Вы находитесь на странице: 1из 330
EMP Ill Using Imperfect Data Donald J. Wheeler Fellow American Statistical Association Fellow American Society for Quality SPC Press Knoxville, Tennessee Copyright © 2006 SPC Press Statistical Process Controls, Inc. All Rights Reserved Do Not Reproduce any material in this book by any means whatsoever without written permission from SPC Press SPC Press 5908 Toole Drive, Suite C Knoxville, Tennessee 37919 (865) 584-5005 Fax (665) 588-9440 www .spepress.com ISBN 0-945320-67-1 316 + xvi pages 175 figures 64 examples 40 data tables 8 appendices 25 reference tables 3 worksheets index 1234567890 vi Chapter 1 Contents About the Author Acknowledgments Preface Frequently Asked Questions 1.1 Questions Regarding Measurements 1.2. Four Ways We Use Measurements 13 A Model for Measurements 14 Observational Studies versus Experimental Studies 1.5. The Organization of This Book PART ONE: CHARACTERIZING THE MEASUREMENT PROCESS Chapter 2 Chapter 3 Chapter 4 Consistency Charts 2.1 Two Technicians 2.2. An Operational Definition of Measurement Consistency 2.3 Which Comes First? The Chart or the Adjustment? 24 Measurement Consistency for Destructive Tests 2.5 Rubber Rulers 2.6 Rubber Parts 2.7 Consistency Charts for Subjective Values 2.8 Summary for Consistency Charts Precision 3.1 Quantifying Precision with the Estimated Standard Deviation 3.2 Quantifying Precision with the Probable Error 3.3 Manufacturing Specifications 3.4 Summary for Consistency Charts Bias 4.1 All Bias is Relative 4.2 But Do] Have to Do the Graph? 43 Consistency of Bias 4.4 Comparing Two Procedures for Relative Bias 4.5 Summary for Part One: Characterizing a Measurement System xiii xiv evan u i 13 15 18 19 2 21 22 23 23 25 30 32 33 33 37 40 41 43 vii PART TWO: EVALUATING THE MEASUREMENT PROCESS Chapter 5 Chapter 6 Chapter 7 Chapter 8 viii Relative Utility and Short EMP Studies 5.1 The Intraclass Correlation Coefficient 5.2 Relative Utility Using Consistency Charts and Product Charts 5.3 A Note on Notation for Measurement Error 5.4 Short EMP Studies 5.5 Measurement Systems That Use Multiple Determinations 5.6 Summary Worksheet for Short EMP Study Basic EMP Studies: Identifying Nuisance Components of Measurement Error 6.1 Basic EMP Studies and Potential Nuisance Components 6.2 Checking for Detectable Bias: Main Effect Charts 6.3 Checking for Differences in Test-Retest Error 6.4 What to Look For in an EMP Study 6.5 Discrimination Ratios and Intraclass Correlation Plots 6.6 Summary of Basic EMP Studies Worksheet for Basic EMP Study Some Basic EMP Studies 7.1 EMP Study Number 33 7.2. An EMP Study for Truck Spoke Dimension No. 2 7.3 AnEMP Study for the Compression Test 7.4 An EMP Study for the Nanometric Gauge 7.5 The Manual Test Stand for Product T143 7.6 A Round-Robin Study for Silica-Producing Volatiles 77 Summary Two-Factor EMP Studies 8.1 Test Method 623 with Product 1108 8.2. Test Method 623 with Product 1108 in a Nested Study 8.3 A Nested Study for Test Method 623 with Product 1402 84 Summary 45 45 48 50 51 54 59 62 63 63 67 71 81 83 91 91 94 96 101 103 106 112 113 113 117 120 126 PART THREE: USING IMPERFECT DATA Chapter9 Chunky Data or Inadequate Measurement Increments 127 9.1 Chunky Data 127 9.2. Fixing Chunky Data 131 9.3. The Detection Rules for Chunky Data 133 9.4 Will the Standard Deviation Statistic Fix Chunky Data? 136 9.5 Chunky Data in an EMP Study 138 9.6 Summary 138 Chapter 10 Censored Data 139 10.1 Assays for Compound M 139 10.2 Life Testing 146 10.3 Compound Y 147 Chapter 11 Characterizing the Product Not Measured 49 11.1 The Problem of Representation 149 11.2 The Lot Fraction Nonconforming 150 11.3 Acceptance Sampling 152 11.4 Within-Batch Uniformity for Measured Properties 156 11.5 Batch-to-Batch Uniformity and Grand Lots 159 11.6 Quality Assurance with Count Data 161 11,7 Grand Lots versus Acceptance Sampling, 162 11.8 Summary 164 Chapter 12 Prediction 165 12.1 The Process Behavior Chart and the Two Voices 165 12.2 What Actions are Appropriate? 167 123. Use the Measurements to Adjust the Process? 168 12.4 The Only Reliable Basis for Extrapolation 170 PART FOUR: EXPLANATIONS Chapter 13 Chapter 14 Chapter 15 Chapter 16 How Many Digits Should You Record? 13.1 How the Measurement Increment Affects the Median Error 13.2. The Uncertainty in the Last Digit 13.3 The Twin Uncertainties in Every Measurement 13.4 Chunky Data and the Measurement Increment 13.5 Summary The Basis for Manufacturing Specifications 14.1 The Idea Behind Manufacturing Specifications 14.2 Watershed Specifications 143 The Bivariate Normal Model 144 The Probability of Conforming Product 144 Summary The Usefulness of Measurement Systems 15.1 A Simple Experiment 15.2. Detecting Shifts: How Signals Get Attenuated 15.3 Detecting Shifts: The Theoretical Probabilities 15.4 Attenuation of Signals from the Measurement System 15.5 The Ability to Quantify Process Improvements 15.6 Four Categories of Data for Process Monitoring 15.7 Assessing the Ability to Quantify Process Improvements 15.8 Summary of Relative Utility Gauge R&R Studies 16.1 A Gauge R&R Study 16.2 The “Percentages” of the Total Variation 16.3 The “Percentages” of the Specified Tolerance 16.4 The Number of Distinct Categories 16.5 The Guidelines 16.6 What Can You Learn From a Gauge R&R Study? 16.7 An Honest Gauge R&R Study 16.8 The Basic EMP Study for the Gasket Thickness Data 16.9 Summary Chapter 17 Concepts and Techniques Used 17.1. The Chart for Individual Values and Moving Ranges 17.2. The Average and Range Chart 17.3 Average and Range Charts with Experimental Studies 174 Detection Rules for Interpreting Process Behavior Charts 17.5 Bias Correction Factors 17.6 Degrees of Freedom and the Coefficient of Variation 17.7 The Analysis of Means 178 Interval Estimates Using Student's t-Distribution 17.9 For Further Reading APPENDICES Appendix 1. Glossary of Symbols Appendix 2. Bias Correction Factor Tables Table A.1 Table A.2 Table A.3 Table A.4 Bias Correction Factors €2,¢4, da, ds, dg Using R to Estimate Variances dy* Moments for Range Distributions do, d3, dg Moments for Distributions of the Standard Deviation C4, C5, C6 Appendix 3. Tables for XmR Charts Table A.5 Table A.6 Degrees of Freedom for the Average Moving Range Degrees of Freedom for the Median Moving Range Appendix 4. Tables for Average and Range Charts Table A.7 Table A8 Table A.9 Table A.10 Table A.11 Table A.12 Average and Range Charts based on R Az, D3, Dg, Ez Average and Range Charts based on R Ag, Ds, Dg, Es Degrees of Freedom for the Average Range Degrees of Freedom for the Median Range Alpha Levels for Average Charts Used in Lieu of ANOM or ANOVA Overall Alpha Levels for Upper Range Limits 239 239 241 243 245 246 247 248 250 252 258 259 260 261 262 263 264 265 266 267 268 269 xi Appendix 5. Tables for Average and Standard Deviation Charts Table A.13 Average and Standard Deviation Charts based on5 3, B3, By, Ex. 270 Table A.14 Average and Standard Deviation Charts based on Azo, By, Br, Eg. 271 Table A.15 Degrees of Freedom for the Average Standard Deviation Statistic 272 Table A.16 Degrees of Freedom for the Median Standard Deviation Statistic 273 Appendix 6. Tables for Range-Based Analysis of Means Table A.17 Analysis of Means (ANOM) 5% Critical Values 274 Table A.18 Analysis of Ranges (ANOR) 5% Critical Values 275 Table A.19 Analysis of Main Effect (ANOME) 5% Critical Values 276 Table A.20 Analysis of Mean Ranges (ANOMR) 5% Critical Values 280 Appendix 7. Miscellaneous Tables Table A.21 Student's t-Distribution 286 Table A.22. Median Errors in PE Units for Different Measurement Increments 287 Table A.23 99% Interval Estimates for Percent Nonconforming in Lot 288 Table A.24 Truncation Effect Scaling Factors TESF 294 Table A.25 Process Three-Sigma Multipliers P3SM 295 Appendix 8. Worksheets S The Short EMP Study 297 H__ The Honest Gauge R&R Study 299-301 B_ The Basic EMP Study Flowchart 303 B_ The Basic EMP Study Worksheet 304-309 Index 311 About the Author Donald J. Wheeler is a consulting statistician who had the good fortune to work with Dr. W. Edwards Deming and David S. Chambers. Dr. Wheeler graduated from the University of Texas, Austin, with a Bachelor's degree in Physics and Mathematics, and holds M.S. and Ph.D. Degrees in Statistics from Southern Methodist University. From 1970 to 1982 he taught in the Statistics Department at the University of Tennessee, Knoxville, where he was Associ- ate Professor. Between 1981 and 1993 he periodically assisted Dr. Deming with his four-day seminars. He is a Fellow of the American Statistical Association and a Fellow of the American Society for Quality. He has conducted over 900 seminars in sixteen countries on five continents. He is author or co-author of 25 books and over 140 articles. Through these sem- inars and books Dr. Wheeler has had a profound impact on companies and organizations around the world. siti Acknowledgments A large part of this book consists of the lessons I have learned from my students. Begin- ning in 1984, when Richard W. Lyday introduced me to the issues of evaluating the mea- surement process, I have had the privilege of teaching and working with engineers, scientists, and technicians as they have learned how to better understand their measurement systems. Over the years, as these students have shared their questions and their results, I have learned how to adapt and refine the earlier techniques to broaden their utility. In my efforts to make this book a practical guide I have enlisted the aid of others. First of all, there are the many individuals whose examples are included in this book. While confi- dentiality requires that many of these remain anonymous, this anonymity does not lessen my appreciation for their contributions. Next, there are those who directly or indirectly con- tributed ideas, suggestions, and no little bit of proofreading, to shape the final form of this book. Special thanks in this regard are due to James Beagle, Systein Evandt, Andrew C. Palm, and Ray Phillips for their careful reviews of the manuscript and many thoughtful suggestions. As practicing engineers, mathematicians, and statisticians they have all collaborated with the author at various times and in various ways to help to keep this book in the realm of that which is practical and useful. Finally, I would like to thank William H. Woodall for helping me clarify some of the material in Chapter Sixteen. While we have all worked to make sure that the material in this book is correct, there will inevitably be some inconsistencies that will have escaped the notice of us all. For these, or other questions, feel free to contact the author at < djwheeler @ spcpress. com >. Preface This project began as a revision of Evaluating the Measurement Process, Second Edition. This revision was planned to include the use of the Analysis of Main Effects and the Analysis of Mean Ranges as follow-up techniques for use with EMP Studies. It was also to include the use of the Intraclass Correlation Coefficient as a measure of relative utility rather than the obscure and poorly understood Discrimination Ratio. Thus, the project began with the simple objective of cleaning up a few loose ends and presenting some improved techniques. However, as the project progressed, the lessons learned in teaching this material during the seventeen years since the Second Edition was published kept finding their way into the manuscript. As I added these various insights to the manuscript it began to grow dramati- cally. Finally, when it grew to over 300 pages, it could hardly be said to be a revision of the original 100 page Second Edition. It had become its own book. So as this book grew from a planned Third Edition of Evaluating the Measurement Pro- cess into a comprehensive book on characterizing, evaluating, and using imperfect data, the title needed to be changed to reflect the new nature of the book. At the same time, I wanted to acknowledge the origin for this book, hence, EMP III Using Imperfect Data. Next came the problem of how to organize the book. The original Evaluating the Mea- surement Process was a how-to manual for conducting measurement process studies, and I wanted to maintain the accessibility of the earlier book. However due to the broader scope of this book some detailed explanations were necessary. In the early drafts these detailed explanations got in the way of the how-to manual. In the end I organized the book into four parts and placed these explanations in Part Four. Part One now contains a greatly expanded treatment of Characterizing the Measurement Process with separate chapters on Consistency, Precision, and Bias. Part Two focuses on EMP Studies with separate chapters on Short EMP Studies, Basic EMP Studies, and Two-Factor EMP Studies. Part Three is Using Imperfect Data. This part includes chapters on Chunky Data, Censored Data, and issues surrounding the extrapolation involved when we use the product we have measured to characterize the product not measured. Part Four explains how many digits should be recorded, and why the Probable Error defines the optimum size for the Measurement Increment. It describes how to establish Manufacturing Specifications, and it defines how to characterize every measurement system as a First, Second, Third, or Fourth Class Monitor. Chapter Sixteen gives a clear explanation of what is happening with a Gauge R&R Study and provides an Honest Gauge R&R Study as. an alternative. Finally, for those who might be rusty on some of the techniques used throughout the rest of the book, Chapter Seventeen provides a quick review of both the concepts and the mechanics of the various procedures used. xv xvi “An element of chance enters into every measurement; hence every set of measurements is inherently a sample of certain more or less unknown conditions. Even in those few instances where we believe that the objective reality being measured is a constant, the measurements of this constant are influenced by chance or unknown causes.” W. A. Shewhart No two things are alike, but even if they were, we would still get different values when we measured them. INTRODUCTION Chapter One Frequently Asked Questions This chapter introduces some concepts that will be used throughout this book. The first three sections provide important information about the scope, purpose, and notation of this book, and should be considered to be required reading. They provide the foundation upon which the rest of the book is built. Section 1.4 makes a distinction between experimental studies and observational studies and the techniques of analysis available for each. Section 1.5 describes the organization of the remainder of the book. 1.1 Questions Regarding Measurements Answers to some of the most frequently asked questions about measurement systems are given and explained in this book. Among these are the following: 1. Are the measurements consistent over time? The only way to answer this question is to repeatedly measure the same thing over time and to know how to interpret the values obtained. A Consistency Chart is an operational def- inition of consistent measurements. It is explained in Chapter Two. 2. What is the precision of a measurement? 3. How many digits should be recorded? 4. Just what is the effective resolution of a measurement? 5. How cana single reading, or the average of several readings, be used to characterize the measured item relative to specification limits? The answer to these questions requires a way to quantify measurement uncertainty and a way to obtain meaningful Manufacturing Specifications. Measures of uncertainty and their uses are covered in Chapter Three, with further explanations provided in Chapters Thirteen and Fourteen. EMP III Using Imperfect Data 6. Are the measurements biased relative to some standard? 7. How can you compare different techniques of measuring a particular characteristic? Chapter Four outlines simple ways to use XmR Charts and Interval Estimates to detect and evaluate biases relative to standards and between measurement methods. 8. How can we characterize the relative utility of a measurement system for characterizing a particular product? 9. What does relative utility mean in practice? 10. When are measurements not satisfactory for monitoring a production process? 11. How much information about the product is contained within the measure- ments? 12. How much of the variation in the measurements for a product stream is due to measurement error? 13. Can we detect detrimental changes in the production process when they occur? In order to answer these questions we will need a practical measure of relative utility that is theoretically sound. Such a measure is presented in Chapter Five. In Chapter Fifteen this measure is used to develop a framework for classifying the utility of a measurement system for use with a given product. 14. Do the measurements display detectable differences from operator-to-operator, machine-to-machine, lab-to-lab, or day-to-day? This will require a way to collect and analyze data to look for the presence of various nui- sance components of measurement error. Such methods are presented in Chapters Six, Seven and Eight. 15. Are the measurements being made with Measurement Increments that are small enough to properly reflect the process variation present? 16. How can I place censored data on a Process Behavior Chart? When the Measurement Increments are too large, the measurements will be chunky and this chunkiness will be detectable on Process Behavior Charts that are based on these mea- surements. Chunky Data and Censored Data are covered respectively in Chapters Nine and Ten. 17. Can you use measurements to characterize product which has not been measured? 18. When can you use measurements to make adjustments to a production process? 1/Frequently Asked Questions 19. When can you meaningfully “accept” or “reject” batches of product based upon one or more measurements? Questions 17, 18, and 19 deal with the use of measurements to characterize and make decisions regarding items that have not been measured. The basis for this extrapolation and the issues involved are discussed in Chapters Eleven and Twelve. 20. What can we learn from a Gauge R&R Study? Chapter Sixteen describes the steps and missteps of this procedure in detail. These questions cover the most pressing issues concerning measurement processes. They are directly concerned with the utility of a measurement process in a specified context, ie., for a specific product. Because the questions are focused in this manner, the answers will often clarify the issues, making significant improvements possible. By focusing on the use of a specific measurement process for a specific product we will narrow our horizon, and within this range of interest it will generally be sufficient to assume that there is only one level of precision and only one level of bias. Moreover, the questions of greatest interest to most practitioners can be handled within this narrow context. Since the purpose of this book is to give simple answers to the most pressing questions, rather than to give detailed and complex answers to solve global problems, some of the traditional hoary chestnuts are completely missing from this book. In particular, the issues of “linearity of accuracy” and “constancy of error” are missing from the list above. This omission was deliberate. Linearity of accuracy is concerned with whether the bias of a measurement system remains the same across a range of measurement values. Constancy of error is concerned with whether the measurement uncertainty remains the same across a range of measurement values. While these broader questions concern the behavior of the measurement process across a wide spectrum of applications, they also require more sophisticated experiments and more complex analysis techniques. Since most industrial applications consist of using a measurement system within a narrow range for a specific product, answers for these global questions are seldom needed in practice. Finally, since there cannot be one answer to all the questions above, simplistic approaches which use a single summary value to characterize the acceptability of a measurement process can never be satisfactory, or even correct. 1.2 Four Ways We Use Measurements Why do you measure something? Since obtaining a measurement takes time and energy, there should be some payback for this effort, some utility gained. To understand this transac- tion and the problems entailed therein it is helpful to make a distinction between four differ- ent uses of measurements. EMP III Using Imperfect Data DESCRIPTION One use of a measurement is to describe the item measured. The need to describe might be motivated by idle curiosity, by a need to be informed, or by a desire to have the data for future use; but whatever the motivation, this basic use of a measurement is to answer the question of how many or how much. When a value is used to describe the item measured it is important to understand the limitations of the value used—the uncertainty attached to the value itself. The source of this uncertainty will primarily be the uncertainty in the measure- ment process. Thus, in order to use data to describe the item measured we will need to have a way to define the uncertainty that should be attached to a measurement. CHARACTERIZATION A second use of a measurement, closely related to the first, is to characterize the mea- sured item relative to specifications. Here the measurement is used to take action on the item measured—to characterize the measured item as being within the specifications or outside the specifications. Once again the major source of uncertainty will be the uncertainty intro- duced by the measurement process. However, instead of attaching our uncertainty to the measurement, we will now need to make adjustments in the specifications in order to make allowances for this uncertainty when we characterize the measured item. Since the common questions surrounding Description and Characterization are easily answered by understanding the variation of the measurement process, the material in Chapters Three and Four will go a long way to clarifying issues with these two uses REPRESENTATION A third use of a measurement is to represent the product that was not measured. Here the objective is to characterize the product stream with regard to specifications. In order to do this we have to extrapolate from the product that was measured to the product that was not measured. Of course the basis for such an extrapolation will be the selection procedure used to obtain the product that we did measure. When a measurement is used to represent items not measured there are several sources of uncertainty. The uncertainty in the measurement process is still present, but now we also have the additional uncertainties related to the repre- sentation. The first of these comes from the product variation: since the product not mea- sured will differ from the product measured we have an additional source of uncertainty not present in the description of a measured item. Additional uncertainties of representation will come from the selection procedure (or the lack thereof in some cases). We will need to have a way of making allowances for all of these uncertainties before we can effectively use data to represent the product not measured. PREDICTION A fourth use of a measurement is to predict. Here the objective is to characterize future process outcomes. The mechanism used is the characterization of the past behavior of the 1/ Frequently Asked Questions process. Here we are no longer concerned with classifying some batch of product as con- forming or not, but rather with understanding what to expect in the future. When data are used to predict we must be concerned with measurement error, the predictability of the pro- duction process, and the combined variation from both sources. Thus, in addition to the sources of variation identified above, there is also the uncertainty associated with the contin- uation over time. Since both Representation and Prediction involve extrapolation from the product mea- sured to the product not measured, or not yet produced, a sound basis for this extrapolation will be required. This basis and related issues will be covered in Chapter Eleven and Twelve. Moreover, since the way you intend to use a measurement will have an impact upon how much you need to know about the sources of uncertainty in that measurement, it is impossi- ble to define a single, simple index for the quality of a measurement system. There are many different aspects of any measurement system. Some of these will affect some uses of data and will have little impact on other uses of data. This book presents simple and effective techniques that will allow you to identify those things that can hinder your use of measurements. Rather than worksheets with canned for- mulas and arbitrary guidelines given without motivation or explanation, the emphasis throughout has been upon developing an understanding of techniques of analysis that are both theoretically sound and of practical use. However, in the interest of facilitating the use of EMP Studies, a series of worksheets are provided in the Appendix. 1.3. A Model for Measurements In order to write and think clearly about measurements and measurement error it is important to have a well-defined structural model. Therefore, throughout this book the following model will be used: a measurement, M, will consist of two major components; the product value, P, and an error term, E. It is the error term that will contain the contributions from the various sources of measurement uncertainty described above. Measurement = Product Value + Measurement Error M=P+E At some later stage we may break the measurement error term down into components, but for now this simple model will be sufficient. For a collection of measurements the average value will be: Average Measurement = Average of Product Values + Average Measurement Error r, in terms of probability models: Mean(M) = Mean(P) + Mean(E) Since we would like for the Mean of the Distribution of M to be equal to the Mean of the Dis- tribution of P, we need for the Mean of the Distribution of E to be zero. When this happens 5 EMP III Using Imperfect Data the measurements, M, will be said to be “unbiased.” When Mean(E) is not zero the product measurements, M, will be biased. Thus, we shall assume that the bias of the measurement system depends upon Mean(E). Ina similar vein, we shall assume that the variation of the product measurements can be characterized by: Variance(M) = Variance(P) + Variance(E) On? = Op + oF The precision of the measurement system will be concerned with the value of the Variance of the Distribution of E. Since Pierre Simon Laplace published his famous theorem in May of 1810 the classic model for the Distribution of E has been the normal distribution. But before we worry about bias, precision, or the distribution of E, we need to answer a more fundamental question, is the measurement system predictable? This property will be referred to as measurement consistency. Thus consistency, precision, and bias all refer to properties of the Distribution of E. * Consistency refers to the predictability of a measurement system. * Precision refers to the variation of a predictable measurement system. * Bias refers to the mean value for the errors introduced by a predictable measurement system. Without a consistent measurement system precision and bias do not exist as well-defined properties. Consistency, precision and bias are of interest when data are used for Description, Characterization, Representation, and Prediction. Product _ Product | Measurement Measurements Values * Errors M=P+E A A A Consistency = Predictability of Measurement Process \ | Precision = Variation of consistent Measurement Process \ Bias = Mean of consistent Measurement Process ‘The Predictability of the Production Process \ is concerned with the consistency of the stream of Product Values, P But we have to work with the consistency of the stream of Product Measurements, M, in spite of the uncertainties introduced by the Measurement Errors, E. ‘Mean(M) = Mean(P) + Mean(E) Variance(M) = Variance(P) + Variance(E) Figure 1.1: A Structural Model for Thinking About Measurements 1 / Frequently Asked Questions When we talk about the predictability of a production process we are interested in the behavior of the product values, P. However, since these values are not directly observable, we have to work with the product measurements, M. When Variance(P) is large relative to Variance(E) this shift is not problematic. But as Variance(P) gets smaller relative to Vari- ance(E), the ability to use the product measurements to characterize the predictability of the production process becomes more dependent upon having a consistent measurement system. Thus, when data are used for Prediction, some measure of the relationship between these two variances will be needed. Another aspect of a measurement is the amount of round-off used when the measure- ment is recorded. The number of digits recorded, and occasionally, the rounding used for the last digit, will define the smallest Measurement Increment. Usually this increment is com- pletely unrelated to the distribution of the measurement errors. As we will see the size of the Measurement Increment can have an impact upon our ability to detect the variation in our product measurements. When this happens the round-off will affect both the quality and the usefulness of our measurements. As needed by the various topics, this structural model will be used throughout this book. 1.4 Observational Studies versus Experimental Studies This book will describe the analysis of data arising from both observational studies and experimental studies. Observational studies are studies where the data are obtained as a by- product of some on-going operation. These data may be deliberately and intentionally col- lected, but they are still a by-product of some process while that process is being operated in an ordinary manner, ie., the process is not being deliberately manipulated in order to collect specific data under different conditions. In other words, in an observational study, the data track the process. On the other hand, experimental data are collected under different conditions where those different conditions were created for the express purpose of obtaining the data. Thus, experimental data will always consist of a fixed amount of data, collected under different conditions, while observational data will consist of data from an ongoing stream, usually col- lected while the known inputs are held constant. Because experimental data are collected under special conditions, they tend to be more expensive than observational data. Moreover, because of the way they are obtained, we expect experimental data to represent the differences between the special conditions. Thus, when we analyze experimental data we are looking for differences that we have paid good money to create and that we believe are contained within the data. Moreover, the fact that we will have to conduct more experiments if we can not find the expected differences will tend to make us choose a less conservative, and more exploratory, analysis. EMP III Using Imperfect Data Table 1.1: Observational Studies vs. Experimental Studies Observational Studies Experimental Studies Additional Data Available Fixed Amount of Data One Condition Present Two or More Conditions Present Should Be No Signals Should Be Some Signals Sequential Analysis Procedure All Data Analyzed at One Time Conservative Analysis Traditional or Exploratory Analysis ‘When analyzing data from an observational study, the fact that the data were supposedly collected under one condition will mean that we do not expect to find any differences within the data. Furthermore, since differences will often indicate unplanned changes in the pro- cess, we will want to be sure about any differences we find. Since additional data will usu- ally be available, we can afford to play a waiting game with observational studies. As each new observation is added to the chart the limits provide a conservative analysis with a very small risk of a false alarm, so that we will have strong evidence of a change before taking action. The Individual Value and Moving Range Chart (XmR Chart) and the Average and Range Chart were created for use with observational studies. When they are used in this way they may be said to be “Process Behavior Charts.” Whenever this phrase is used in this book it will be referring to an observational study. ‘As seen in Figure 1.2 there are several different analysis techniques that can be used with experimental studies, while some techniques may be used with either type of study. A brief description and review of several of these analysis techniques and related concepts is given in Chapter Seventeen. Observational Studies Experimental Studies | _ | Average & ‘Average and Range Charts Range (used as Process Behavior Charts) Charts — Individual & Moving Range Charts (used as Process Behavior Charts) Figure 1.2: Analysis Techniques for Observational and Experimental Studies 1/Frequently Asked Questions 1.5 The Organization of This Book The remainder of this book is divided into four parts. Parts One and Two will serve as a guide to actually performing various types of measurement system analysis, while Part Three examines the broader problems of using data in practice. Part Four contains related explana- tions and supporting documentation for the techniques given in Parts One and Two. In Part One, which consists of Chapters Two, Three and Four, the topics of consistency, precision and bias are discussed. Absolute characterizations of the quality of a measurement system are developed and explained, and the role of a Consistency Chart is described. Part Two describes the Role of EMP Studies and consists of Chapters Five through Eight. Short EMP Studies are presented as ways of characterizing the relative utility of a measure- ment system for a given application. Basic EMP Studies let you characterize the measure- ment process and also consider the effects of a potential nuisance component of measurement error. Two-Factor EMP Studies let you characterize the measurement process while you simultaneously consider the effects of two potential nuisance components of measurement error, Six case histories for Basic EMP Studies are given in Chapter Seven, Chapters Nine through Twelve make up Part Three. Under the heading of Using Imper- fect Data these chapters consider the effects of Chunky Data, how to use Censored Data to construct meaningful Process Behavior Charts, the problem of characterizing the product that was not measured, and the problem of prediction. Part Four contains the detailed explanations behind some of the techniques presented in earlier chapters. Chapter Thirteen answers the question of how many digits are appropriate for any measurement. Chapter Fourteen describes the origin of the rules for establishing Manufacturing Specifications. Chapter Fifteen justifies and explains the relative utility cate- gories of First, Second, Third and Fourth Class Monitors. Chapter Sixteen presents a detailed explanation of the auto industry's Gauge R&R Study Finally, the techniques and concepts used throughout this book are summarized and reviewed in Chapter Seventeen.

Вам также может понравиться