 ( A R C . . . 10.2. A D . 2 .15. and 3 givcs thc array in table 18. 1 2 5 1 5 1 4 1 2 3 5 1 2 2 2 2 2 5 3 2 5 4 1 s 1 1 1 1 1 1 2 3 4 5 .14. . . 13 (BC‘.3. ABCL)) 7. 9.FOURLEVEL FACTORS A N D IWOI. .14.1 1. 4’ in the I. I). 3 5 . . A B D . 15 ( A C . . . B .EVEL FACTORS 327 S Table 18. A B ) 4.( 16) Array 1’ 1 I P 1 2 3 4 1 R 1 2 3 4 2 1 3 4 3 4 1 T 1 1 1 2 2 2 2 3 3 3 3 4 4 4 4 2 3 4 1 2 3 4 1 2 3 4 2 4 3 2 1 1 2 3 4 3 4 1 2 4 3 2 1 2 1 2 3 4 4 3 2 1 3 1 4 3 3 4 1 4 3 2 We assign the five factors at four levels to the columns in the following way: S P Q R T 1. 11. Changing the levels t o 0. 1 4 . R C D ) 1 . 3 ( A . 0 Table 18. 2.
7 281.6 278.5 349.1. The authors were interested in the ratio of Col. A .9 204.1 361.1.9 250.3 153.6 235. B.0 286. and C . In a paper.3 154.8 227.15. In example 18.13 can also be used to construct an orthogonal array for six factors. The first column gives the approximate value of the stress. EXERCISES 18. Carry out the analysis of variance. each at five levels.8 187.3 176.6 231.328 DESIGN OF EXPERIMENTS: FACTORJAL EXPERIMENTS A T SEVERAL LEVELS 18.1.5 201.4 206. I 251.6 159.7 216.8 230.8 250. Oikawa and T.5 278. I 246.2. Such an array is shown in table 18.5 185.7 198. I .8 177.5 295.5 294.0 169.1 245.0 201.3.8 267.5 232.15. 10.5 269.2 1 1 1 1 1 1 1 1 1 2 2 2 2 2 2 2 2 2 3 3 3 3 3 3 3 3 3 1 1 1 2 2 2 3 3 3 1 1 1 2 2 2 3 3 3 1 1 1 2 2 2 3 3 3 1 2 3 1 2 3 1 2 3 1 2 3 1 2 3 1 2 3 1 2 3 1 2 3 1 2 3 .3 175.1 369. T.7 213.8 167. 191. in 25 runs. are given in the last three columns.3 236.5 234.H 170. Oka (1987) report the use of a 3 x 3 x 3 factorial to investigatc differences between two mcthods of estimating membrane stress.6 154. FACTORS AT FIVE LEVELS The method of section 18.7 171.1 217.0 208.6 293.6 369.1 235.0 174. we analyzed the response ctch rate in the cxperiment by Yin and Jillie. the second column gives the value from a more detailed analysis.6 304.3 158. An application appears in thc exercises.0 152. The levels of the three factors. 2/Col. uniformity and selectivity.5 186. Repeat the analysis for the other two responscs.
x* 0 1 +I 0 0 +1 0 0 1 0 +I 1 0 1 0 1 *. 1 +1 1 x.ItXERCISES 329 18. t1 I 1 0 x3 36 12 14 21 y 6. This an important point about the design. the first factor is the power in Mtorr at levels 10. 4. 7. This is an example of a BoxBenkhen design with three factors and three center points for a total of 15. 3. X z . X.3. predicted values. . and residuals. I. Fit the model There arc substantial tvalues for each of the predictors and a largc value of H2. This is overkill. Print out the observed values. 9. 20. be due to a high level of noise (note the value of s). x. perhaps. x.) Fit the complete quadratic model to the data in exercise 18. and X 3 arc uncorrelated. 0 1 0 0 +I 0 0 7 20 68 y 35 26 11. 8. 15. 12. which might lead you to conclude that the model gave a very good fit to the data.. Show that X. 0 +1 +1 1 0 +I +1 1 x 3 28 0 10. Thus.3. o 0 0 +1 y 16 48 +I 0 43 68 30 The values taken by the X variables arc coded values for the levels taken by the factors. It is adapted from an experiment carried out in the development of semiconductors. 2. x. Continue this procedure until you have a model in which every coefficient has a significant rvalue.4. 5. therc are several terms whose coefficients have small rvalues. x.3. 13. (Continuation of exercise 18. 18. = (power . Drop one of the two terms with the smallest tvalue and rerun the regression.lO)/lO. 15. 14. There arc a few residuals that seem to be largebut that could.
076 1. 18.64 81.150 0. Notice that there is a marked reduction in the size o f the residuals in the second fit.14.I . They are related to coniponents of the circuit. (1987).066 8. The array that was used is the first five columns of the array in table 18.850 0. Which are the important factors? Which levels of those factors give the largest voltage? 18.03 85. The response is the outlet tempcraturc minus 360 degrees. Y.391 2.15.97 Y A 1 B 2 2 3 2 1 3 2 C 1 1 2 2 Y 58. Notice that the estimated standard deviation.3 and compare the fit to the results of exercise 18.685 0.25 19.464 1. The observations in the order that is given in the table are as follows: 0.3 3 DESIGN OF EXPERIMENTS: FACTORIAL EXPERIMENTS AS SEVERAL LEVELS 18. Work out the details of the derivation of the lattice array for 4' in table 18. The experiment involves a heat exchanger.085 2.5.SW 0. 18.005 5.164 6.423 4.189 0.326 0.4). The response.326 3. (1987) also report the data for an L ( 9 ) array for three factors.522 3. would the cnginccr bc justified in recommending that A should bc at level 3 and that the levels of the other factors are immaterial'? .18 67.200 3. There arc five factors.471 1.3.085 2.6.71 19.64? Would your conclusions be changed if you used as the response In( Y ) or CY'? If the objective was to minimize Y. 125.8.917 0. Fit the model to the data of exercise 18.248 1.254.704 1. (Continuation of exercise 18.7.90 125. The following data set comes from a simulation study of an output transformerless (OTL) pushpull circuit rcported by Wu et al.153 7. Compare the predicted values and the residuals in the two fits. s. particularly for the two points with the highest responses ( y = 68).82 Which arc t h e important factors7 Do they exhibit linear or quadratic trends? Are these conclusions dominated by the single large response. has dropped to less than half its earlier value.3. A 1 1 2 3 3 B I 3 2 3 1 C 2 3 3 3 I 54. Wu et al. is the midpoint voltage of the system. each at five levels.177 1.78 14.
An example of these designs was given in the previous chapter. This is another area in which statistics has played a major role. under thc guidance of Deming. attention has usually been focused on a response such as the yield.Statistical Methods in Engineering and Quality Assurance Peter W. The agronomist wants to compare different varieties of wheat and different levels of fertilizer with the idea of optimizing thc yield: can we develop a hybrid that will yield more bushels per acre. workcd hard at solving it. Taguchi. Inc CHAPTER NINETEEN Taguchi and Orthogonal Arrays 19.1. This is an area that was not o f major conccrn in thc West until recently. He has been mcntioned several times in previous chapters. In earlier chapters. however. recognized the problem after World War 11 and. our attention has been attracted to a third fieldmanufacturing engineeringin which the objective is not t o squeeze thc extra bushel or the extra octane number. One of the leaders in the field is Japanese engineer and statistician G. espccially the series of experimental dcsigns introduced by Box and his student D. M. This is the topic of response surface methods. ‘The Japanese. Much of this chapter is based on his contributions. which were developed by Box and others in the 1950s and 1960s. Sequences o f welldesigned experiments lead the engineer to the operating conditions that optimize the yield and also map the region around that inaximum so that the opcrator can adjust for such departures from optimum as the gradual loss of potency of a catalyst. Many engineers are now using thcse methods. John Copyright 0 1990 by John Wiley & Sons. INTRODUCTION In the traditional strategy of experimentation in engineering. but to produce thousands and thousands of widgets that are as close to somc specifications as possible. we painted a picture of the old days when the engineer had a process rolling along smoothly in a state of statistical control. In the last decade. 331 . agriculture. The chemical engineer wants to know which levels of pressure and tempcraturc will optimizc thc amount of product coming out of the reactor. Behnken. and other sciences. or can we improve our husbandry to coax a higher yield from our land? They have had enormous succcss in achieving their aims. H.
k. You carziiot inspect quality into a process. The other player is not content with the criterion of being cither on o r off the green. She takes thc hole itself for her target and aims for the flag with the criterion of getting as close to the hole as possible. we would like to refine our manufacturing process to have zero bias and the smallest possible variance. the inspectors cannot alter that figure. build quality by designing it into the production process. but to be within spccification. So. we mean carrying out researchanddevelopment experiments to find the best operating conditions for manufacturing not just one item but thousands. By their diligence.and the variance. Although we cannot improve quality by inspection. they can reduce the percentage o f substandard widgets that leave the plant. what can you do? You can begin by rejecting the old simplistic notion that to be outside specification is bad. and the target value. but they cannot directly improve quality. Replace it by the new ideas of process capability that were discussed in section 9. in theory.332 TAGUCHI A N D ORTHOGONAL ARRAYS between the engineer and thc factory gate was a squad of inspectors who. The first is to remind the reader that even though a process can be rolling along smoothly under control. We want to control both of them. is satisfactory. The mean square error is the sum of two parts: thc squarc of the bias. and she works continually to reduce it even furthcr. we can. which is the diffcrcncc between the actual process mean. When we talk about designing quality into the process. which is to take their prototype and make it manufacturable. within its natural limits. or robust design. putting. The sample median can be prefcrrcd to the sample mcan as an estimator of E( Y) because it is robust against cxtremc observations (outliers). One of the major contributions of the Japanese approach is to make the engineer think in terms of aiming for a target value.h ' One might think of two golfers ). Too often the word design is used in a restricted sense to mean the work of a group of specialist engineers.. who design new items and perhaps build a single prototype model. i. If you are producing 20% defective widgets. those limits may bear little relationship to the actual specifications. One aims for the green: she either meets the specification of being on the green. Statisticians use the term robust to mean insensitive to departurcs from set conditions.25 in connection with the criteria C. her mean square error becomes smaller. This is called offline quality control. po. Taguchi calls this the system design phase. All they can do is decrease the fraction of that 20% that goes out to the customer. and work to reduce the mean square error. E( y . sometimes far away physically and intellectually from the shop floor.e. po. playing iron shots from the fairway. The second is that the inspectors can spot most of the bad batches and turn them back for rework or for scrap. Wc arc more concerned with thc next step. Wilcoxon's nonparametric rank test can be preferred to the . as her game improves. spotted all the bad widgets and allowed only good ones to leave the plant. or she misses the green and is out of specification. in which case she is rcady to go into the next stage of production.. doing rework in the rough. Two things must be said at this stage. even by only a whisker. p . and C6. and must.
We discuss parameter design in the next sections. for dispersion. which we mentioned earlier. for 2 location. 19. 19. and tolcrancc design. which we call the dcsign factors. In the final stage. and some other statistics.4. The Taguchi approach to robust design makes thc engineer focus on rcducing the variability of the process. The engineer talks about y as the responsc. One of the variables that controls the output is the input ratio at a certain point in the circuit. s . and the sample variance. The process involves several factors. A STRATEGY FOR PARAMETER DESIGN Suppose that we are manufacturing an electric circuit and that we are intcrcsted in the output of a certain stage of the circuit.A STRATEGY FOR PARAMETER DESIGN 333 twosample ttest because it is insensitive to departures from the assumption of a normal distribution. which are advocatcd by Taguchi. 19. the variable that is being observed is often called a performance characteristic. we conduct experiments to determine the “best” nominal settings (or dcsign parameters) for these factors. y. . The two tnost commonly used performance characteristics for a sample of data are the sample mean. THE THREE STAGES OF PROCESS DESIGN Taguchi suggests that there are three stages of dcsign for a process: system dcsign. We may also use transformations of them to measure the dispersion. parameter design. called signaltonoise ratios. the engineer finds the tolerances around the nominal settings that were dctcrmined by thc parameter design.2. Wc may learn that some of the parts used in the process require very closc tolerances. Thcrefore we may save cost by using in the second group cheaper parts than wc had originally projected. tolerance dcsign. The agronomist usually talks about the yield of a plot. He advocates a systcmatic approach to achieve that aim. In the prescnt context. In the parameter design stage. while others are not so demanding. such as ln(s2).3. TERMINOLOGY The generic observation has different labels in different fields. Taguchi has led the way in making engineers think of carrying out experiments in the manufacturing process to make the product robust to variation both in the parts that are used in the manufacture and in the environmental conditions in which thc item is to be employed.
'I'he target value. corresponds to a ratio. Plot of output vs. the output is increased to v : but the sensitivity of y to variation in the ratio is decreased because of the nonlinearity of the response curve. Ratio Figure 19. but not the variance. This suggests a design strategy for parameter design experiments. point.4.1 shows a plot of output against the ratio. that we can vary to adjust the output back to y. and thcn usc the others to adjust the performance characteristic to its target value. Taguchi makes extensive use of orthogonal arrays. We will perform a factorial cxperinient to investigate the influence of these factors. xu = 13. changing thc levcls of some other factors affects the variance.1. such as a capacitor. we will learn that changing the levels of some factors affects the process mean.. without losing the advantagc of lower sensitivity to small changes in the ratio that wc discovered at x = 20.4.334 'IAGUCHI A N D ORTHOGONAL ARRAYS Figure 19. As part of this strategy. Perhaps we can find another part in the circuit. If the ratio is increased to x = 20. From it. We should use the latter group to reduce the variance a s much as wc can.5. As the ratio varies about the value xo. y o = 55. ratio. the slope of the curve is quite steep at that 60. the output varies about y o . .
6.5.6 1N.0 71. and then adjust the levels o f A and B to get the desired value of the process mean. Suppose that we have three factors in our process and that we decide to look at two levels of each.2 50. i.7 76. We can aniilyze the data in two ways. First. wc take as the response the means of the sets of five observations. the best levels o those factors. or dispersion. 19. because il is morc nearly normally distributed..0 65. increasing R increases 7. That tells us which factors have a major influence on the location o f the process.8 50.0 92.1 64. on the other hand. We start with a simple example that illustrates the basic principlcs.5.0 60.2 40. we used the sample variance as the measure o f variability. ' s rather than s2 itself.5 91. the average output voltage.3 91. and as the measure o the process f Table 19.5 46. we takc as the perforrnancc charactcristic the estimated variance at each of the eight sets of conditions. Many statisticians prefer to use the natural logarithm of . The data are shown in table 19.4 .7 46.F.g. We find that A and I1 are the important factors. is modest. We learn that changing the levels of A OJ B has little effect on the variance.6 70.0 01.2 76.8 47.1. The effect of changing the level of C. e. In this example.. SIGNALTONOISE RATIOS In the cxamplc of the previous section.2 01.0 65.e.S 76. In the second analysis.1 71.SIGNAJ.5 78.TONOISE RATIOS 335 19. or parameters.5 91. The combined analyses suggest that we should run the process at the low level o f <: to cut down on thc variability.5.3 717 66.5 88. changing the level of C produces a significant change in the variance. though positive. we obtain the same result with either choice.5 47. Increasing '4 decreases . which is greater at the high level of C.0 60. PARAMETER DESIGN EXPERIMENTS The purpose of the parameter design experiment is to determine which of thc factors are important in the manufacturing process and to find the f optimum set o f working conditions.7 92.3 49. We run the complete factorial (all eight points) and wc make four runs at each of the eight sets of conditions.1 A ~~ 13 c 1 1 1 +I 1 +I 1 1 +1 +I 1 I 1 1 +I I +1 1 +I +I +I +I +I +I 60.
ARRAYS mean. The last one. Why we should be considering the variance of log( y) rather than looking at 7 or log(?) is not clear..V. but not as a gencral procedure. which he cls signaltoal noise ratios. is a transformation of the mean square error when the targct value is zero. It can be argued that the first of them. The use of the signaltonoise ratios has been a matter o f dispute. the objective of the engineer is to optimize the process by choosing the set of working conditionsthc levels of the experimental factorsthat makes the signaltonoise ratio as large as possible.. SN. It is clearly a transformation of the wellknown coefficicnt of variation. In all cases. (ii) when the objective is to make y as large as possible. the second is related to the mean square error for l l y when its target value is zero.. It can be shown that this ratio is mathematically equiviilent to the logarithm of the variance of log( j ) . It has been said that he has developed over 70 different such signaltonoise measures! The three most commonly used of Taguchi’s signaltonoise ratios are these: SN. = 10 log. however.” “larger is better. is more difficult to justify. For a set of data C. That is why somc of them are prefaced with minus signs.336 TAGUCHI AND ORTHOGONAI. They are given the slogans “smaller is better.  (y) 1 where n is the number of points involved. (iii) when y is to attain a specific target value. Pushing that a bit further. .” and “on target is best. I’aguchi advocates several other measures. It may well be more appropriate in somc cases.’’ The fact that the ratios are expressed in terms of log. rather than natural logarithms is not important. = S Y’ * it is often expressed as a percentage. which is the ratio of thc standard deviation o f the response to its mean. when the objective is to make y as small as possible.
S. They are the knobs on the black box that is our process. Suppose that thcrc are N vertices in the inner array and n obscrvations in each outer array. control. but we do not necessarily want to have Y as largc as possible. and 120 ohms. each at three levels. Therefore. With three design factors at two levels each. We obtained our values o f the variance by making four observations at each of the eight vertices of the cube. each at two levels. and the overall variance of the n points. The first step in planning the experiment is to select thc design factors and to choose their levels. which exceeds most budgets. this time in the noise factors. arid it has also been the subject of articles by Box and othcrs. With more factors and niorc levcls. Examples of thcsc arc the temperature in the work place. the total number of possibilities is only eight. There were only three design factors. we want small values o f s. contains no mention o f the target value. The design factors are factors that we can controlmore or less. However. We can also calculate one or more signaltonoise ratios. We could also treat as a noise factor the resistance of the component that was mentioned earlier by perturbing it by a few ohms above and below its nominal values. Certainly. INNER AND OUTER AKKAYS The example in section 195 was simple in two ways. .’rogre. the expcrimcntal design is more complicated. We consider two types of factor: the design factors that we have already discusscd and noise lactors. This can turn out to be an awful lot o f observations! At each vertex. with six design factors. 100 ohms. the total number of possibilities is 729. .J. the humidity. In a letter to the editor o f Qudiry i. The analysis proceeds in two steps. representing environmental noise. The outer array is another orthogonal array. s:. SN.Ss. we calculate two performance characteristics o f the outer array: the average observation. which makes it fiard to justify as it measure o f meeting that target. giving a total of Nn observations in the complete cxpcriment. and so we were able to run the complctc factorial experiment. and the pollen count.s.7. or do not choose to. These are the noise factors. That is the sort of thing that might happen if we used cheaper parts. 19. There are other factors that affect the performirncc of thc final product that we either cannot. The overall experiment consists of running the outer array once at each point (vertex) of the inner array. we choose a fractional factorialone o f the orthogonal arrays. perhaps the L( 18) lattice for six factors in 18 runs. We can decide that fsctor A will be a certain resistor with three levels: 80 ohms. This lattice for the design factors is called the inner array. Hunter (1987) has presented a strongly written and convincing argument against using this criterion.INNER AND OUTER ARRAYS 337 Furthermore. Maximizing the quantity Y’ls’ places a premium on points with large values of j iind small values o f s . j j .
9526 22.9163 22. The outer arrays consist of nine points each. we can use analysis of variance. Because the array is orthogonal.1.8. but do not alter the dispersion.67 42.08 11. and use the adjustment factors t o get j back on target.8302 8.(i 2ld8 11.21 45.3058 23. We note that the linear contrasts for A and both the linear and quadratic contrasts for B are Table 19.1 shows the design of the N.07 39.3533 10.91 45.. s’. s f . The factors are assigned to the first five columns of the lattice. which we can call the signal factors.47 46.7598 23. inner array and the summary statistics.b5 41.6859 11.5. for thc outer arrays carried out at cach of the 18 vertices.9051 23.1767 7.82 39. following the pattern o f exarriplc 18.1639 9.28 46.4048 22. f sl 0 0 1 I 2 2 2 1 2 0 1 2 0 2 0 2 0 1 2 0 1 1 1 0 0 0 1 1 2 1 2 0 2 0 I 2 0 1 0 1 1 2 2 0 1 1 0 2 2 0 1 2 0 2 0 1 2 2 2 1 1 1 1 0 1 2 39.8. If that is so.51 43.4675 21. wc will put thc dispersion factors at thc lcvcls that minimizc thc variability.338 TAGUCHI A N D ORTHOGONAL ARRAYS Now wc havc the inner array with two performance characteristics observed at each vertex. and S .l .37 39. 9.2546 22.8258 SN .6673 8.15 46.4237 8.3950 9.and the third with E and one of the others.7558 23.77 42.8870 10. just as in section 19.7.0s 46. and we make an analysis for each.9026 22.8.5.67 10. that affect y. wc will find some factors.o n c with factors A and B .8245 22.8668 10.3092 22.( 18) lattice describcd in section 18.80 45.30 2 0 1 2 2 1 2 0 I 0 2 2 0 1 40. The L(18) Lattice for the Inner Array A 0 B 0 C 0 D 0 E 0 1 2 1 2 0 __. but do not affect the location.6670 9.7382 23. Wc begin with an analysis of the data for the response j . The easiest way is to f make three runs on a twoway analysis o variance program.5696 1.3SM 21.3150 22. AN EXAMPLE OF A PARAMETER DESIGN EXPERIMENT We consider an cxample in which there are five factors that are possible control factors.82 43.1985 9. 19. and some other factors that control the dispersion. Tablc 19.8275 7.1. the second with C and D.7165 24. It is decided that these factors be considered at three levels cach and that the inner array consist of the I.8737 23.3032 7.5976 21.8222 Y. I f naturc is kind.
If the objective is to maximize y.7 43. namely.73 9. OJ its orthogonal fractions. s?.5 43. one can come to the same conclusions from looking at a table of the averages at the various levels. possibly. to a lesser extent. but that Table 19. table 19. The table of averages of In($’). to a certain extent. E arc signal factors. Averages for j 0 1 2 45.78 9. and that the linear component of E would be significant if we took a = 10%.2 that factors A and B.8 43.8 43.7 43.9 4.AN EXAMPLE OF A PAHAMElEK DESIGN EXPEHIMEN’T 339 Table 19.5 43. The reason for this is that the use of the factorial experiment. E have a big influence upon j .9 43.8.47 9.09 9.16 0.7 42. Analyzing the variance response.92 0.70 8. Averages for sf A B c E I) 9. A = 2 and B = I .8 42. The use of a signaltonoise ratio like S N .2.8. A reasonable set of recommendations would be to run the process at the middle level of D. makes life sinipler for the engineer because there is only one criterion to be considcrcd. ‘Table 19.85 .0. It is not always really necessary to carry out the analysis of variance procedure.65 0.and to adjust the level of j by altering A and B.77 9.8 44.71 0. One chooses the “best” vitlucs for each o f the factors A and B. and.2.8. The averages are shown in table 19.89 9.4. and.3. and Leaves the choice of the level of C to other considerations. R . the linear component o f E are important for controlling the variability. This tells us that A .8.5 significant at the 5% level. It is clear from an examination of table 19.10 9. We already saw that in the more obvious cases.3.8.1 n B c I) E 40. enables us to estimate effects by averages. If the tar@ value for p is about 43.7 41. shows that factor D and.32 11. we should certainly move to the middle level of E because that brings us near t o the target with a further reduction in variability. the decision process is easier. togcthcr with D = E = 1.9 42.8.24 9. points in the same direction. It is clear that one should try to operate at the middle level of I ) .
In deciding which of the factors should bc varied to minimize the variance and which should be varied to adjust thc response.98 23. We should run the outcr array at the set of values for the factors in the inner array that we are going to recomrncnd.340 .266 2.9.270 2.4. we relied on an orthogonal arrayan orthogonal maincffccts plan.287 2.202 simplification may not always be realistic.163 2.for each factor. Perhaps we need to expand our inncr array t o make it a resolution IV fraction. we will.74 23. In this example.227 2.8.266 2.94 22. Averages of In(sz) 0 A B 1 2 2.8. the result is a set of operating conditions that appear to be the best. I TAGUCHI A N D ORTHOGONAL ARRAYS Table 19. 0 I 22.37 22.282 c : E 1) 2. they are A = 2 and B = C=D =E1. We could decide that the optimum conditions for our process are those that maximize S N .06 22.286 2. They are the bcst in thc sensc that if we run the process at those conditions. we sssumed that there were no twofactor interactions. or a resolution V Table 19. we hope.37 22. THE CONFIRMATION EXPERIMENT When all this has been said and done. The results could be disastrous.00 22. are given in table 19. manufacture product that is close to the target with a minimum of variability.5s 22.21 22.202 2. in which case we have to stop and think (and be thankful that we did not make a rash rccomrnendation that would stand as a memorial to us).263 2.18 2 23.405 2.219 2.262 3.5. But we could have been wrong! This leads us to the last stage of the experimcntal plan for parameter designthe confirmation experiment. We need some more data.8. But there may be t i pitfall.96 A B C ri E 22.241 2.65 22.88 23. Averages for SN. The averages for the ratio SN.66 . 19.37 23.5. In the analysis.
but we might just as well continue to use lowcost parts for A . Until about 60 years ago. and not with precision parts in a white coat laboratory setting. we are going t o have to use bcttcr parts in some places. hence. If the outer array on our confirmation experiment has shown a satisfactory level of variability. and the outer array at the confirmation run provides some data about the area around that set of conditions where we will be working. we find how much of the variability is associated with each factor. This time the noise factors are noise in the parts themselves. Not only will improving the precision of A not help us much. we may find that “wiggling” the level of A does not contribute much to the variability of the final product. relatively speaking. TOLERANCE DESIGN Having settled on the parameter values of our process. If factor A is a resistor at a nominal level of 100 ohms. and save a lot of money. in the noise array. 100 ir 2. we can go into production with the lowcost parts and a corresponding savings in production costs. a factor with levels. We run a second outer array centered on the optimum parameter conditions. we have given a short introduction to the design of experiments. Taguchi recommends that the parameter design experiment be done with lowpricc components that have. which has its strengths and its weaknesses. 19.FINALE 341 fraction. we have focused on the approach of Taguchi. we turn to tolerances. The results could confirm our earlier conclusions. In this chaptcr. When the array has been run. and. somc variability. on the other hand. whcre we have to make our further improvements. and calls for some serious thinking and statistical planning. experimenters were taught to change one variable at a time. That happened in the example discussed in section 18. This classical approach to experimentation has two . This can become costly in terms of both money and time.11. If the variability is too high.3. resistors for that itcm.10. it becomes. by the nature of their prices. FINALE In the last three chapters. But which parts need to be improved? A tolerance design experiment is an experiment to find out where the variability occurs and where the adjustments should be made. We may find that A is critical and that the best way to reduce variability is to use very prccisc. 19. We also have a basis for the next stage of the investigationtolerance analysis. and morc costly. We have our optimum conditions. say.
The strength of thesc designs lies in their efficiency and their ability to detect interactions. were using them in chemical enginccring research. but now the A effect is estimated by “16 + uc . Now that practicing engineers are aware. The z3 half replicate also has four points. Davies) first published in 1954. the engineer runs the base point. particularly George Box. and bc. ‘I’hc growth in the usc of these ideas in engineering in the United States was slow. The classical oneatatime design consists of four points: ( l ) .( I ) with variance 2u2. and the statistical indcpcndcnce o f the estimates were achieved by considering. unlike the Japanese. the estimates in the oneatatime experiment are correlated. or doubling of efficiency. ARRAYS drawbacks: inefficiency and the inability to detect interactions. some of the more zcalous disciples of ’I’aguchi get overheated and proclaim that all the ideas of multifactor experimentation arc Japanese. the estimates of A . it was the standard book from which most industrial statisticians learned about design of experiments. particularly in the scmiconductor industry. we have two choices. and C in the factorial experiment arc uncorrelated. Rccausc the designs arc orthogonal. It is still difficult to find room for statistics coiirses in most engineering curricula. mainly inplant short courses. and then investigates A by keeping B and C fixed and varying A alone. also at Rothamstcd. With three factors at two levels and four runs. 6. we arc able t o compare avcragesthe average of all the points at high A minus the uvernge of all the points at low A . For a long time. L. At the end o f World War 11.hc . Unfortunately. The etficiency comes from the balance.(1)]/2 with variance m ? : instead of one point at high A minus one point at low A . The mcthods that they were using were gathered in a book (&. The halving of the variance. seveIal factors simultaneously. we have the average of two at high A and two at low A.342 TAOUCHI AND ORTHOGONAI. Fractional factorials were developed by Finncy. edited by 0 . was developed at Rothamsted Agricultural Station by Fisher and Yates in the 1930s. ( l ) . It is still very good reading for both engineers and statisticians. The estimate o f the A effect is u . during World War 11. of the importance of statistics. as we know it. Many of them are hearing about these things for the first time in the framework of the ‘I’aguchi methodology. ac. Furthermore. The main reason has been that. American engineers are learning about designed experiments through programs arranged by thcir companies. and were invented by Taguchi ’ .sign und Anulysis of Industrial Experiments. u. The ideas of inultifactor experiments took a while to move from agriculture to engineering. it followed and built upon Fisher’s carlier introduction of the procedure for the analysis of variance. nb. and c . and varying. This property is sometimes called the hidden replication property of factorial designs. scientists at thc laboratories of Imperial Chemicals Industries (ICI) in Great Britain. B. American engineering education has not included much emphasis o n statisticid methodology. and of the gains that Japanese have made through intensive use of statistical methods. Multifactor experimentation.
we have outlined a systematic approach t o experimentation.S. but remember: not only will the interactions (if any) be aliascd all over the place. 19. newspapers conveyed the absurd allegation that the western statisticians are still advocating oneatatime experiments and are unaware of the principles of multifactor experimentation.9.0 0. you can hardly avoid learning something! Good luck! OKTHOGONALITY = BALANCE = SIMPLICITY = EFFICIENCY EXERCISES 19. This data set comes from Karlin (1987). Another caveat concerns variability. In the last three chapters.5 0. There really arc interactions.EXERCISES 343 himself. A policy o f picking the winner among the means of three levels o f it factor may get you into trouble if there is no significant difference. An article in one of the major U. 1 1 +1 1 1 1 1 tl 41 1 4. the important thing about welldesigned experiments is that they arc efficient and simple. and use orthogonal arrays. It can too easily get out of hand and rcsult in engineers blindly following the instructions in the handbook while their brains and enginecring judgment are left hanging on the hook outside thc door. Because of thc orthogonality. It is tempting to load up an L(32) lattice with 31 factors. Setting up a multifactor experiment in an orthogonal array is a straightforward matter.1. and you need to watch out for them. Because they are orthogonal factorial experiments.5 9.5 5.5 3. Hut thcrc is more than that. That is one of the reasons for the confirmation experiment in section 19.5 4. He is reviewing a software package and he includes this example of a design with five factors in an L ( 8 ) lattice and tivc observations at each point. Somc rabid disciples also go so far as to deny the existence of interactions.1.5 .0 3.2. which it claimed were developed outside the West. the calculations are straightforward. It is an example for a ‘Taguchi parameter design. Such blind partisanship by zealots is regrettable.0 8. But some caveats must be mentioned again.5. If you do it properly. No mattcr what language you speak. Fisher had to discourage some of his f agronomist colleagues from going off halfcocked and claiming that effects werc significant when they were not. One of the important uscs of the analysis of variance is to obtain a valid estimate o the error. the results are easy to present and easy to understand. including the calculation of all three signaltonoise ratios. Carry out the details of tlie analysis of the data set in table 19. but you will have no good handle on the variability.0 7. arguing that they do not exist in tlie real world.
94 7.5 9.62 7.5 1.and CD were the only interactions of interest.78 8.69 7.0 9..88 7.50 7.59 They took as their signaltonoise ratio 2 = 10 log.50 7.0 5.69 7..59 7. The data follow: B I 2 1 2 1 2 1 2 C BC D BD CD 2 1 2 2 1 1 1 1 1 2 2 1 1 2 1 2 2 1 1 1 1 I 1 2 2 1 2 2 1 1 2 I 2 1 2 2 2 2 2 2 E 1 2 2 1 2 1 1 2 7.25 7. quench oil temperature.00 8.is the free height of the spring.0 1.0 10. Its targct value is eight inches. hcating time. They wcre prepared to accept that BC.81 7.44 7.6 8.75 7. They investigated eight factors in an t ( 16) array.( $) . C.63 7.25 7.5 6. 2 What levels of the factors would you recommend? Which factors would you adjust to minimize the variance'? Having done that. Which would you regard as the important factors? 19.5 11.56 7.5 4.5 7. they mcasurcd the thickness of the epitaxial layer at five places on each of 14 wafers (70 measurements in all). holddown time. and E. and 0.5 19.0 17.81 7.56 7. D.56 High 0 7. Pigniatcllo and Ramberg (1985) reported an experiment on the heat trcatment of leaf springs on trucks.5 1.5 9.44 7.5 15.4.0 5. For each run. The rcsponse.5 4.5 16.344 1 TAGUCHI AND ORTHOGONAL ARRAYS +1 1 +I 1 +I +1 1 1 1 +I Cl +I +I +l +I +1 1 1 +1 1 +1 +1 1 1 +1 1 +1 1 +1 0.5 7.75 7.50 7.12 7.0 13. Kackar and Shoemaker (1986) used designed experiments to improve an intcgratcdcircuit fabrication proccss. susceptor rotation method.50 7.88 7.00? 19.81 Low 0 7.3.18 7. The engineers considered five factors in the manufacturing process: B.0 6.50 7.78 8.5 7. Y.32 7. were three runs at each set of conditions.09 7.code .0 7.18 7.44 7.56 8.5 Suppose that you were to use the signaltonoise ratio (SN).0 2. high heat temperature.which was repeated at both levels of 0. C.62 7.69 7. The factors were: A .D.15 7. The dcsign of the experiment was an t ( 8 ) array in There factors B .5 4.56 7.56 7.50 7.18 7.06 7. which factors would you adjust to move the response to y = 8. fi.5 2.88 7.56 7.transfer time.5 0. B.44 7. BD.0 4.88 8.
Hc assigned the fourlevel factors to three columns of the lattice.860 14.1.878 14.843 14.5 0.0.5 117.5.014 Y .4969 0.415 14. G .1989 19.6505 .5 1 0.442s .5 p m ? A 1 1 1 1 1 1 1 1 2 2 2 2 2 2 2 2 B 1 1 1 1 2 2 2 2 1 1 1 1 2 2 2 2 C 1 1 2 2 1 1 2 2 1 1 2 2 1 1 2 2 D 2 2 1 1 1 1 2 2 1 1 2 2 2 2 1 1 E 1 2 1 2 1 2 1 2 1 2 1 2 1 2 1 2 F 1 2 1 2 2 1 2 1 2 1 2 1 1 2 1 2 G 1 2 2 1 1 2 2 1 2 1 1 2 2 1 1 2 H 1 2 2 1 2 1 2 2 1 2 2 1 2 1 1 2 14. at what levels of the fourlevel factors should you run your process'? Docs it matter? A 1 ~ C L ) l i F G H 1 1 1 1 1 1 1 1 2 1 1 1 I 2 2 2 2 1 1 1 1 1 2 2 2 2 2 1 2 3 4 3 4 1 2 4 1 2 4 3 1 2 4 3 1 1 2 1 2 2 1 2 1 1 1 2 2 1 2 1 1 2 2 1 1 2 2 1 1 2 2 2 l Y 1 3. Which factors should you control to minimize t h e variance? Which factors should you then adjust to obtain a mean thickncss of 14. for example.deposition time. nozzle position.165 13. If the bit failed during the eighteenth cycle. and H . scc chapter 18. Y was recorded as 17.3267 0.5 1 2.5 1 0. C. 14 of the 32 bits failed during the first cycle.6270 0.972 14.888 14..037 13.021 13.1. D. Y is thc number of cycles to failure of the bit. E. F €ICI etch temp.5 1 0. He had eight factors at two levels each and two factors at four levels.5 . The data follow. HCI flow rate.8563 0.757 14.1190 0 . Phadke (1986) used an L(32) latticc in an experiment to comparc the lives of router bits.907 13.821 14.932 13.4230 0.5.4369 0. arsenic gas flow rate.1.3131 0.2292 .3467 0.4307 0.5 1 117.EXERCISES 345 of wafers. The orthogonal array follows together with the values of 9 and s Z at each point. Which were the important factors? Given that large values of Y are desirable.032 14. deposition temperature.6154 .5 1 0.0.880 14.862S log ( s 2 ) 0.
5 0.5 .5 17.5 0.5 0.5 14.5 0.5 3.5 3.5 0.346 TAGUC'III AND ORTHOGONAL ARRAYS 2 2 2 2 2 2 2 I 1 1 I 1 1 1 1 2 2 2 2 2 2 2 2 1 1 1 2 2 2 2 1 1 1 2 2 2 2 I 1 1 2 2 2 2 1 I 2 2 2 1 1 1 1 1 1 2 2 2 2 2 2 2 1 1 1 2 1 1 .5 0.5 2.5 17.5 0.5 17.5 3.5 3.5 2.5 17. 1 3 2 1 2 1 4 3 1 2 3 3 4 I 2 4 3 2 1 2 1 4 3 2 4 3 1 2 4 3 1 2 4 3 1 2 4 3 2 4 3 1 2 4 3 1 2 1 2 2 1 2 1 2 1 2 2 1 2 1 2 1 2 2 1 2 1 1 1 2 1 2 2 1 2 2 1 2 1 1 2 2 1 1 2 1 2 2 1 1 1 1 1 2 1 1 2 2 1 1 I 1 2 2 1 1 2 2 2 2 1 1 2 2 1 1 1 1 1 1 1 1 1 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 0.5 3.5 17.5 17.5 0.
M. (19Nh). and Huntcr. “A sampling inspection plan for continuous production.g Equations to Data: Computer Ana1y. 0. 8. Dodge. Church. . 229246.. Staristicsfor F:xperimenters. Wiley. 2nd cdition.. (1959). Wilcy. Jr.” Technomutrics. 2nd edition. F. Irwin.” Techtzometrics. H. and Wood. (1951). (IYXO). ti. “Use of halfnormal plots in interpreting factorial twolevel experiments. Dykstra. C.Statistical Methods in Engineering and Quality Assurance Peter W. (1959). Nine of his papers were collected in a memorial issue of Journul of Quality Technology. Wiley. F. Dodge. A. no. ”Chain sampling inspection plans. F. H. F. 5 . 264279. Davies. 1. 321342. Daniel. vol. vol. Dodge. vol. 1 1 . S . New York.” lndusrrid Quality Control. N. (. (1981). New York. New York.57. N. (19SSb)..uRrtzic:s. “The fractional replication of factorial arrangements. Daniel.s of Mll~tijhclorDutu. “Additional continuous sampling inspection plans.” Indicwial (2uality Control. 14. . H. .vol. “Analysis of data when the response is a curve. (l%l).. . New York. Sampling Insyection Tables. Draper. F. Hunter. Cornell. 9. 347 . 5. W.” hdustriul Quulity C‘onfrol. M. (Ed. C. (1943. no. Ihdge. 931. H. 712. D. New York. Homewood.. Inc References Box. L. G. Otto (195Y).” Annals of Mathematical Slaristics. H. J. Applied Regression Analysis.l. Dodge. b2tin. Finney.E. and Torrey.. and Smith.” Annals I( of EuF. G. Experimerits with Mixrures. and Komig. Wilcy. 2nd edition. (1977). 4.) (1956). 1 I . Dodge.si.vol. 12. Duncan.. vol. II li nois. “Skiplot sampling plans. P. 2nd edition. (19M).1955a). 291301. F. vol. Wilcy. J. V O ~ . l945). (1978). no. G. Hafncr. 1. Quulity C. John Copyright 0 1990 by John Wiley & Sons. New York. 5th cdition. Alonzo C. J. Design and Analysis of Inditstrial Experiments. 6370. 35. t i .’’ Tcchnometrics. *’Partial duplication of factorial experiments. H. 1013. 7.’otztrol Industrial Strrtistics. F. vol. A.
” Teclznomefrit. Oikawa. 4757. vol. 5150. John. G. “Control chart tests based on geometric moving averages. Wiley.1. (1087). 3950. “Transformations for attribute data. N. 100115. S. 15. “ldentifying useful differences in n multiplehead machine. vol. “Robust Design: A costeffectivc method for improving manufacturing processes. 5168.” Journul of L)ucility Technulugv. P. It. S. vol.. “Software on review. . vol 65. J . New York.” Journal of Quality Techriulugy. Ott. New York. and Snee. M. (1971). A. series B. J . Nelson. (19%). 20. E. issue 2. ’I. . Pigniatello. “Discussion of a paper by R. M.” Journal of Quality Technology. S. Karlin. MacRcth. Asian Productivity Organization. f’age.” Qualify Progress. 41. Nelson. (1985).” Juurnul of Quality 7echnulogy. vol. Plackett. C. (1973). and Rambcrg. Gordon ( lU89>. It.134. vol. S. D. “Design optimization case studies. 188192. S. Page. E. 833447. vol. Lucas. I. (1987). “Letter to the editor. vol. M. (1976). “Cumulative sum control charts.” Biomelrika. “Continuous inspection schemes. B T . MacinilIan. 5556. UNIPUB. and Oka. 6 5 . 78. “New techniques for approximating the strcss in padtype nozzles attached to a spherical shell.. Kackar. 4 .” A . R. . J . Ishikawa. 239250.12. “Combined Shewhartcusum quality control schcmcs. “The el’fect o f nennormality on the control limits of X charts. E. L. C. W. (IOSX). (19834. lacas. M. J.vol. 8 . ( 1982). (1961). (1985). 3. May 1989. and Shoemaker.” Biutnrtrika. R. 5 . January. Montgomery. J .3411 REFERENCES Hunter. (198S). 8 .” Tec:hnumetrics.” A . “‘lbe design of optimum multifactorial experiments. vol 1.” Setnironductor Internarional. (IYX7). J.. (lO46).” Jourrial of Qualify Techtcology. (19S9). Peter W. 20.. Robcrts. 130. R. vol. 305325. Hunter. and Nclson. L. “The dcsign and use of cuniulativc sum quality control schemes. 15. 344360. E. ”The deceptivencss of moving averages. vol. vol. vol. Technical Journal.” Journal of Qualit). 17. I . Lucas. S. Schilling. W.” Techtiornetrics. 19. vol. li. N . vol. 33. no. J. May. Introduction to Starislical Quulily Control. E. 5457. May. New York. S. 99.” Trunsucfions of the Amcrican Society uf Mechanical Engineers. (1986). Scheffc. TechnoloKy. (1973). 1831815. Phadke. J. Kackar.” Quality Progress.” Joiirrinl of Qunliry Technology. IS. Guide lo Qualiry Control. I. 198206. ( 1983b). M . P. S. (1976). ’I’echnicul Juurnul. S. J. 203210. (1976). “Experiments with mixtures. (lYS4). T. “A modified Vmask control scheme. vol. D. 20. T h e exponentially weighted moving average. issue 2. Stutisticul Design rind Analysis of Experiments. and Burman.” Journal uj’ Qucrliry ‘/ic*hnulogy. vol. K. T.” Juiirnul of [lie R u y d Sfulisticd l Society.s. T. & 7‘. 18.. “Check thermal effects in photoresist coating process.
“An investigation of OAbased methods for parameter design optimization. W. D. Exploratory Vala Analysis. (1931). “Developing blending models for gasoline and other mixtures. England. Massachusetts. E.. J.” Solid State Technology. (1982).. Mao. “The design and analysis of factorial experirncnts. J. (1987). Sequential Anulysis.. S.. Gerald %. J . Tukey. AddisonWesley. Wu. E. Wald. C. vol. 119130. “A quick compact two sample tcst to Duckworth’s spccifications. Shewhart. F.” lmpcrial Bureau of Soil Science. W. (1981). (1947).” Qucility Progress. G. Madison. Economic Conrrul of Quuiiry of Munufactured Product. Don W. (1987). May. and Jillie. 127. Yin. 1. 3148. A.” Technometrics. Snee. 23. . vol. 17. Wiley. (1977). (1984). Acceptance Sampling in Quality Control. New York. 2225. ‘I‘ukey.REFERENCES 349 Schilling. and Ma. Van Nostrand. W. G.” Technomrtrics. Dekker.132. Center for Quality and Productivity Improvement. F. vol. Reading. (19S9). New York. Schilling. Yates. 24. “Orthogonal design for process optimization and its application in plasma etching. “An overview of acceptance control. Harpenden. University of Wisconsin. A. F. New York. S. (1937).” Report No. R. S.
6554 .8 I33 3389 1.97500 .8 5 1 9 .5478 . 179 6 .94520 .5040 S438 3332 .69I 5 .96784 I .7 18 .97062 97670 .8749 3944 .96926 .7224 .08 .90320 .82 I2 . The Normal Distribution*'+ O r D b 4 z 1 .6736 .958I8 .4 .6331 .92785 .6700 .8665 3869 .5987 .6844 .3 3643 341 3 .6664 .91621 . John Copyright 0 1990 by John Wiley & Sons.I .96562 .96638 .7454 .93189 .6628 S910 .7019 .7995 3264 so00 S398 s793 .96485 .5714 .7967 3238 .5871 .96407 . 6 .75I 7 .659 I S596 .96246 .97558 .Table 1.02 .97381 1.O .6406 .8340 5359 5753 .6808 SO80 .94630 .1 I .94295 .a80 .95637 .7 123 .06 S199 .92364 3554 3770 2962 .5319 .1 0 .96712 .91774 .7642 .94408 .94738 .oo .93448 . 1466 9 .5279 .7852 .761I .90490 .7764 3051 3315 .5675 .92922 .9 .9744 3810 .7257 .8289 .2 s160 s557 S948 .91924 ( r  2 ! 1. M.97193 .94179 .9331 9 .6368 .95728 .7389 .3 .7823 3106 .853 I .7734 3023 .7580 .94950 .7703 .97257 .8438 Statistical Methods in Engineering and Quality Assurance Peter W. Inc 1.8508 3729 3925 .04 .7939 .3 0 .94845 .7673 .95 154 .92647 .09 ?r t x .8830 90147 .7549 .4 .95352 . 9 .62 7 1 .5 I20 .90988 .7324 .8888 90658 . I .5 .7910 .92507 .7357 .95449 96327 .93822 .94062 .97320 3485 3708 3907 .8186 .96856 .95543 .8365 3599 .96080 .5 16 .93943 .o I .93699 .90824 .95254 .7088 .7794 3078 .05 .6443 .92073 .8997 .7486 .2 I .7 .07 .7 190 .96995 .97I28 .95907 .91149 .8577 3790 3980 .93056 .7422 .7881 .6950 .6879 .6026 .6103 .95994 . 3849 .5239 .6064 .6255 .6293 .8 .55 I 7 .6141 . 157 7 .96164 .7054 .91309 .846 I 3686 .6772 .6985 .7291 .5636 .92220 93574 .97615 3621 .651 7 .95053 .
08 .9*7814 .9’6376 .9*4558 .9’5604 .9‘4988 .9’6319 .98422 .9’2378 .9*5385 .9’3963 .9’3129 .9’7299 .9’8965 .09 .9’6928 .9’8347 .6 3.9’2451 2.4 .oo .933590 .9834I .98300 .9’4024 .9’7398 .9’2112 .9’8146 .9’7673 .9’7523 . by A.9’4766 .9’8694 .o1 .9’8359 .98214 .9’6752 .9’8559 .97982 .9’6427 .9’8650 .9‘4777 .9379!B .934429 .98 169 .97778 .9’8999 .98840 .9’1836 .9‘0039 .9’6736 .9’0358 .9’4623 .9’6982 3.9*5926 .9’7493 .9’5499 .9*6696 .9’6103 .9’I 260 .938637 9‘0799 .9’8074 .9’6869 .935658 .9*8511 .9’7365 .9’0324 . I 2.9’7599 .9’8739 .943848 .1 3.9’663 I .9‘3052 .9*3593 .9’1802 .98382 .9’42M .8 3.9’841 I .9’6636 .9’3363 .9‘7843 4.937585 2.9‘6964 t From Srafisficul Tables and Formulas.p2468 .9‘7948 .98679 .98500 .9’4457 .9’581 I .9’5166 .98899 .928856 .938879 .938689 .N z  .9’2656 .9’7282 .9‘8736 .9‘3327 .9’4915 .9‘7090 .98645 .9*6406 .9’709 1 .9’8922 .04 .9’7197 .98574 .938787 .9’1344 .9‘4132 .9‘2 159 .9‘0426 .9*8817 .9’2240 . Hald and the publishers.9’7445 .reproduced by permission of Professor A.0 2.9’821 5 .7 2.2 2.98077 .9’3431 .02 . .9’5339 9’6533 .9‘7546 .3 2.9’8074 .932886 .97882 .9’5959 .97725 .9* I 158 .9’8583 .7 3.9’7020 .933810 .4 .9’4297 .9’7922 .98610 38928 .9’7197 .9’1106 .9‘2765 .9’8305 .3 3.98983 .9’2024 .947748 .946554 .6 2.9’5975 .9’8134 .9’6242 .9 .928462 .9*7211 . Hald.934991 . John Wilcy and Sons.98956 .2 3.9’8250 .9*5190 .9’0863 9’3053 .9’6093 .9’8409 .97831 .O 3.93473 .98778 .987 13 .9’8605 .9’8193 ’ 3.9’0097 .9’8964 .98030 .9’5855 .0  .9‘5731 9 6833 . New York (1952).9’8777 .98870 .9‘6833 .9’8012 .938527 .9’3790 .9’6505 .98537 .9’8834 .9*4331 .9’4614 .9*7327 .98 124 .97932 .9’5335 .9‘4094 .937674 .9*6253 .9’1553 .9’2636 9’4810 .9‘7439 .937759 9’8469 .9’2857 .5 3.9‘ I 504 .9‘5573 .9’7 1 10 .9’3244 .9’0957 .9’6207 .9’7842 .03 .9’8893 .07 .9’0613 .9 .9*5753 .9%%0 .06 .98745 .9‘7744 .9‘7882 .98257 .8 2.9*7649 9’8282 .9’ 1576 .05 .9* 1838 .9‘3613 .98809 .9’0646 .9’8930 .98461 .9’5201 .5 2.9‘6092 .
819 2.714 1.761 1.77 1 2.201 2.684 I .645 1.750 2.363 1. t From Biorncrriku Tables /or Srarisriciuns.289 1. reproduced by permission of the publishers.692 .734 1.699 .602 2.688 .756 2.727 .764 2.718 .1 10 2.341 1.345 1.763 2.83 1 2. .703 1.706 .500 2.333 1.131 2.462 2.72 1 1.68 1 .025 12.674 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 oc.326 .262 2.69 I .31 1 1.069 2.765 .355 3.056 2.965 4.695 .860 1.365 3.704 2.323 1.012 2.539 2.688 .552 2.677 .457 2.68 1 2. Hartley.'SABLES 353 Table 11.925 5.576 5 6 7 8 9 .048 2. S.467 2.638 1.998 2.718 2.980 1.753 I .650 2.845 2.920 2.318 1.365 2.706 4.687 .861 2.132 2.708 1.397 1.476 1.567 2.541 3.358 2.816 .740 1.658 I .604 4.05 6.145 2.005 63.328 1.686 .313 1.303 3. Vol.77 1 1.697 .060 2.671 That is.533 1.o1 31.101 2.685 .306 2.960 .120 2.316 1.415 1.711 .657 9.390 2.473 2.706 1.701 I .499 3.717 1.01 5 1.725 1.679 .776 2. interpolation should be carried out using the reciprocals of the degrees of freedom.685 .182 2.356 1.683 .821 2.943 1.02 1 2.779 2.179 2.947 2.878 2.143 2.337 1.074 2.086 2.694 .689 .353 2.617 2. values of r.898 2. edited by E.807 2.841 4.797 2.711 1.703 .684 .314 2.833 1.447 2.055 3.886 1.315 1.485 2.383 1.where rn equals d e g r m of freedom and .325 1.310 1.319 1.896 2.683 .782 1.423 2.624 2.330 1.25 1.977 2.169 3.250 3.821 6.660 2.684 .479 2.440 1.Pearson and H.042 2.350 1.518 2.000 1.707 3.747 3.583 2.Ooo .895 1.571 2. 60 120 30 40 I .746 1.078 I . 0.787 2. The tDistribution*'t'+ x 1 2 3 4 .372 1. Where necessary.064 2.303 1.700 .106 3.282 .314 1.528 2. I (2nd edition) Cambridge University Press (1958).741 1 3.296 1.686 .228 2.080 2.683 .684 . and for this the function 120/m is convenient.492 2.032 3.690 .052 2.729 1.697 I . .160 2.796 1.508 2.812 1.921 2.045 2.093 2.32 1 1.
5748 I 5.0705 I 2.8602 1 2 3 4 392704 x lo'' .1910 31S264 32.4831 2 1.56418 8.02389 7.05347 3.8325 14.8523 15.484419 3.3193 32.35 1846 .1413 30.206990 1 1.63273 6.60321 3.87944 10.872085 1.5476 20.2672 35.60094 5.73264 3.297I I0 982069 x .26214 6.2 1034 1 1.7496 18.48773 393214 x lo' .96 I64 8.40379 5. The Chisquare Distribution*'t .1170 .0863 16.26481 6.5779 3 1.8454 30.989265 1.6660 23.2 2935 5.0902 2 1.8 I94 31.6883 29.8381 14.9550 23.734926 14.16735 2.67 I76 9.8013 34.8693 30.2995 29.2 15795 .84398 5.07468 2.07382 3.0100251 .0261 22.1 190 27.5871 28.37776 9.57063 .2962 27.2177 21.8053 36.5966 12.99 147 7.@so .56503 4.40776 7.17973 2.89 I86 6.237347 1.68987 2.4884 28.26094 7.1433 6 5 9 7 8 3.9958 26.1564 38.9190 18.6848 24.62872 15 16 17 18 19 4.90655 7.59 1 6 157088 x .90766 7.831211 1.8 1473 9.4494 16.7185 37.087912 .07I72 12 .I02587 .3367 24.3449 13.1908 16.94030 4.411740 .70039 1.23075 8.3251 I 12.0201007 .0228 20.69724 6.5073 16.344419 1.8119 18.5582 I 3.025 .57056 4.7356 26.22603 5.995 .239043 1.00874 5.84 146 5.6751 2 1.3070 19.010 .24697 3.5822 10 11 12 13 14 2.0671 15.7569 28.Table 111.5893 25.m 6.362 1 23.0128 17.1882 26.66043 3.39046 10.950 .646482 2.4753 20.2170 27. I4224 5.I 14832 .9200 23.63539 2.63490 9.2093 24.975 .9 90 .145476 1.7250 26.554300 .4087 34.1 5585 2.01491 7.2767 7.1435 .5346 19.9999 33.34840 11.81221 6.675727 .710721 5.10691 4.81575 4.0506356 .
8852 40.299 140.8508 11.1539 88.5733 15.7576 57.5585 46.0471 80 70 51.5312 101.8484 1.7642 43.9798 .9968 41.7083 37.5400 61.9321 40.879 113.6907 76.8076 12.64272 9.9295 13.329 1 4 11 2.950 .3574 40.9278 48.1963 67.050 .11 46 4 15.26042 9.6525 38.9535 22.6466 74.4789 36.where m represents degrees of freedom and is approximately normally disFor m < 100.4900 91.081 9 46.4848 1 1.1133 41.5197 11.7585 67. edited by E.43386 8.39I5 69.7807 38.16% 35. For m > 100.3079 16.6705 33.4613 13.1 197 13.0757 39.7729 55.9232 43.1603 1 I .2565 8.2894 41.7659 79.48I 7 18.2976 I 13.I879 25 26 27 28 29 43.5662 38.990 .629 118.2752 51. Cambridge University Press (1958).OU 20 2 1 22 23 24 30 40 50 60 16.1944 44.1513 16.145 124. .7065 27.5569 40.3356 53 6 20 .I532 65. values of i:.6417 46.8564 39.03366 8.7067 37.3372 42.7222 31.7541 70.3141 45. linear interpolation is adequate.6465 4I .19567 10.7 66.4926 26.2182 49.28293 10.26040 8.807 104.0231 106. .169 That is.7867 20.Table 111 (continued) 975 .321 128. t From Biometrika Tables /or Srarisricions.9823 I 1. S.8439 14. I (2nd edition).88623 9.1813 45.9517 100 90 43.9907 35.4104 32. Parson and H.561 100.5093 34.4607 45. so that percentage points may be obtained from Table 11.3417 7I .54249 10.9792 59.6449 50.433I 32.5913 12.1720 59.3380 13.Hartley.1725 36.005 14.89720 9.3276 45.4151 44.6384 42.% 9 .5346 10.4202 83.5048 79.4010 42.215 116. Vol.9933 52.1260 77.59083 10.2219 90.1643 29.5648 14.7908 24.0905 13. 6 135.9244 35.3791 16.2899 49.9279 17.5240 12.44I8 53.3641 37.7393 60. a (r.6885 12.4011 10. reproduced by permission of the publishers.5879 50. 0.3794 34.198I 12.136 129.9630 48.425 1 12..8786 13.010 .342 95. = / and unit variance.1211 7.7956 44.8922 63. tributed with mean .0648 48.
9609 2.9327 1.2252 3.1423 2.9878 1.2987 2.2183 2.1279 2.5222 2.8833 2.0472 2.6056 2.1745 2.1075 2.6602 2.2081 2. I638 2.9577 2.5894 3.4887 2.7055 3.1843 2.7265 2.3427 4.3026 57.4898 2.8841 I .9269 1.9610 1.1808 2.3416 2.7632 2.2193 2.0909 2.5383 2.9292 1.3163 2.024 I 2.6927 2.4694 2.0553 2.3040 19 20 21 22 23 24 25 26 27 28 29 40 60 120 co 1.0913 2.0262 3.1 185 2.0284 2.46 I8 2.2850 3.7265 2.9819 1.6952 2.3473 2.5053 2.2831 2.7741 3.9549 3.2830 2.9790 4.7247 2.3077 2.0232 2.1949 2.3467 2.2517 3.9427 1.909 1 2.2438 2.2888 3.856O 1.9407 1.2858 2. I030 2.6241 2. I953 2.1655 2.8473 3.1539 2.1618 5.4160 2.3891 2.01 39 2.2341 2.2906 2.9649 I .4579 3.9271 2.9515 1.8918 1.9457 1.12% 2.0785 2.8289 1.9189 1.4530 3.3092 5.1908 3.5263 9 .9486 2.9959 I .9238 2.9374 2.3940 2.5362 2.5509 2.24 I 58.8064 2.0605 2.8725 1.8871 2.9345 1.1131 3.1760 2.0862 2.6315 1 39.6239 2.8274 2.8743 1.8747 1.2735 2.6053 2.4624 5.0732 3.3772 2.5383 4.6055 2.6682 2.9836 1.0580 2.3327 2.9357 3.1 280 2. m 5.8807 2.2926 9.3614 2.8238 1.3387 2.3668 5.1 I49 2.4512 2.6702 30 .8841 1.7797 3.204 58.3069 2.6843 1.2662 3.5613 2.2828 2.4337 2.9899 2.1073 3.2135 2.8595 2.0410 1.7929 1.1220 2.2730 2.9923 1.4633 3.1774 2.5283 2.0751 2.3393 2.5493 2.9480 1.5612 2.1017 2.4045 3.3246 4.9269 1.1571 2.9449 5 6 7 8 59.3473 2.7675 1.2426 2.4606 2.3649 2.3801 2.3255 9.7380 1.9177 2.0566 2.9668 1.7516 2.1582 2.1765 3.3274 2.8959 1.9747 2.4374 2.1422 2.0098 3.9803 1.4801 2.7277 2.1300 2 0838 4 55.0741 2.2489 2.3679 3.500 8.1494 2.5603 2.4403 2.0397 2.276 I 2.1783 2.8354 2.9531 1.3908 4.9714 1.4404 2.9605 2.7748 1.3512 2.3805 5.33 10 2.4955 2.356 Table 1V.9091 1.6195 3.6683 2.7914 2.0645 2.5106 2.7760 3. The I’ Distribution*’? .035 I 2.2663 2.0492 1. 1 2 3 4 5 6 7 8 9 11 TABLES 2 3 53.5746 2.5191 2.7849 2.0506 4.0084 1.9063 1.2847 5.8947 1.7220 1.1931 2.0546 2. I524 2.9826 1.3947 2.5202 3.7167 LO 12 13 14 I5 16 17 18 2.2574 3.5216 2.3603 3.048 1 3.9949 1.8652 I .593 9.3970 2.833 9.0922 2.8498 1.7478 2.0145 2.9968 1.1022 3.5028 2.8129 2.2261 2.1582 2.439 9.0070 2.0730 2.906 9.0613 2.858 9.5448 4.2434 5.8194 1.6106 2.0045 1.0065 2.2400 3.2065 2.3075 2.2446 2.2333 2.4140 2.0822 2.9245 2.349 1 5.3932 2.6446 2.0880 2.8068 2.9012 28939 2.5893 2.9001 1.3170 2.864 49.1958 2.5893 2.1094 2.0379 20171 9 59.9985 1.9830 2.0047 1.0604 3. I362 3.9188 1.
1512 5.6034 1.8658 1.1049 2.5871 1.8200 1.8111 1.6155 1.7827 1.2983 2.1843 2.740 9.3916 5.4163 2.8449 1.705 61.3830 2.3226 2.5437 1.49 13 5.6524 1.2974 2.7537 1.8593 1.2067 28363 2.9323 1.8359 I .8550 1.3162 2.5947 2.6890 1.4829 9.8924 1.7175 I .7223 I .4928 2.1671 2.695 1 1.9333 1.84M 1.284 I 2.9854 1.7973 1.I0 L FmI.6377 I .7667 I .3952 1.7551 1.4784 1.7964 1.1072 LO26 I 1.4373 1.0115 1.5934 1.a74 1.4746 9.1050 27222 2.8990 1.9997 1.91 I9 1.6012 I .8451 1.33% 2.6074 I .8462 1.9377 1.7146 1.5555 2.8656 1.5351 2.2320 2.3391 22085 2.5925 1.8319 1.3769 1.7756 1.6433 1.4248 I .6308 1.7395 1.8852 3.0954 2.6065 1.7439 1.6322 2.8090 1.7312 1.9625 1.6073 1.4755 1.028 1 2.1845 3.8036 3.7816 I .1337 5.8775 1.5732 I .5570 1. I554 2.5313 1.7492 1.1592 2.7938 1.7590 I .4663 9.8174 3.9576 1.408 1 9.2768 2.6224 1.9399 1.5036 1.4041 2.1402 27620 25142 2.8955 3.2304 3. I228 27423 2.6771 1.1905 2.1376 2.7298 1.7873 1.5805 1S686 1.5987 12 15 20 24 62.7070 1.5575 1.78% 3. I 10 60.6659 1.7592 1.8368 I .8624 1.5176 1.6714 1.1425 5.4670 1.8759 1.2646 1.6988 3.7627 I .4906 1.6252 1.8450 1.0070 1.7025 2.2380 2.9597 1.91 17 1.6228 1.5472 1.8913 1.4871 I .6856 1.783I 1.2482 2.5450 1.6852 1.6147 1.2682 2.6864 1.9315 1.8090 1.3789 2.8168 1.4496 5.9722 1. a t 2 1.0554 1.220 61.8000 2.6554 1.7808 1.9043 1.6768 I .8 183 2.6673 I .5825 1.7191 1.6465 1.6407 1.6721 1.265 9.8388 1.7021 1.1741 2.6567 I .8712 2.7182 1.7506 1.4246 2.4821 I .7159 1.7507 1.2 I56 5.1474 2.5715 1.7989 1.2400 I .7753 3.4708 2.5490 1.0537 2.4672 1.6560 1.5376 1.1686 3.1000 2.6624 I .3676 1.7895 I .9577 1.5380 2.8195 1.3832 30 62.748 I 1.1317 2.8310 3.9827 1.0597 2.0966 2.9197 I .4094 1.5020 2.5435 1.6389 I .91I7 1.44 I3 5.0593 2.0095 3.4564 1.1878 2.7805 1.574 I 1.668 I 2.7382 1.2007 2.7607 3.1926 1 2.7867 1.8272 1.2547 40 60 120 a ) 60.6589 1.1681 3.0818 1.2926 2.6052 1.8362 1.1764 3.0360 1.0009 1.4247 9.8 142 1.4642 2.3419 I .7596 1.0516 1.9369 2.5198 1.SO56 1.9721 1.9367 1.8750 1.0762 2.9557 1.M76 I .6569 1.2951 I .061 63.8572 1.W36 I .%58 62529 62794 63.6356 1.6468 1.3203 I .1727 1.7812 2.8903 1.7083 1.8103 1.7306 1.9 I 9 9 3.6032 1.4579 5.8280 1.1 597 3. I230 2.3614 2.9770 1. I784 2.0171 1.5753 2.328 9.7232 1.5090 1.7019 1.6574 1.TABLES 357 Table IV (continued) 0 a = .5107 1A 7 2 1. I573 2.6272 1.8689 3.9047 2.5668 1.5862 1.9243 1.6910 1.002 9.6759 3.2435 2.4989 I .7193 1.7302 1.6890 1.5327 1.oooo .0532 2.195 9.8647 1.541 I 1.7708 1.8728 1.2915 1.9043 1.8868 1.6662 1.2087 2.8108 1.9861 1.7059 I .4206 .2003 5.8443 3.
8452 8.3359 2.7426 2.4904 2.1709 4.5990 2.8853 3.4741 2.5521 9.2821 2.7642 2.1274 3.6672 4.1059 3.7278 2.7066 2.05 1 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 40 60 I 20 m 3 4 5 6 7 8 ~ __ 9 161.9223 2.0946 2.5435 2.2417 4.4495 2.4940 4.58 230.4226 2.9201 3.3205 2.7763 2.0164 1.8868 8.6458 2.7669 2.4221 3.9646 4.0280 3.8167 2.8123 7.6283 2.4590 4.0986 4.9604 2.0204 2.8100 2.4472 2.4047 2.4943 2.2107 2.6337 3.3053 2.71 224.9277 2.1240 2.358 Table IV (continued) 2 TABLES 2 = .0096 4.3002 2.4105 3.4381 3.3738 3.1968 3.3883 2.296 19.7257 3.5102 2.8321 2.1830 4.0718 2.7725 4.7870 3.7413 2.371 19.4591 2.8056 3.5252 2.1355 3.5219 3.6331 3.8661 2.9715 3.4324 2.0503 4.5727 2.2229 2.0401 1.9874 5.3343 2.0848 4.1960 4.3928 2.5874 3.7144 2.6060 2.5876 2.6823 3.3567 3.3852 3.2900 2.3719 5.8477 2.5337 4.9588 1.3248 4.6848 2.3748 2.7587 2.6767 3.4471 2.5868 2.2501 2.1599 3.8962 2.2655 2.1922 4.3883 6.0088 2.5454 2.8626 3.6802 2.8183 4.0410 5.2839 3.3541 3.9134 2.4928 3.513 19.4638 2.6896 2.2389 3.9957 5.4668 3.0662 3.9384 4.7374 4.54 18.9153 2.5581 2.2927 3.9443 6.45 199.8387 2.7014 2.3661 2.7964 2.7086 6.6613 2.9013 2.247 19.0717 29480 2.3593 2.4095 4.1468 3.7955 2.3258 3.2490 2.3690 3.3551 2.6987 2.8401 2.5915 3.2560 6.9467 2.4876 2.7472 4.7401 2.84 15 5.5005 3.2317 3.000 19.2066 3.7389 3.0725 3.1433 4.0069 2.77 238.3371 2.3874 3.8660 3.5480 2.3732 2.7083 3.5336 2.6408 2.4903 3.8524 2.50 215.6001 4.4227 2.3881 3.6207 2.7581 2.1791 3.1504 3.5377 2.2252 4.3201 2.4205 2.3404 3.4453 2.1203 3.8486 2.8443 4.2565 4.2592 3.2662 21802 2’0970 2.1172 9.0135 8.5806 3.9988 6.8759 4.3463 2.4434 3.5140 2.2296 3.128 9.543 1 4.3808 4.0942 6.99 236.5911 2.6079 5.6572 2.8799 .353 19.3277 3.2100 4.35 I3 4.5767 2.4 139 4.9647 2.0867 2.2141 4.3468 4.7861 5.4563 2.0990 3.3965 2.4768 2.0254 2.9961 2.1665 2.5546 3.5914 5.2793 4.4205 2.9340 2.2360 2.6613 2.8951 2.1028 3.9503 4.1122 3.3 177 5.3439 3.3009 4.6049 5.3419 2.4422 2.0491 3.2172 3.6143 2.4817 3.385 10.9582 2.7571 4.975 I 2.330 19.16 233.5914 6.7109 2.0012 3.4028 3.9406 8.4513 4.2597 4.2766 9.7729 2.1631 6.4780 3.8378 3.6875 3.88 240.2039 3.1 174 4.5491 2.0556 3.164 19.7141 2.9823 3.31 58 3.1750 2.2874 3.7905 2.2782 2.6030 2.3683 2.5277 25082 2.6400 2.9912 2.1789 3.6987 2.5719 2.2540 2.2913 2.0984 3.0123 2.
5055 2.9464 1.05 10 12 15 20 24 30 40 60 120 m 241.0629 2.1479 2.9446 1.7047 3.3210 2.2524 2.4445 3.0107 1.7534 2.4935 2.3522 2.3275 2.0889 2.5331 2.1058 2.5494 8.7505 1.8742 3.1757 2.2967 2.2258 2.4589 2.0716 2.0707 2.6878 5.6223 1.6602 8.7396 1.2214 4.7307 1.7029 8.0391 2.6541 1.4630 2.0096 1.462 8.9299 1.1497 2.9643 1.0053 2.6090 2.2966 2.9644 5.7398 3.5272 3.0428 28259 2.7740 2.0102 1.6188 3.8963 1.1503 2.5342 2.6037 2.5944 8.9838 1.9174 1.0148 1.7444 1.3% 19.446 19.4035 2.5436 2.4105 3.32 19.6365 3.0600 3.7689 1.6835 1.7522 4.8578 5.09 251.9380 1.8687 1.8415 3.7831 1.2033 2.5705 4.9005 1.0267 2.8432 1.2547 2.0794 2.8895 1.3893 1.0061 2.4045 2.8337 1.0589 2.5108 3.3879 2.1071 2.6021 2.0584 2.3519 1.91 245.7351 4.2962 2.1323 2.0476 2.2341 2.3984 3.3758 3.9464 1.8055 1.6084 1.6587 1.2298 2.9645 1.2043 2.4638 3.1307 2.8364 1.1834 21649 2.8027 1.0540 20283 2.0772 1.2776 2.9842 1.5173 4.3479 2.8217 1.7480 1.6981 1.4480 2.5581 3.7067 2.3410 2.2664 2.3472 3.9669 2.5943 1.4247 2.8025 5.9999 3.5437 2.9147 1.8637 2.0075 1.9005 27372 2.5379 2.1040 2.6928 1.3392 2.8543 1.1900 2.2747 2.3940 4.6211 2.1508 2.487 19.1242 2.9139 1.9898 1.8874 1.429 19.9117 5.2468 2.6664 4.3077 2.2840 3.1778 2.9105 1.5720 8.9192 1.9381 3.9365 2.9302 1.0264 1.8307 4.471 19.9010 1.0960 2.9796 1.5343 1.1179 2.8842 1.5309 24259 2.4957 3.4663 2.5089 1.2197 22043 2.3487 2.6996 2.05 250.7684 1.4591 4.3180 4.479 19.6866 2.01 249.6385 5.6710 2.8389 1.454 19.0658 2.1141 2.1938 2.1898 2.3080 2.34421 2.9938 1.8450 2.1555 2.6581 5.0035 1.0825 2.25 254.0275 2.3842 2.8117 1.5543 1.4952 1.2878 2.9165 1.2365 2.2504 2.6906 1.7488 1.1045 2.7001 1.2354 2.2036 2.95 248.6166 8.8657 1.9586 1.8424 1.oooO .5766 1.6281 4.2686 2.7331 1.1077 2.7170 5.7929 1.1768 2.5705 2.3043 3.4314 3.3803 2.0558 2.8082 3.14 25220 253.7743 3.9926 1.2539 1.2756 2.9168 1.4499 2.1906 2.1601 2.6464 2.1507 2.7446 8.9390 1.6491 1.7851 1.3779 2.4753 2.6377 1.7475 2.0729 2.7876 2.2674 2.7744 5.9782 2.7897 1.0921 2.6609 2.5265 8.9276 2.7138 1.8780 1.8649 1.4117 2.1152 2.413 19.2064 2.7570 1.0166 1.1477 2.3082 2.9317 1.8203 1.7537 1.2184 3.3807 2.1282 2.9605 1.9604 1.3404 3.6688 3.0050 1.TABLES 359 Table IV (continued) 2 = .7110 1.1 141 2.88 243.8533 1.8718 1.0411 2.1646 2.8536 2.1373 2.4290 1.8380 1.7459 5.6373 1.9130 2.8361 1.5747 3.9681 1.7872 2.6717 1.8920 1.4673 1.3650 3.4901 2.6777 3.6169 2.2304 2.9736 1.496 19.5801 2.8409 1.9245 1.7918 1.7855 8.0712 2.7186 2.2230 2.4202 2.8128 1.
360
Table IV (continued)
a
TABLES
4
.025
I
2 3 4
1
2
3
5
6
7
a
9
25 26 27 28 2 9
20 2 1 22 23 24
15 16 17 1 8 19
1 0 11 12 13 14 .
5 6 7 8 9
647.79 38.506 17.443 12218
m5 .0
10.007 883 .11 8.0727 7.5709 7.2093 6.9367 6.7241 6.5538 6.4143 6.2979
39.w 16.044 10.649
6.1995 6.1151 6.0420 5.9781 5.9216 5.8715 5.8266 5.7863 5.7498 5.7167
4.7650 4.6867 4.6189 4.5597 4.5075 4.4613 4.4199 4.3828 4.3492 4.3187 4.2655 4.2421 4.2205 42006
4.2909
5.4564 5.2559 5.0959 4.9653 4.8567
8.4336 72598 6.5415 6.0595 5.7147
956.66 963.28 948.22 899.58 9 7 11 3. 864.16 921.85 39.373 39.387 3.5 935 39348 39.331 39.165 39.298 1.7 443 1.0 511 1.4 450 1.3 475 14.624 14.885 15.439 89047 9.0741 9.6045 8.9796 9.1973 9.3645 9.9792 7.7636 6.5988 5.8898 5.4160 5.0781 7.3879 6.2272 5.5226 5.0526 4.7181 7.1464 5.9876 5.2852 4.8173 4.4844 43361 4.0440 38 11 .9 3.7667 3.6634 35764 3.5021 3.4379 3.3820 3.3327 3.2891 6.9777 5.8197 5.1186 4.6517 4.3197 68531 5.6955 4.9949 4.5286 4.1971 6.7572 5.5996 4.8994 4.4332 4.1020
4 I 528 . 4.0768 4.0112 3.9539 3.9034
4.8256 4.6300 4.4742 4.3472 4.24I 7
4.4683 4.2751 4.1212 3.9959 3.8919 3.8043 3.7294 3.6648 3.6083 3.5587
4.072I 3.8807 3.7283 3.6043 33014 3.4147 3.2767 3.2209 3.1718 3.1283 3.0895 3.0546 3.0232 29946
3.3406
40 60 120
00
30
5.6864 5.6586 5.6331 5.6096 5.5878
5.5675 5.4239 5.2857 5.1524 5.0239
4.1821 4.0510 3.9253 3.8046 3.6889
3.5894 3.4633 3.3425 3.2270 3.1 1 1 6
3.6943 3.6697 3.6472 3.6264 3.6072
3.8587 3.8188 3.7829 3.7505 3 7 1I .2
3.5147 3.4754 3.4401 3.4083 3.3794 3.3530 3.3289 3.3067 3.2863 3.2674
32501
32151 3 I835 . 3.1548
3.2499 3.1261 3.0077 2.8943 2.7858
3.0265 29037 2.7863 2.6740 2.5665
3.1287 3.1048 3.0828 3.0625 3.0438
29685 2.9447 29228 2.9027 2.8840 2.8667 2.7444 2.6274 2.5154 2.4082
28478 28240 28021 2.7820 27633
2.7460
3.0074 2.9686 2.9338 292 .04 28738
3.2934 3 2I94 . 3 I 556 . 3.0999 300 .59
3.9498 3.7586 3.6065 3.4827 3.3799
3.1987 3.1248 301 .60 3.0053 2.9563
3.8549 3.6638 3.5118 3.3880 32853
5.5234 4.8232 4.3572 4.0260
6.6810
3.7790 3.5879 3.4358 3.3120 3.2093 3.1227 3.0488 29849 299 .21 2.8800
2.9128 2.8740 2.8392 2.8077 2.7791 2.7531 2.7293 2.7074 26872 26686
263 .28 2.5068 2.3948 22875
26513 25289 2.4117 22994 21918
2.6766 2.6528 26309 26106 25919 25746 2.4519 23344 2.2217 2.1136
28365 2.7977 27628 27313 272 .07
'TABLES
361
Table 1V (continued)
I
= .025
1 0
12
15
20
24
3 0
40
60
120
co
968.63 976.71 984.87 997.25 1001.4 1005.6 1009.8 1014.0 1018.3 993.10 39.398 39.43 1 39.4 15 39.456 39.456 39.473 39.481 39.448 39.490 39.498 14.419 14.337 14.253 14.167 14.124 34.081 14.037 13.992 13.947 13.902 8.8439 8.7512 8.6565 8.5599 8.5109 8.4613 8.411 1 8.3604 8.3092 8.2573 6.6192 5.4613 4.761 1 4.295 1 3.9639 3.7 I68 3.5257 3.3736 3.2497 3.1469 3.0602 2.9862 2.9222 2.8664 2.8173 2.7737 2.7348 2.6998 2.6682 2.5396 2.6135 2.5895 2.5676 2.5473 2.5286 2.51 12 2.3882 2.2702 2.1570 LO483 6.5246 5.3662 4.6658 4.1997 3.8682 3.6209 3.4296 3.2773 3.1532 3.0501 2.9633 2.8890 2.8249 2.7689 2.7 196 2.6758 2.6368 2.6017 2.5699 2.54 12 2.5149 2.4909 2.4688 2.4484 2.4295 2.4120 2.2882 2.1692 2.0548 1.9447 6.4277 5.2687 4.5678 4.1012 3.7694 3.52 17 3.3299 3.1772 3.0527 2.9493 2.862 1 2.7875 2.7230 2.6667 2.6171 2.573 I 2.5338 2.4984 2.4665 2.4374 2.4110 2.3867 2.3644 2.3438 2.3248 2.3072 2.1819 2.0613 I .9450 1.8326 6.3285 5.1 684 4.4667 3.9995 3.6669 3.4186 3.2261 3.0728 2.9477 2.8437 2.7559 2.6808 2.6 I58 2.5590 2.5089 2.4645 2.4247 2.3890 2.3567 2.3273 2.3005 2.2759 2.2533 2,2324 2.2131 2. I952 2.0677 1.9445 1.8249 I .7085 6.2780 5.1172 4.4150 3.9472 3.6142 3.3654 3.1725 3.0187 2.8932 2.7888 2.7006 2.6252 2.5598 2.5027 2.4523 2.4076 2.3675 2.3315 22989 2.2693 2.2422 2.2174 2.1946 2.1735 2.1540 2.1359 2.0069 1.8817 1.7597 1.6402 6.2269 5.0652 4.3624 3.8940 3.5604 3.3110 3.1176 2.9633 2.8373 2.7324 2.6437 2.5678 2.5021 2.4445 2.3937 2.3486 2.3082 2.2718 2.2389 2.2090 2.1816 2.1S65 2.1334 2.1 121 2.0923 2.0739 1.9429 1.8152 1.6899 1.5660 6.1751 5.0125 4.3089 3.8398 3.5055 3.2554 3.0613 2.9063 2.7797 2.6742 2.5850 2.5085 2.4422 2.3842 2.3329 2.2873 2.2465 2.2097 2.1763 2.1460 2.1183 2.0928 2.0693 2.0477 2.0276 2.0089 1.8752 1.7440 1.6141 1.4835 6.1225 5.9589 4.2544 3.7844 3.4493 3.1984 3.0035 2.8478 2.7204 2.6142 2.5242 2.4471 2.3801 2.3214 2.2695 2.2234 2.1819 2.1446 2.1107 2.0799 2.0517 2.0257 2.0018 1.9796 1.9591 1.9400 1.8028 1.6668 1.5299 1.3883 6.0693 4.9045 4.1989 3.7279 3.3918 3.1399 2.9441 2.7874 2.6590 2.5519 2.461 1 2.3831 2.3153 2.2558 2.2032 2.1562 2.1141 2.0760 2.0415 2.0099 1.9811 1.9545 1.9299 1.9072 1.8861 1.8664 1.7242 1.5810 1.4327 1.2684 6.0153 4.8491 4.1423 3.6702 3.3329 3.0798 2.8828 2.7249 2.5955 2.4872 2.3953 2.3163 2.2474 2.1869 2.1333 2.0853 2.0422 2.0032 1.9677 1.9353 1.9055 1.8781 1.8527 1.8291 1.8072 1.7867 1.6371 1.4822 1.3104 1.oooO
362
Table IV (continued)
'TABLES
1
1 2 3 4 5 6 7 8 9
10 11 12 13 14
2
3
4
5
6
7
8
9
40522 4999.5 5403.3 5624.6 5763.7 5859.0 5928.3 5981.6 6022.5 99.000 99.166 99.249 99.299 99.332 99.356 99.374 99.388 98.503 27.672 34.116 28.710 28.237 27.911 27.489 27.345 30.817 29.457 15.207 14.976 14.799 14.659 21.198 15.522 15.977 18.000 16.694 16.258 13.745 12.246 11.259 10.561 10.044 9.6460 9.3302 9.0738 8.8616 8.6831 8.5310 8.3997 8.2854 8.1850 8.0960 8.0166 7.9454 7.8811 7.8229 7.7698 7.7213 7.6767 7.6356 7.5976 7.5625 7.3141 7.0771 6.8510 6.6349 13.274 10.925 9.5466 8.6491 8.0215 7.5594 7.2057 6.9266 6.7010 6.5149 6.3589 6.2262 6.1 121 6.0129 5.9259 5.8489 5.7804 5.7190 5.6637 5.6136 5.5680 5.5263 5.4881 5.4529 5.4205 5.1785 4.9774 4.7865 4.6052 12.060 9.7795 8.4513 7.5910 6.9919 6.5523 6.2167 5.9526 5.7394 5.5639 5.4170 5.2922 5.1850 5.0919 5.0103 4.9382 4.8740 4.8166 4.7649 4.7181 4.6755 4.6366 4.6009 4.5681 4.5378 4.5097 4.3126 4.1259 3.9493 3.7816 11.392 9.1483 7.8467 7.0060 6.4221 5.9943 5.6683 5.4119 5.2053 5.0354 4.8932 4.7726 4.6690 4.5790 4.5003 4.4307 4.3688 4.3134 4.2635 4.2184 4.1774 4.1400 4.1056 4.0740 4.0449 4.0179 3.8283 3.6491 3.4796 3.3192 10.967 8.7459 7.4604 6.6318 6.0569 5.6363 5.3160 5.0643 4.8616 4.6950 4.5556 4.4374 4.3359 4.2479 4.1708 4.1027 4.0421 3.9880 3.9392 3.8951 3.8550 3.8183 3.7848 3.7539 3.7254 3.6990 3.5138 3.3389 3.1735 3.0173 10.672 8.4661 7.1914 6.3707 5.8018 5.3858 5.0692 4.8206 4.6204 4.4558 4.3183 4.2016 4.1015 4.0146 3.9386 3.8714 3.8117 3.7583 3.7102 3.6667 3.6272 3.5911 3.5580 3.5276 3.4995 3.4735 3.2910 3.1187 2.9559 2.8020 10.456 8.2600 6.9928 6.1776 5.6129 5.2001 4.8861 4.6395 4.4410 4.2779 4.1415 4.0259 3.9267 3.8406 3.7653 3.6987 3.6396 3.5867 3.5390 3.4959 3.4568 3.4210 3.3882 3.3581 3.3302 3.3045 3.1238 29530 27918 2.6393 10.289 8.1016 6.84001 6.0289 5.4671 5.0567 4.7445 4.4994 4.3021 4.1399 4.0045 3.8896 3.7910 3.7054 3.6305 3.5644 3.5056 3.4530 3.4057 3.3629 3.3239 3.2884 3.2558 3.2259 3.1982 3.1726 2.9930 2.8233 2.6629 2.5113 10.158 7.9761 6.7188 5.9106 5.3511 4.9424 4.6315 4.3875 4.1911 4.0297 3.8948 3.7804 3.6822 3.5971 3.5225 3.4567 3.9381 3.3458 3.2986 3.2560 3.2172 3.1818 3.1494 3.1195 3.0920 3.0665 2.8876 2.7185 2.5586 2.40'73
15 16 17 18 20 21 22 23 24 25 26 27 28 29
19
30 40 60 120 w
5.3904
TARLFS
Table IV (continued)
u = .01
363
10
12
15
20
24
30
40
60
120
00
6055.8 6106.3 6157.3 6208.7 6234.6 6260.7 6286.8 6313.0 6339.4 6366.0 99.491 99.501 99.458 99.466 99.474 99.483 99.399 99.416 99.432 99.449 26.125 26.316 26.221 27.229 26.505 26.411 26.690 26.598 27.052 26.872 13.463 13.652 13.558 13.838 13.745 14.546 14,198 14.020 13.929 14.374 10.051 7.8741 6.6201 5.8143 5.2565 4.8492 4.5393 4.2961 4.1003 3.9394 3.8049 3.6909 3.5931 3.5082 3.4338 3.3682 3.3098 3.2576 3.2106 3.1681 3.1294 3.0941 3.0618 3.0320 3.0045 2.9791 2.8005 2.6313 2.4721 2.3209 9.8883 7.7183 6.4691 5.6668 5.1114 4.7059 4.3974 4.1553 3.9603 3.8001 3.6662 3.5527 3.4552 3.3706 3.2965 3.2311 3.1729 3.1209 3.0740 3.0316 2.9931 2.9579 2.9256 2.8959 2.8685 2.8431 2.6648 2.4961 2.3363 2.1848 9.7222 7.5590 6.3143 5.5151 4.9621 4.5582 4.2509 4.0096 3.8154 3.6557 3.5222 3.4089 3.31 17 3.2273 3.1533 2.0880 3.0299 2.9780 2.9311 2.8887 2.8502 2.8150 2.7827 2.7530 2.7256 2.7002 2.5216 2.3523 2.1915 2.0385 9.5527 7.3958 6.1554 5.3591 4.8080 4.4054 4.0990 3.8584 3.6646 3.5052 3.3719 3.2588 3.1615 3.0771 3.0031 2.9377 2.8796 2.8174 2.7805 2.7380 2.6993 2.6640 2.6316 2.6017 2.5742 2.5487 2.3689 2.1978 2.0346 1.8783 9.4665 7.3127 6.0743 5.2793 4.7290 4.3269 4.0209 3.7805 3.5868 3.4274 3.2940 3.1808 3.0835 2.9990 2.9249 2.8594 2.801 1 2.7488 2.7017 2.6591 2.6203 2.5848 2.5522 2.5223 2.4946 2.4689 2.2880 2.1154 1.9500 1.7908 9.3793 7.2285 5.9921 5.1981 4.6486 4.2469 3.9411 3.7008 3.5070 3.3476 3.2141 3.1007 3.0032 2.9185 2.8442 2.7785 2.7200 2.6675 2.6202 2.5173 2.5383 2.5026 2.4699 2.4397 2.4118 2.3860 2.2034 2.0285 1.8600 1.6964 9.2912 7.1432 5.9084 5.1156 4.5667 4.1653 3.8596 3.6192 3.4253 3.2656 3.1319 3.0182 2.9205 2.8354 2.7608 2.6947 2.6359 2.5831 2.5355 2.4923 2.4530 24170 23840 2.3535 2.3253 2.2992 2.1142 1.9360 1.7628 1.5923 9.2020 7.0568 5.8236 5.0316 4.4831 4.0819 3.7761 3.5355 3.3413 3.1813 3.0471 29330 2.8348 2.7493 2.6742 2.6077 2.5484 2.4951 2.4471 2.4035 2.3637 2.3273 22938 22629 2.2344 2.2079 2.0194 1.8363 1.6557 1.4730 9.1118 6.9690 5.7312 4.9460 4.3978 3.9965 3.6904 3.4494 32548 3.0942 2.9595 2.8447 2.7459 2.6597 2.5839 2.5168 2.4568 2.4029 2.3542 2.3099 2.2695 22325 2.1984 2.1670 21378 2.1107 1.9172 1.7263 1.5330 1.3246 9.0204 6.8801 5.6495 4.8588 4.3105 3.9090 3.6025 3.3608 3.1654 3.0040 2.8684 2.7528 2.6530 2.5660 2.4893 2.4212 2.3603 2.3055 2.2559 2.2107 2.1694 2.1315 2.0965 2.0642 2.0342 2.0062 1.8047 1.6006 1.3805 1.oooO
364
Table I (continued) V
1
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 16211 198.50 55.552 3 1.333 22.785 18.635 16.236 14.688 13.614 12.826 12.226 11.754 11.374 11.060 10.798 10.575 10.384 10.218 10.073 9.9439 9.8295 9.7271 9.6348 9.5513 9.4753 9.4059 9.3423 9.2838 9.2297 9.1797 8.8278 8.4946 8.1790 7.8794
31
TABLES
= .005
2
2 m 199.00 49.799 26.284 18.314 14.544 12.404 11.042 10.107 9.4270 8.9 122 8.5096 8.1865 7.9217 7.7008 7.5138 7.3536 7.2 148 7.0935 6.9865 6.8914 6.8064 6.7300 6.6610 6.5982 6.5409 6.4885 6.4403 6.3958 6.3547 6.0664 5.7950 5.5393 5.2983
3
21615 199.17 47.467 24.259 16.530 12917 10.882 9.5965 8.7171 8.0807 7.6004 7.2258 6.9257 6.6803 6.4760 6.3034 6.1556 6.0277 5.9161 5.8 177 5.7304 5.6524 5.5823 5.5190 5.461 5 5.4091 5.361 1 5.3 170 5.2764 5.2388 4.9759 4.7290 4.4973 4.2794
4
22500 199.25 46.195 23.155 15.556 12.028 10.050 8.805 1 7.9559 7.3428 6.8809 6.521 1 6.2335 5.9984 5.8029 5.6378 5.4967 5.3746 5.268 1 5. I743 5.091 1 5.0168 4.9500 4.8898 4.835 1 4.7852 4.7396 4.6977 4.659 1 4.6233 4.3738 4.1399 3.9207 3.7151
5
23056 199.30 45.392 22.456 14.940 11.464 9,5221 8.3018 7.47 I 1 6.8723 6.42 17 6.07 11 5.7910 5.5623 5.372 1 5.21 17 5.0746 4.9560 4.8526 4.7616 4.6808 4.6088 4,544 1 4.4857 4.4321 4.3844 4.3402 4.2996 4.2622 4.2276 3.9860 3.7600 3.5482 3.3499
6
23437 199.33 44.838 21.975 14.513 11.073 9.1554 7.9520 7.1338 6.5446 6.1015 5.7570 5.48 19 5.2574 5.0708 4.9134 4.7789 4.6627 4.5614 4.4721 4.393 I 4.3225 4.2591 4.2019 4.1500 4.1027 4.0594 4.0197 3.9830 3.9492 3.7 129 3.4918 3.2849 3.0913
7
23715 199.36 44.434 21.622 14.200 10.786 8.8854 7.6942 6.8849 6.3025 5.8648 5.5245 5.2529 5.03 13 4.8473 4.6920 4.5594 4.4448 4.3448 4.2569 4.1789 4.1094 4.0469 3.9905 3.9394 3.8928 3.8501 3.81 10 3.7749 3.7416 3.5088 3.25 11 3.0874 2.8968
8
23925 199.37 44.126 21.352 13.961 10.566 8.6781 7.4960 6.6933 6.1159 5.6821 5.345 1 5.0761 4.8566 4.6743 4.5207 4.3893 4.2759 4.1770 4.0900 4.0128 3.9440 3.8822 3.8264 3.7758 3.7297 3.6875 3.6487 3.6 130 3.5801 3.3498 3.1344 2.9330 27444
9
24091 199.39 43.882 21.139 13.772 10.391 8.5138 7.3386 6.4511 5.9676 5.5368 5.2021 4.935 1 4.7 173 4.5364 4.3838 4.2535 4.1410 4.0428 3.9564 3.8799 3.81 16 3.7502 3.6949 3.6447 3.5989 3.5571 3.5186 3.4832 3.4505 3.2220 3.0083 2.8083 2.6210
60 120
5,
40
30
TABLES
365
Table IV (continued)
a = .005
10
24224 199.40 43.686 20.967 13.618 10.250 8.3803 7.2107 6.4171 5.8467 5.4182 5.0855 4.8199 4.6034 4.4236 4.2719 4.1423 4.0305 3.9329 3.8470 3.7709 3.7030 3.6420 3.5870 3.5370 3.4916 3.4499 3.4117 3.3765 3.3440 3.1 I67 2.9042 2.7052 2.5188
12
24426 199.42 43.387 20.705 13.384 10.034 8.1764 7.0149 6.2274 5.6613 5.2363 4.9063 4.6429 4.428 1 4.2498 4.0994 3.9709 3.8599 3.7631 3.6779 3.6024 3.5350 3.4745 3.4199 3.3704 3.3252 3.2839 3.2460 3.21 11 3.1787 2.9531 2.7419 2.5439 2.3583
15
24630 199.43 43.085 20.438 13.146 9.8140 7.9678 6.8 143 6.0325 5.4707 5.0489 4.7214 4.4600 4.2468 4.0698 3.9205 3.7929 3.6827 3.5866 3.5020 3.4270 3.3600 3.2999 3.2456 3.1963 3.1515 3.1104 3.0727 3.0379 3.0057 2.781 I 2.5705 2.3727 2.1868
20 24836 199.45 42.778 20. I67
12903
24
30
25044 199.47 42466 19.892 12.656 9.3583 7.5345 6.3961 5.6248 5.0705 4.6543 4.3309 4.0727 3.8619 3.6867 3.5388 3.4124 3.3030 3.2075 3.1234 3.0488 2.9821 2.922 1 2.8679 2.8 187 2.7738 2.7327 2.6949 2.6601 2.6278 2.4015 2.1874 1.9839 1.7891
40
25148 199.47 42308 19.752 I2530 9.2408 7.4225 6.2875 5.5186 4.9659 4.5508 4.2282 3.9704 3.7600 3.5850 3.4372 3.3107 3.2014 3.1058 3.02 15 2.9467 2.8799 2.8 198 2.7654 2.7160 2.6709 2.62% 2.5916 2.5565 2.5241 2.2958 2.0789 1.8709 1.6691
60
25253 199.48 42149 19.61 1 12.402 9.1219 7.3088 6.1772 5.4 104 4.5592 4.4450 4.1229 3.8655 3.6553 3.4803 3.3324 3.2058 3.0962 3.0004 2.9 159 2.8408 27736 27 132 26585 2.6088 25633 25217 2.4834 2.4479 24151 2.1838 1. U % 1.7469 1.5325
120
25359 199.49 4 1.989 19.468 12.274 9.0015 7.1933 6.0649 5.3001 4.7501 4.3367 4.0149 3.7577 3.5473 3.3722 3.2240 3.0971 2.987 I 28908 2.8058 2.7302 2.6625 2.60 I6 2.5463 2.4960 2.4501 2.4078 2.3689 2.3330 2.2997 2.0635 1.8341 I .6055 1.3637 and
m
24940 199.46 42.622 20.030 12780 9.4741 7.6450 6.5029 5.7292 5.1732 4.7557 4.43 15 4.1726 3.9614 3.7859 3.6378 3.51 12 3.4017 3.3062 3.2220 3.1474 3.0807 3.0208 2.9667 2.9 I76 2.8728 2.8318 2.7941 2.7594 2.7272 2.5020 22898 2.0890 1.8983
25465 199.51 4 1.829 19.325 12.144 8.8793 7.0760 5.9505 5.1875 4.6385 4.2256 3.9039 3.6465 3.4359 3.2602 3.1115 29839 2.8132 2.7762 2.6904 26140 2.5455 2.4837 2.4276 2.3765 2.3297 2.2867 2.2469 2.2102 2.1760 1.9318 1.6885 1.431 I 1.oooo
9.5888 7.7540 6.6082 5.8318
5.2740 4.8552 4.5299 4.2703 4.0585 3.8826 3.7342 3.6073 3.4977 3.4020 3.3178 3.243 1 3.1764 3.1165 3.0624 2.9685 2.9275 2.8899 2.8551
3.01 33
2.8230 2.5984 2.3872 2.1881 1.9998
That is. values of F
where (m,. m l ) is the pair o degrm of fradom in F f
t From 'Tabla of percentage points of the Inverted Beta (F)Distribution," Biometrik Vol. 33 (1943). pp. 7388. by Maxine Merrington and Catherine M.Thompson; reproduced by permission of E. S Pcarson. I f necessary. interpolation should be carried out using the reciprocals of the degree o . f freedom.
Statistical Methods in Engineering and Quality Assurance
Peter W. M. John Copyright 0 1990 by John Wiley & Sons, Inc
Index
A
A2, 147 Ac, see Acceptance number
Acceptable quality level (AQL), 180, 183. 198 Acceptance number, 17Y, 199 Acceptance region, 113 Acceptance sampling, 1,7 Algorithm, Yates', 292 Alias, 297298 Alpha risk, see Hypothesis testing Alternative. see Hypothesis testing American National Standards Institute, 197 American Society for Quality Control, 197 Analysis of covariance, 267271 Analysis of variance, 228,251,264 ANOVA, see Analysis of variance ANSI, see American National Standards Institute ANSVASQC %I .4, 197 ANSI/ASQC Z1.9, 197, 201 AOQ, see Average outgoing quality (AOQ) AOQL, see Average outgoing quality limit (AOQI,) Approximations to probability densities: binomial approximation to 0 hypergeometric, 4 normal approximation to binomial, 72 normal approximation to rank sum, 135
Poisson approximation to binomial. 182 AQL, see Acceptable quality level (AQl1 Arcsin. 171 Arcsine transformation, 171 Arithmetic moving average chart, 203, 207 ARL, 154. See also Average run Icngths (AW Array, see Orthogonal arrays ASQC, see American Society for Quality Control Assignable cause, 113, 146, 169, 173 ATI, see Averagc total inspection (ATI) Attributes, 295 I control charts for, 164175 Average outgoing quality (AOQ), 184 Average outgoing quality limit (AOQL), 184186 Average run lengths (ARL). 154,204, 206 Averagc total inspection (ATI), 186. 189
B
B J ,Bj. 156, 157 Backsolution, 23 1 Balance. 274,342 Bayes theorem, 42 Behnken, D.H., 321322.331 BchrensFisher problem, 132 Bernoulli trials. 3637 Beta distribution, see Distributions Beta risk, bee Hypothesis testing
367
303 Degrees of freedom.f. 107 Binomial charts. 7. see Distributions Chain. 268 Conditional density functions. see Hypothesis testing CSP (continuous sampling plans).4. 101.7. Jr.230. J.C. I64 Defective.98. 159. 67 Conditional probabilities.297.332 Cp&.98. 164171 ccharts. 158160 Causality. see Degrees of freedom Discrete variables.295296 Davies. 88.292295.66 Blocking. 165166 p charts.331. see Assignable cause Causeandeffect diagram. 178 Defining contrast. 172175 cumulative sum. 172175 Central limit theorem. 298. 6769.232 Confirmation experiment. 276 BLUE (best linear unbiased estimate). See ulso Defining contrast Control lines. see Boxandwhisker plots Boxandwhisker plots. 336 Coin tossing model.2 132 14 I charts. 190 d2 (for control charts). see Distributions Bivariate distributions.368 Biased estimates. 189 ~ 4 148.S. 326. 321322 Boxplot. 77. W Density functions.E. 298 Chain sampling plans.5269 Contrast. 147 D3. 2951 .248. 228. 106. 160161 np charts.279280 Demerit.Ci. 197 Cumulative distribution function.251 Covariance.. 1.217. 203. 293 Coeflicient of variation. INDEX C CIS c2.. 159. 60 Charts.207.342 BoxBehnken designs.14 ccharts. 104.274. 174 Deming. R. A. alias. 314315317. 113.218 analysis of. A. 41 Confidence intervals.. 101 Burman. 164.. 104. 10. 2 18 Cause. 203. 2225.156. O.P. 164175 binomial. see Distributions Church. 218. 146147 Correlation coefficient.. 203207 attributes. 72. 203. 7.332 Capability of a process. 138. 33 Concomitant variables.321322. C. d2 (for double sampling).3 1 1.. 122.248 Box.157 .209213 E W A .331 DeMoivre. 126.P. 29. 13. 2 142 I5 geometric average. 164171 Binomial distribution. 80 CFR (constant failure rate) model. 113 arithmetic average.L. 114. 159 Critical region. W97. 13 Defect. 156 bar charts. W. 146156 Chisquare distribution. 159 CPU (upper capability level).. see Cumulative sum control chart D d1. 20213 Cusum chart. see Analysis of covariance Covariate. 44. 187 Characteristic life. 288. 203.. 297 Constant failure rate. 340 Conforming 178 Codounded. 81. C?. see Plackett. Set ulsu Standard deviation d. Cp. 342 Deciles. 25.8.E.04. 148151 s charts. 267 CPL (lower capability level).148149 Ilaniel. see Probability density functions Deviation. 167170 R charts.284. 207. 54 Cumulative sum control chart. assignable. see Distributions Continuous variables.
220 Gamma distribution.238 Duckworth test.. 129. 30 Weibull. 1 IS. s w Distributions Gaussian distribution.94. N. 8283 multinomial. see Distributions Geometric moving average chart. 248 Geometric distribution. 285.112 binomial. normal geometric.5237 F Factor.. see Exponentially weighted moving average chart Expectation. see Distributions GaussMarkov theorem.125.265 G Galton. 237238 Extrapolation. loo uniform (continuous). 5. I .66. 18.F.83. 212 Houghton.197 DodgeRomig sampling plans. I 6 1 beta risk.73.. uniform (discrete). 13 112 207.81. see lnterquartile range 284 Hunter. 324 Hypothcsis testing. 71.1 0 182 4. 52. 172176.85. 30. 3. D. 7. Hypergeometric distribution.273 Factorial experiments..189. Sir Francis...189 Dotplot..89. 102.100.5640.342343 gamma. I 0 0 GrdecoLatin squares.INDEX 369 Farrell. AJ. 129.123. 52.214215 Exponential model. 40.264. W.R. 115. chisquare.4647. minimum.25fi257 fit. 137 113.201 113. 1 0 3 Double sampling 190 Draper. 1 1 W. 323324 Graphs.3337.181 alternative hypothesis.6164.267. See also Tests acceptance region.o f . 7287. Gossct. 147.75. 186.4649. 82.. fitting by regression. see Distributions HyperGrdecoLati 11 squares. 114.124. 280 . 14 Fisher.71. see I>istributions.284 Hurdle. 55. %. 255 23. Hunter. J. 19. 31.235. 71. 30 normal. 271.G. 177..J.141. 203.72. 285 Distributions: beta. 113127.182 Student's I.108.25. 85.125127. 1151 I6 alpha risk.296 Fishbone diagram. E. 238 318.89 lognormal. 1321. 203. I39 F.6365.207. 0 . 303 Fourlevel factors. 118.69.102. 324325 Fractional factorial experiments. 186187.181 critical region. 59 Hinge. fitting by regression. Poisson.112 Gaussian.R.213214 H Halfnormal plot 295 Hazard rate. 6061 Dodge.S.62. 2223 Histogram. 285289.6 Hspread. H. 4546 hypergeometric. 1 Fence. 217. 214215. 8182. G o o d ~ ~ ~ . 78.186.77.292 EWMA.S. W. 197. J. Foldover designs. 15. 23 Finney. 136 Duncan.113 Exponential distribution.80. see Distributions Exponentially weighted moving average chart. 1 151 16 E Effect.25 Homoscedasticity. exponential. 227 Horizontal Vmask. 3738. 120. ("Student"). Sir Ronald A. 293 Hyperbola.
197. 166.18. 177. 179 sentencing. 1 1..337338 L(27). 45. 5657. K. 72. 71 Mathematical expectation. 332 Median.7678 Moving average charts. 6769. 249 Modal interval. Wilcoxon's test Marginal probabilities. 174 Linear estimator. 199200 Interaction.201 MILSTDl235B.58.. 201 tightened.we Stemandleaf diagrams Least squares. 156.320 L(l6). see Limiting quality (LQ) LSD (least significant difference). 337 Inspection level.370 Hypothesis testing (Continued) null hypothesis. 1 15 type I1 error. 132.303 Interaction plot. 118. 187. M Mann. 12. 1081 10 Mean.289.M. 41 Independent variables.299. 120 INDEX I I charts. 121 type 1 error. J. 101..2157.332 MILSI'Dl05D. see Confidence intervals Ishikawa. 170171 Leaf.. see Expectation Maximum likelihood estimates. see Mean square J John. 276. 177 disposition action. 198 Inner arrays. 45.327 L( 18).243. 74 Linear combinations of random variables. 6365. 197 Minitah.2223. 105 Lot. 299 L(9). 122. 78. 5961. 13 LCL (lower control limit). 179 Lot tolcrance percent defcctive (LTPD).189201 MILS'I'D414. H. 4 Kurtosis.107 generating function.78.75. method of. 261.218. 251.312. 223. 2. see Tests. 1151 16 sample size. 199 normal. 87. 134.304.M. 1161 17 power. 318 L(36). 203 MS. 15.6. 3 183 I9 L@). 183 LTPU. 159. P. 83. 22 Interval estimates. I80 LQ. 209 (Lo). 297. 7677. 106107 Moment generating function. 319 .68. 18. 70..W. . 89 Likclihood ratio. 319.80.see Lot tolerance percent defective (L'I'YD) Lucas. 1415. 323 Lattices. 101 Mean square.300301. 287 Intcrquartilc range. Lord.7173.M. 7678 Moments. . 6466. 191 Limiting quality 198 Limits. 70. 267 LSL (lower specification limit). 67 Juran. 910. 81 Indifference quality. J.284285 Joint density function. 106.77. 304. 158. 15 Moment estimators. 6466. 3 12 L( 12).256 Mean square error.179. I 15.71. 75 estimators.245 Lifetime. 160161 Independent events. 300301. 199200 switching rules.7 I . 120 rejection region. 24 K Kuizen. 87 L Latin squares. I99 rcduced. 4 Kelvin.
333 Plackett.247 Normal distribution. 274 Ott.R. 233 Residuals.. S. see Hypothesis testing N Q Quadratic blending model. J. E. 53. 223. 27 I Quartiles. 304.271 Pooled estimate of variance. see Distributions Polynomials: fitting by regression. I ? Performance characteristic. fitting by regression. see Hypothesis testing Reliability function. 204. see Rejection number Rectification. see Capability of a process Multinomial distribution. 253. 121 Oneatatime design 342 Onesided intervals. 295296 Normal scores. 178 Nonconformity. 296 Page. 59 Residual plots.290 Rejection number.S. 79 Kandom variable. 177 MVUE (minimum variance unbiased estimate). 113 Parameter design. 179. 146147 R charts. E.220.. 285 Nonparametric tests. 158 Neyman. 320 Plot: as an experimental unit.27 I orthogonal. see Tests Rational subgroup. 83. see Hypothesis testing Probability density functions. 258259 Ogive. 58 0 OC. 259 Quadratic model... 253 Qualitative factor. 209 Nelson. 213 R . 333340 Pareto diagrams.W. 10. see Distributions Normal plots. 91 Nonadditivity.S. 169. 277 Outer arrays. 129 Normal equations. 303304.172 Nondefective. 105 Nelson. See also Lattices Orthogonality. 8385 npcharts. 198 Orthogonal arrays. 337 Outliers. 131. 171. 125 Percentiles.23 I Resolution of a fractional factorial. 91 P Radians. 164. 255.294 Population. 164. 7.S. K. E. 209 Parameter. 271 Nonconforming. 88. 133 Parabolic masks. 18. 186 Regression. 287 Point estimates. 271 Quantitative factor... 297. 148151 Rc. 178 Nonorthogonal. 16.8890.INDEX 37 I Pearson. see Distributions Multiple sampling plans. 24 p charts. 171 Kandom samples. 1651 66 Null hypothesis.22.270. 120. 147. YO Poisson distribution. R. 67 Process capability. 264 interaction.74 Power.209 Paired ttest.340 Roberts. 1I7 Oneway layout. 93 Onesided tests. 13.R. 104.see Operating Characteristic curve Octane blending. I I12. 30 Range. 167170 Pearson. 52 conditional density functions. See d s o lnterquartile range Rank tests. 224. P. 179 Rejection region. 265 Operating Characteristic curve. L.S..73. 184..
910. 15 Statistic. 197 187. 18 Romig.201 MILSTD1235B.179. see Tests. 332333 T Taguchi. chisquare test for. 156 Scheffc. 87 Skiplot sampling plans. 217. set' Stopping rules for control charts Rules for switching between levels of inspection.331343 'Ially sheets. 177. normal. 137. H. see Bernoulli trials Trimmed mean. H.21 Signaltonoise ratios. 171 Treatment. 177178. H ... see Tests. 158. . 122. 13 Target value. 92 Twoway layout. 31 chisquare test.we Hypothesis testing S Sample. 187. 179187. 4. 101. 122. see Normal scores Screening experiments. . 230. 264 Trials.. 129. 13.9. 134 TINT. 197.203 Stratification.372 Robust design. 187 Smith.189201 MILSTD414. 191193.277 Standard deviation. R.W.238 Snee. 189 Rule. 101 Xolerancc design 333..229 Runs: experimental points. 171 U . 9.D. 2122.235. 74.. 114 UCL (upper control limit). 340 Torrey. W. 145.. 154. 187 Scatter plots.206. see Hypothesis testing Type I1 crror. 189 MILSTDIOSD.II. 177. 1251 27.. see Distributions Sum o squares. 129 Duckworth test.264 Wilcoxon's test. 57.18 TRMEAN.4. 158. 131. 174. 332333 Robust statistics. 133. 265 of points on quality control charts. 258. 301302 Sequential probability ratio test.284286. 129. 17. 191197 Seven tools. 197. J. 126 Ties in rank tests.209 [test. 39 Sampling plans: ANSVASQC Z1. 165. 2022. Ishikawa's. see 'rests 'Tukey.201 chain sampling plans.332 ztest. 197 skiplot sampling plans.229.332 Tests. 123. 13 Student's distribution. 258 Schilling. 133 rank tests.. 119120. 98. 39 with replacement. 136136. 156. 264 MaimWhitney test. 201 Rule of thumb. 131 Twosided intervals. 104. 25 Stopping rules for control charts. E. 76. 76 System design. 12. 12 Itest. Wilcoxon's test sequential probability ratio test. 204 INDEX Statgraphics.218 s charts.G. Transformations. M. 187. 186 Scores.297. G. stopping. 183. 251. 190 Skewness. 12 Stemandleaf diagrams.264 f Symmetric. see Tests Sequential sampling. 187 CSP (continuous sampling plans). 190 Sampling: without replacemcnt. 159. 13 Shewhart. 141 Twosample rtest. 228.136 2 X 2 tables. 186. I.G. 197 DodgeKomig sampling plans. 165.203. 335 Single sampling plans. 10 Sample size. 273 Type I error.N. 8. Wilcoxon's test paired ttest. 197 ANSVASQC Z1. 136 Ftest.
29. 342 yaks’ algorithm. F. See Distributions Upper confidence intervals. 132 Ftests for.. 131. 131. 10.75.38. 336 Vmask. 209 V xbar. coefficient of. 123.R. 132 BehrensFisher test. 159 Variables: concomitant 268 continuous. 336 use as SN ratio. Wilcoxon‘s test Wilcoxon’s test. 154. D. 2 17 Unbiased estimators. see Tests Wood. 295 1 Variance. see Tests .. 178 ztest. 146156 X Yates. 7. 93. 105 Unequal variances. 147 xbar charts. 284. 137 Uniform distribution. Yo xbarbar. F.294 Variation.292. 97102.137.294 estimation of. see Distributions Whiskers. 103 USL (upper specification limit). see Boxandwhisker plots Whitney.64..31 I.97102. 292 Y Z Zerodcfects. 9. see Tests.INDEX 373 Warning lines.78.71.3249 discrete.203 Weibull distribution.
This action might not be possible to undo. Are you sure you want to continue?
We've moved you to where you read on your other device.
Get the full title to continue reading from where you left off, or restart the preview.