SURVEY SAMPLING Leslie Kish Professor of Sociology Program Director, Surcey Research Center Institute for Social Research The University of Michigan JOHN WILEY & SONS New York + Chichester + Brisbane + TorontoCopyright © 1965 by John Wiley & Sons, Inc, All Rights Reserved. Heproduction or translation ot any part of this work beyond that permitted by Sections 107 or 108 of the 1976 United States Copyright Act without the permission of the copyright owner is unlawful. Requests for permission or further information should be addressed to the Permissions Department, John Wiley & Sons, Inc. 12:13 14:15 Library of Congress Catalog Card Number: 65-19479 Printed in the United States of America ISBN 0 471 48900 xPreface This is designed to be a simple book on sampling methods, with emphasis on and illustrations from surveys of human populations. Although I have designed samples for a variety of nonhuman populations, my experience has been mostly in social applications; the book emphasizes these areas, and it draws most of its examples from them. However, essentially the same methods can be used for sampling animals, plants, minerals, physical products, accounts, and inventories. Sampling special- ists find that their skills possess broad generality and transferability; thus this book can also be useful to medical men, biologists, chemists, engineers, and accountants. Sampling plays a vital role in research design involving human popula- tions: it commands inereasing attention from social scientists and practitioners. 1 include not only the disciplines of economics, sociology, anthropology, psychology, and political science, but also the professions of public health, biostatistics, education, social work, public admin- istration, and business administration. Sampling problems are equally material to practitioners engaged in marketing, commerce, and industry. While writing this book, I tried to keep a definite audience in mind and to maintain a dialogue with them. My first consideration was for quantitatively oriented students in the social sciences and allied fields. T believe that sampling methods should become roughly the third course of statistics for students in the social sciences. It should follow a course in the fundamentals of statistical reasoning, and another devoted to major statistical tools. | hope that this book can serve as a text for professors of social statistics who are not specialists in sampling. Second, I have also aimed to provide a reference book on samplingvi Preface methods for professors, researchers, and officials who want to understand survey sampling without necessarily becoming sampling. specialists. Improving their methods is crucial, since most samples in survey research represent occasional small or medium-sized efforts by researchers who are not specialists. T have not written this for my colleagues in sampling, but chiefly for my [ellow social scientists, whose specialties differ from mine, Reading one book will not turn anyone into an expert sampling statistician; but the careful reader should become able to design and execute valid samples of moderate dimensions and difficulty, to avoid selection biases, and to achieve reasonable efficiency. He should also become more adept at evaluating the sample results he encounters, to judge their validity, their mits of inference, applicability, and precision. The reader will be able to design satisfactory rather than optimum samples—to satisfice (as Herbert Simon says), rather than to optimize. T paid less attention than others to some technical efforts for getting the last 2 percent efficiency out of samples. These efforts may be worthwhile for large, complicated samples—where it may be best to consult an experienced sampling statistician, In that case, the researcher with sufficient knowledge to meet him half-way will be better able to utilize the scarce time of the sampling consultant, Social research places special emphasis on the comparative and analytical uses of samples. Hence, I have stressed statistics of subclasses in the sample, rather than focusing all attention on the entire sample. Subclasses are treated for each of the major designs; and the comparison of subclass means receives frequent attention. Other issues relating to analytical statistics from complex samples are also discussed. These discussions frequently utilize research developed under a grant (G-7571) from the National Science Foundatio: The book is oriented toward providing a working knowledge of practical sampling methods, with an understanding of their theoretical background. The necessary working formulas are given. I have tried to state clearly the assumptions underlying the formulas, and to outline the potentialities and limitations inherent in the assumptions. I often use illustrations to explain in detail the meaning of formulas and definitions. T also provide a variety of examples, with computations laid out in painstaking detail. The book contains many practical procedures, the “domestic arts” of sampling along with its science; the valuable “tricks’ that are usually learned only in apprenticeship. Perhaps the principal task of any book or course on methods should be the drastic shortening of the period of apprenticeship in a discipline. Throughout the book the reader canPreface vit discover my attempts to synthesize complete sets of coherent rules for conduct ordinarily explained—if at all—with ad hoc rules-of-thumb, My efforts are perhaps most evident in methods for dealing with frame problems. They receive a great deal of emphasis, especially in an carly summary (2.7) of frame defects and techniques for treating them, and in a detailed treatment in Sections 11.1 to 11.4. Throughout the book the reader's attention is called to possible frame defects and their effects on sample design. Problems invite the reader to participate in finding and overcoming defects he is likely to encounter in research situations, Sampling methods are not developed here for their own sake, but as means to ends originating in substantive research problems, especially in the social sciences and their applications, They are designed to provide valid, scientific, and economical tools for those research problems, Hence, the approach is directed less toward delivering a neat theoretical package than toward practical research projects in all their complexity, while preserving « fundamental simplicity of presentation. To achieve simplicity I frequently resort to useful approximations. Briefly, this book attempts to preach what reasonable sampling statisticians actually practice, rather than an arid dogma. For example, the factor (N — DJN, if inconvenient, can be ignored when the population size N is large. The book emphasizes estimated variances that the researcher actually computes, rather than their expected values. And so on. Tdid not try to make this book “self-contained"—an illusory aim. For the algebraic steps of derivations the interested reader is directed, by frequent references, to relevant passages in several fine textbooks. In these references I sought neither to discover original sources, nor the prestige of the most elegant mathematical treatment, Instead, | sought the most relevant, readable, and available sources, preferably textbooks; there the reader can find further references. However, | am glad that, on the urging of friends, | abandoned my original intention to avoid derivations completely. My derivations are few and simple, yet they cover the essentials of sampling theory adequately for readers who understand certain fundamental concepts of statistical theory. Sampling theory was condensed by eliminating some repetition and detail, The basic derivations are in Sections 2.8, 4.5, 4.6, 5.6, 6.6, 8.5, and in scattered remarks, These and other technical portions carry a (#); continuity can be maintained without them. My original intention was to formalize and disseminate the notes that my students and I have evolved from teaching a one-semester course. This restricted but vital core is presented in the simple, nontechnical, carefully organized and related sections of Chapters 1 to 8 in Part 1: Fundamentals of Survey Sampling. To complete the course, [ addviii Preface Sections 9.1 to 9.5 on area sampling, plus most of Chapter 13 on non- sampling errors. Brief selections may also be added from Chapters 11 and 12, which deal with special techniques. These two chapters and the technical portions of others have served as the base for advanced courses in special problems. However, they are designed primarily to serve as sources of reference for individual, technical problems, These sections are mostly self-contained, rather than inter- connected. 1 hope that researchers will often find in one of them ready answers to specific technical problems. Those who can avoid the complexities of cluster sampling will find a self-contained treatment of element sampling in Chapters I to 4. To this can be added a modest treatment of the essentials of cluster sampling: 6.4B, 6.5A and B, 7.2, 7.3, and 9.1 to 9.3. These could comprise a third of a course, incorporated into a full year course in statistics. For a briefer introduction to sampling, teachers have used the followiag sections of 81 pages: 1.0-1.7, 2.6, 2.7, 3.1, 3.4, 3.5, 5.1, 5.4, 6.4B, 6 5A, 7.2, 7.3, 9.1-9.3. {hope that readers will especially like the problems. They should give much credit to my students, who eliminated the worst, stimulated the best, and sharpened their phrasing; also some credit to my many colleagues and clients, whose questions and research projects gave birth to most of these problems. I tried hard to make them life-like; hence, the student's crucial task is to discover what the principal issues are, then pose appro- priate questions. When this is well done, the answers come rather smoothly. The focus is on principal issues, rather than on fine nuances. This was made possible by eliminating cumbersome details, while retaining the essence of actual problems, In addition to testing the student's ability to apply the book’s methods, these problems perform a double task. They deepen understanding of those methods by extending them to new and varied areas. Furthermore, the extensions themselves represent valuable methodological tools. Work on these problems should yield half the value the student can extract from this book. Of a subject devoted to methods, we can say with Sophocles: “One must learn by doing the thing; for though you think you know it, you have no certainty, until you try.” It is with genuine pleasure that | acknowledge my debt to many friends and colleagues. Since Yates’ book in 1948, six fine books on sampling have appeared in English, and { have learned and taught from each in turn, Deming, Cochran, Hansen, and Hurwitz have been personal teachers, friends, and consultants. 1 was fortunate to have, at different times, Roe Goodman, Benjamin Tepping, and Irene Hess as colleagues in the Sampling Section of the Survey Research Center. Angus Campbell and Rensis Likert, in creating and directing that Center and the Institutefor Social Research, have provided valuable support and opportunity. Questions, problems, and issues posed daily by the social scientists at the Center provide a uniquely stimulating atmosphere. 1 also benefited from daily contact with colleagues in the Sociology and Mathematics Depart- ments. Of the many statisticians who have given me personal stimulation, support, and criticism, 1 must single out L. J. Savage, Bruce Hill, and Howard Raiffa; William Ericson gave me valuable criticism on much of the final draft, tightening many a loose phrase or formulation. They stimulated and helped me to make this book more suitable as a text in departments of statistics than I first planned. V. K. Sethi and C. T. Tharakan eliminated many technical errors. Above all, I was fortunate to find in my wife, Rhea, an expert and devoted editor, who has greatly improved the book’s style. Primary and essential stimulation came from the many students in about forty courses in sampling 1 have given in fifteen years. They represent @ good cross section of researchers, teachers, practitioners, and students from the social sciences, mathematics, and allied fields. The students in the seventeen yearly Summer Institutes on Survey Techniques have been mostly officials, scientists, and professors; they came from all over the globe, representing yearly an average of ten countries. ‘Their experience, searching questions, suggestions, and arguments have con- tributed substantially to the origin, form, and content of this book. Leslie Kish ‘Ann Arbor, Mich. June 1965Contents Part _ FUNDAMENTALS OF SURVEY SAMPLING 1, Introduction . 10 Mw 12 13 14 1s 16 17 Guide to the Introduction . Survey Design and Sample Design A Taxonomy of Survey Units and Concepts Population Values and Statistics Statistical Inference in Surveys ‘A Taxonomy of Selection Procedures Criteria of Sample Design toe Mechanical Selection, Randomization, and Frames 2. Basie Concepts of Sampling. _- 20 2 22 23 2 25 26 27 728 Some Basic Symbols Procedures for Simple Random Sampling (SRS) Mean and Variance of SRS... : oo The “fpe”; Sampling without Replacement Proportions; the Means of Binomials Relative Error : The Design of Economic Samples Frame Problems Basic Formulas and Derivations fo 2.8A Some Fundamentals 2.8B Variance in Simple Random Sampling Without Replacement 28C The Expectations of Some important Types of Samples 2.8D Variances for Linear Combinations 3. Stratified Sampling =. Bo 3 32 33 34 Definition and Purposes ‘The Weighted Mean and Its Variance . Mean and Variance for Stratified Element Sampling Proportionate Sampling of Elements 3.4A The Mean and Variance of Proportionate Samples | 3.4B The Design of Proportionate Samples 3.4C Post-Stratification or Stratification after Selection . 49 @ 7s 15 1 80 82 82 86 xixii Contents 3.5. Disproportionate Sampling or Optimum Allocation 3.6 Forming the Strata 3.6. Classifying Sampling Units 3.6B Stratification with Random “Quotas? 3.6C Errors in Sorting 3.6D Overlapping Lists 3.6E Objectivity and Regubirity Unnecessary 3.6F Homogeneity within Strata 3.6G Utilizing Available Variables 3.6H Elaborate Controls Unnecessary. 3.61 Number of Strata “3.61 Effect of Increasing Numbers of Strata 3.6K Equal Allocation : *3.6L. Optimum Stratification 4, Systematic Sampling: Stratification Techniques 4.1 Procedures and Uses of Systematic Selection 4.1 Uses of Systematic Samples 4.1B Problems with Intervaly : 4.1C Variances for Systematic Samples 4.2. Problems of Systematic Selection . 43 Paired Selections 4.4 Replicated Sampli 4.5 Subclass Means, Totals, and Comparisons in Stratified Samples, 4.54 Subelass Means, Totals, and Their Variances *4.5B Derivations of Variances for Stratified Subclasses. *4.5C_ Optimum Allocation for Subclasses *4.6 Precision of Proportionate and Disproportionate Samples 4.6A Precision of Proportionate Samples 4.68 Precision of Optimum Allocation 5. Cluster Sampling and Subsampling 5. Nature of Clusters 5.2. Random Choice of Clusters. 5.3 Subsampling in Multistage Sampling 5.3A The Aims of Subsampling 5.3B Features of Simple Replicated Subsampling SA Effects of Clustering: Intraclass Correlation, Roh 5.5. Stratified Cluster Sampling . 5.6 Components of the Variance 5.6A Components of Two-Stage Random Sampling 5.6B The Intraclass Correlation 5.6C Tables for Analysis of Variance 6. Unequal Clusters. 6.1 Problems of Unequal Clusters . 62 Random Selection of Unequal Clusters; Epsem Subsampling 6.3 Variances for Random Unequial Clusters 6.4. Stratified Sampling of Unequal Clusters 644 Any Number (a,) of Clusters 6.4B Paired Selection of Clusters 6.5. Statistics for Stratified Unequal Clusters : 6.5A The Difference of Two Ratio Means : 92 98 98 99 99 100 100 100 tot 102 102 102 104 m3 113 1B 1s 17 120 123 127 132 132, 136 138 139 139 142 18 148 1st 154 154 156 161 16s 166 166 170 173 182 182 184 187 190. 190 193 195 195Contents 6.5B_ Simple Variances for Complex Samples 6.5C_ Variances for Systematic Primary Selections 6.5D Expansions with Ratio Estimates . GSE Separate Ratio Estimates... 6.6 The Ratio Mean 6.6A. Variance of the Ratio Mean 6.68. Bias of the Ratio Mean 6.6C The Difference of Two Ratio Means 6.6D The Produet yr and Relvariances 7. Selection with Probabilities Proportional to Size Measures (PPS) 7.1. Control of Sample Size . 72. Control of Subsample Size with PPS. : 7.3. Paired Selection of Primaries TA. Selecting Primaries without Replacement 7.4A Single Selection from Random Half-Strata *7.4B Selecting Two Primaries without Replacement : 7.A4C Systematic Selection of Primaries with PPS. 7AD Single Selection from Each Stratum “TAE Three or More Primaries per Stratum 75 Probabilities Proportional to Size; Problems and Procedures 7.54 Selection with PPS Measures 7.3B.Subselection within Primary Units 7.5. Variations in Actual Subsample Size 7.5 Sampling Fractions, Stata and Measure Adjustments 7.SE Undersized and Oversized Clusters 7.5F_ Fractional Systematic Selection Intervals 8. The Economic Design of Surveys 8.1 Planned Precision Based on Unit Variance 8.2 Estimates of Unit Variance; Design Effect 8.3 Models of Cost Functions 8.3A A General Model of Cost Factors 8.3B Specific Cost Function for Cluster Samples 84° Practicality : : 8.4A Simple Designs . 8.4B Practical Field Instructions 8.4C Random Numbers for Field Work 8.4D Sequential Control of Sample Size "8.5 Optimum Designs : *8.6 Techniques for Computing Variances 8.64. Replicated Subsampling 8.6B Collapsed Strata 8.6C Special Techniques for Variance Computations 8.6D Precision of Variance Fstimates 8.6E Components of the Variance Part Il SPECIAL PROBLEMS AND TECHNIQUES 9. Area Sampling 9.1 Area Frames for Dwellings. 9.2 Preparing Maps . 197 202 203 206 206 206 208 209 2u1 27 217 220 23 227 227 29 230 231 232 234 234 238 239 2at 243 246 254 254 257 263 263 268 2m mm 214 276 207 279 282 282 283 286 289 291 299 361 301 304xiv Contents 93 94 98 9.6 97 98 Selection Rates for Blocks and Dwellings 5. ‘Some Practical Problems foe 9.4A Compact Segments versus Listed Dwellings 9.4B Problems of Block Size 9.8C Supplements for the New, the Missed, and the Unusual 9.4D Repeated Selection from a Listing 9.4E Modification to Three or More Stages 9.4F Maps, Photos, and Similar Materials Compact Segments 9.54 Procedures for Compact Segments : 9,5B_ Instructions for Segmenting : 9.5C_ Instructions for Interviewing in Segments 9,5D Segmented Listing Listing Dwellings 9.6A Tastructions for Listing Dwellings 9.6B Selecting Listed Dwellings 9.6C Identifying Dwellings at Sample Addresses Creating Segments from Listed Buildings 9.7A Special Features 9,7B Instructions for Selecting Blocks 9.7C Instructions for Listing Buildings 9.7 Segmenting Blocks and Selecting Segments 9,7E instructions for Sampling the Segments ‘Sampling City Directories 98A City Directories as Frames for Dwellings 9.8B Selecting from Directories . 9.8C Instructions for Sample Dwellings at Selected Addresses 9.8D Instructions for Block Supplement for Missed Dwellings 10. Multistage Sampling =... 10.1 10.2 103 10.4 Three-Stage Area Samples A National Sample of Primary Areas (Counties) Sampling a County 10.3A Measures for Major Strata 10.3B. Uniform Design for All Strata 10.3C Separate Procedures for Strata 10.3D The County as Part of a National Sample The Current Population Survey 11, Sampling from Imperfect Frames mW 12 Empty Listings and Foreign Elements; Variable Frame Densities Duplicate Listings: Overlapping Frames 11.24 How to Avoid the Problem 11.2B Creating Unique Listings 11.2C Weighting the Selection with I/p, 11.2D Matching from Overlapping Lists Selecting Persons from Dwellings; Small Unequal Clusters L1.3A. Dwellings as Clusters of Persons. 11,3B Selecting One Adult from the Household 11.3C Epsem Selection of Persons Rare Elements; High Skewness 308 313 313 316 316 318 319 320 322 322 325 329 331 334 334 338, 339 343 343 344 344 348. 349 352 352 354 355 356 359 359 363 370 370 372 374 374 377 384 384 388 388 389 390 304 396 396 398 401Contents xv 114A Selection Techniques for Rare Traits... 404 1L4B Highly Skewed Populations =. we 410 11.4C Screening with Field Ratings. oe. 412 11.5 Supplements for Incomplete Frames... 0. IS 11.6 Observational Units of Variable Sizes. ee 48 11.7 Weighting Problems . : fo at IL,7A Aims and Procedures of Weighting : 424 11.7B Losses from Random Duplication and Elimination | 427 N.7C Losses from Oversampling Siata =. a8 11.8 Estimating Totals - 433 118A The Advantages of Ngover Fy... ss 33 11.8B Totals from Subclasses of SRS. Foe a I1.8C The Product Estimate Ay of Aggregates 436 12, Some Selection Techniques. so. 0 12.4. Two-Phase Sampling in General and in Stratification... 440 12.1 The Economics of Two-Phase Sampling . 440 12.1B Stratification in Two-Phase Sampling. =. =... 443 12.1C Complex Two-Phase Selections. =. 444 *12.2. Two-Phase Ratio and Regression Means coe 446 12.24 Two-Phase Ratio Means . foe 446 12.2B Two-Phase Regression Estimates 448 12.2C Comparisons of Regression, Ratio, and Stratified Means in Two-Phase Sampling. - 450 12,3 Regression Means; Comparisons to Ratio and Difference Means | 431 12.34 Regression Means ee A8T 1238 Comparisons with Ratio and Difference Means ||. 454 12.3C Stratified Regression Means. : 435 12.4 Correlations from Overlaps in Repeated Surveys . 487 12.4A. Variances of Differences between Two Overlapping Samples | 457 12.4B Gains from Overlaps in Measuring Change + 462 12.4C Sums and Current Estimates from Overlapping Samples 466 12.5 Panel Studies and Designs for Measuring Changes 469 12.5A Panels versus New Samples. : 469 12.5B Partial Overlaps in Rotation Designs. =. =... TI 12.5C Choice of Overlapping Units : 4n 12.5D Splitting a Large Survey into Repeated Samples; Time Sampling ~ . 474 12.6 Continuing Sampling Operations . 6 ATT 12.6A Master Frames. ee AB 12.6B Techniques for Cumulating Information | | |. 480. 12.6C A Stratum for “Surprises” fe 481 12.6D Inertias of Continuing Operations =. 482 12.7 Changing selection Probabilities. . 2. 483 12.74. A Procedure for Minimum Changes of Selections. + 483 12.7B Modifications of the Procedure. . 485 12.7C A Simple Procedure for Burgeoning Units. . 486 12.7D Shifting Units between Strata 487 +128 Multiple Stratification, Lattice Sampling, Controlled Selection. | 488, 12.9 Standard Errors for Medians and Quantiles . - 495 12.10 Trinomials and Matched Binomials : 497xe Contents "12.11 Standard Errors for Combinations of Ratio Means. sot I2.11A Linear Combinations of Ratio Means oe. SOL 12.11B Double Ratios; Comparisons; Indexes 2. 508 Part Ill RELATED CONCEPTS . _. 507 13. Biases and Nonsampling Errors. : oe 509 13.1 Relation of Bias to Variable Error : : 309 13.2 Sources of Survey Errors : si4 13.2A Biases and Variable Errors to : 54 13.2B Descriptions of Diverse Errors 519 13.2 Some Effects of Errors. 2 S24 13.3, Noneoverage; Incomplete Frames, Missing Units 527 13.4 Nonresponse oo 532 13.4 Sources of Nonresponse oo 532 13.4B Effects of Nonresponse - 535 134€ Some Nonresponse Results and Call-Back Data : 536 13.5 Control of Nonresponse _ 548 135A Methods and Aims of Control. : + 548 13.5B Call-Backs $50 13.5C Subsampling Nonresponses 556 13.6 Four Proposed Remedies for Nonresponse 587 13.6A Estimation of Effects : 587 13.6B Substitutions for Nonresponses + 558 13,6C The Polity Scheme soe 559 13.6D A Replacement Procedure. + 360 13.7 Quota Sampling . 562 *13.8 Effects of Bias on Probability Statements - ~ 566 14, Some Issues of Inference from Survey Data. : . sm 14.1 Computation and Presentation of Sampling Errors . 314 14.2 Analytical Statistics for Complex Samples . + 582 14.3 Some Remarks about Statistical Inference oo 587 14.4 Experiments and Surveys : . 593 14.5. Multiple Objectives; Multipurpose Surveys. : 597 APPENDIXES . . Lo . A Summary of Symbols and Terms B Unit Analysis of Statistical Formulas © Remarks on Computations. D Table of Random Numbers . E List of a Population of 270 Blocks References . Answers to Selected Problems Index soepart 1 FUNDAMENTALS OF SURVEY SAMPLING1 Introduction 1.0 GUIDE TO THE INTRODUCTION This chapter differs from the others—as introductions to technical books often do. The rest of this book is devoted to specific methods of sampling, and the detailed treatments should serve the reader as clear guides to new methods. This chapter, however, gives only a brief coverage ‘of important matters that lie outside the narrow province of sampling, But since they are adjacent to it, mutual understanding of these ideas, based on a common vocabulary, is essential. Readers with various back- grounds may find some sections obvious and others too condensed. They can skim over the former and pursue the references to further reading in the latter. Generally, it would be best to read this chapter once before e after the rest of the book—and parts of it in between. troduction provides the why, what, and whence for the book: some rationale and motivation for the uses of “good” sampling methods; a bird’s-eye view of the basic problems and methods for meeting them; and an indication of how the approach of population sampling fits into survey methods and into the general quest for scientific knowledge. [attempt here a systematic classification of the units, tools, methods, problems, and goals of population sampling. Where good definitions are available | use them; for new definitions I try to conform to good usage and to reason, Where | had to coin new terms, I warn the reader to distingiish these novelties from the tried and trusted standard term: Because this chapter deals with fundamentals rather than techniques, 34 Introduction Ch, 1 and because it deals with them briefly yet wholeheartedly, it is also unavoidably more controversial than the rest of the book. 1.1 SURVEY DESIGN AND SAMPLE DESIGN Sample design has two aspects: a selection process, the rules and operations by which some members of the population are included in the sample; and an estimation process (or estimator) for computing the sample statistics, which are sample estimates of population values. The overall design of surveys includes other important aspects that can be called jointly the survey objectives (@ The definition of the survey variables shoutd specify the nature of the characteristics, the rules of classification categories, and the units for expressing them. It must also specify the extent and content of the survey population. (6) The methods of observation (measurement), including both data collecting and data processing, give operational meaning to the survey variables and determine the nature of the survey data. (©) The methods of analysis, statistical and substantive, reduce the survey data to results that can be comprehended and utilized. (@) The utilization of survey results may sometimes take the form of specified decisions, based on the survey results and other relevant information, More frequently, the results become part of the public fund of knowledge, and the researcher has only vague understanding of the future use of his results. (©) The desired precision of survey results may be clearly stated for samples designed for a specified statistical decision. More often the survey aims are many and vaguely stated, yet the researcher can find some broad limits of desired precision. Commonly, however, instead of specifying precision, the researcher must work from a reasonable allowed expenditure, and adjust accordingly the aims and the scope of the survey. This occurs in the design of surveys with many objectives, none of which is of predominant importance. The survey objectives should determine the sample design; but the determination is actually a two-way process, because the problems of sample design often influence and change the survey objectives. We shall encounter examples of the ways in which survey objectives and sample design interact to produce overall survey designs. A dialogue between the researcher and the sampler must occur before any aspect of the survey design is “frozen,” because a change in one aspect may dictate a change in others, Instead of a dialogue, the decisions may involve a larger cast:Sec. 1.1 Survey and Sample Design 5 sampler, researcher, and “consumer”; and the last, perhaps the grantor of the project, may feel behind him the silent pressure of the “ultimate consumers” of the data—the members of a profession or, perhaps, a wider public. The dialogue may occur silently within one head, if the researcher and sampler are one; but the dialogue should nevertheless take place. Most samples are prepared by statisticians and other researchers who are not primarily sampling specialists. Nevertheless, its helpful, although sometimes difficult, to separate sampling design ftom the related activities involved in survey research. The sample design covers the tasks of selection and estimation for making inference from sample value to the population value. Beyond this are the problems of making inferences from the survey population to another and generally broader population, with measurements free from error. Imagine a file of N cards, each of which contains the desired measure- ment (Y,) or measurements (X;, Y,, Z;,...) for one of the population elements. Computations on the entire file would produce the constant population value. Sampling would consist of selecting a fraction of the cards, then computing statistics from the sample to estimate the population value. For example, the sample mean 7 can be computed for estimating the population mean ¥. Moreover, the standard error of 7 can also be computed to estimate the average variability (the root-mean-square error) of j from Y. The actual deviation (j — Y)is unknown; only the average fluctuation, the standard error, can be estimated. Different sampling designs would result in different standard errors, and choosing the design with the smallest error is the principal aim of sampling design. The sampling fluctuation of (7 — Y) must be distinguished from the measurement bias that can affect the population value Y, based on a complete census of the file. The average accuracy of the measurements of the sample elements is equal to the average accuracy of the cards in the entire file. Comparing the sample value to the population value of the entire file, based on a complete equal coverage (Deming, 1960, p. 50]. separates sampling errors from the nonsampling errors of measurement It compares the sample value to an imaginary population value, obtained under similar essential survey conditions. In practice the file does not exist, and the measurements are made only after selection and only on the sample. But t ference is not crucial for thedistinction that places the measurement process outside thesampling design proper. Obtaining the element values (Y,, etc.) and placing them on the cards belong to the measurement process (collecting and processing). If several complete files were prepared under similar essential survey conditions, their population values would differ slightly from each other,© Introduction Cnt due to errors in the individual measurements. Even the number (NV) of cards in the file can be subject to errors of coverage. The population values would fluctuate slightly around a common expected value, determined by the essential survey conditions. To include the appropriate amount of this, variable nonsampling error in the computed total survey error is sometimes possible, although difficult (13.8). Finally, imagine several complete files, each prepared under different essential conditions, so that the accuracy of the cards differs from file to file. But the cost of obtaining access to the files differs also. These differences are not strictly within the sampler’s special domain, but he should not disregard them. The overall survey design should consider the problems of choosing the best file, together with the best sample design from the chosen file. The economic design of surveys requires the joint consideration and planning of sampling and nonsampling errors (Chapter 13). Since the sources and magnitudes of measurement errors in any science are subjects as deep as the entire science, to treat the errors of measurement in specifi detail would be an impossible task. Hence, knowledge about measurement errors must come from specialists in the substantive field(s) involved in the observations. The researcher, on the other hand, may not know enough sampling to design a valid and economic sample, and may consult a specialist in survey sampling. Survey sampling is a specialized skill within statistics because deep knowledge of the art and science requires full-time professional preparation and attention, Furthermore, knowledge of sampling design can be applied equally well to many fields of scientific, public, and business research. The sampler also concerns himself with other aspects of the survey objectives, and participates in their definitions and clarifications. In some cases he may even possess expert knowledge in some of those aspects, Conversely, the researcher may have ideas about the design of samples. Nevertheless, it is useful to make a separation between the two functions. Knowledge of measurement errors must come from specialists in the different fields involved in the observations. The sampler’s own statistical skills allow him to raise the right questions at the proper time, but not to answer them. He qualifies for Tukey's definition of an expert as somebody “who thinks with other people's brains.” 1.2 A TAXONOMY OF SURVEY UNITS AND CONCEPTS The elements of a population are the units for which information is sought; they are the individuals, the elementary units comprising thepec. 4.4 Umurs anu Conceprs 1 population about which inferences are to be drawn, They are the units of analysis, and their nature is determined by the survey objectives. The population is defined jointly with the elements: the population is the aggregate of the elements, and the elements are the basic units that comprise and define the population. The population must be defined in terms of (1) content, (2) units, (3) extent, and (4) time, For example, in the design of a survey of consumer expectations we may desire to specify: (1)all persons, (2) in family units, (3) in the United States and its territories, (4) in 1965, Often the desired population must be redefined to obtain a practicable survey population. For example, the above might be redefined as: (1) all persons above 18 years of age living in private dwelling units, (2) in spending units, (3) in the continental United States without Alaska, Hawaii, and the territories, (4) on January 1, 1965. The surrey population actually achieved may differ somewhat from the desired target population. The chief difference frequently arises from nonresponses and noncoverage. Strictly speaking, only the survey population is represented in the sample. But this may be difficult to describe exactly, and it is easier to write about the defined target population. One survey may yield information about several diverse populations. Different contents may be covered by the same survey. For example, a survey of home accidents may yield information about separate accidents, persons injured in accidents, families incurring accidents, and homes with accidents. Different units may be formed from the same data. Consumer data may be presented in terms of persons, spending units, families, or households. Different extents usually appear in the form of subclasses for which survey results are commonly prepared. The divisions may be geographic, as for regional data, or not, as with occupational or age subclasses. A subclass is a portion of the sample regarded as a sample from the corresponding portion of the survey population. Domains denote subclasses specifically planned for in the sample design. Different times may be represented in a survey when information is obtained about two or more periods, These, for example, could be current and past month, or past month and past year. We avoid using universe as a synonym for population. It denotes a hypothetical infinite set of elements generated by a theoretical model. This may be an ideal operation repeated endlessly, such as the endless tossing of a perfect coin, Behind every survey population stands some hypothetical universe, explicit or implicit, definite or indefinite. For example, a statistic about the U.S. adult population on a fixed date is expected to have relevance for other dates and, often hopefully, for other cultures. The sample provides statistical inference to the surveypopulation. Inference to hypothetical universe is more complex and not statistical (at least, not in the same sense). Characteristics of population elements are transformed to variables Y, by the survey operations of measurement, Some literature deals directly with the statistical population of the variables Y,. But I prefer to say that the ith element has the variable Y,. This permits us to talk of the many variables (¥), Xj, Zi, Win Pp ete.) of the same element. We can also consider relationships between variables of an element, changes of variables, and accuracy of measurements of variables, A statistic based on the variables found in a sample results in a random variable that we call a variate [Kendall and Buckland, 1957]. Sampling units contain the elements, and they are used for selecting elements into the sample. In element sampling, cach sampling unit contains only one element; but in cluster sampling, any sampling unit called cluster may contain several elements. For example, a sample of students may be obtained from a sample of classrooms, or a sample of dwellings from a sample of blocks. The same survey may use different kinds of sampling units, and in multistage sampling a hierarchy of sampling units or clusters is used, so that the element belongs uniquely to one sampling unit at each stage. For example, a sample of persons in a state may be taken by successively selecting counties, townships, segments, dwellings, and finally persons. The population is also an aggregate of the sampling units, specified for each stage. To improve the selection, the population is commonly divided into subpopulations called strata. The aggregate of the strata comprises the population. Listing units (briefly, listings) are used to identify and select sampling units from lists. Sometimes detailed procedures are needed to convert ings into sampling units; for example, converting listed addresses into dwellings and households. The problems may be serious if the elements fail to have unique identification with the listings. For example, a sample of families from telephone listings may involve serious diffi- culties (2.7). Observational units are the units from which the observations are obtained. In interview surveys they are called respondents. Observational units are often the elements, as in surveys of attitudes. But the two may be different. The household head may give data about all persons in the household, or about all children; in the former case the respondents belong, with others, to the population; in the latter they do not belong. In a survey of school children the respondent may be the teacher. Hansen, Hurwitz, and Madow (1953, 11, Chs. | and 2] define and develop well the relationships of several survey units.1.3 POPULATION VALUES AND STATISTICS Empirical research may be performed in different ways: by haphazard observations, controlled observations, experiments, or surveys. This book concentrates on sampling for surveys, Survey research is aimed at estimating specified population values. A population value is a numerical expression that summarizes the values of some characteristic(s) for all N elements of an entire population; it isa summary measure of some feature of the distribution of the variable(s) in the defined population, The basic example in survey sampling is the population mean Y = & ¥,/N, where Y, is the value for some variable of the ith element in the population. Population values closely allied to the mean are proportions, medians, and other quantiles, and the population total or aggregate. Still other popula- tion values measure relationships; the most common of these is the difference between two means; further examples are coefficients of regres- sion and correlation. Any population value is determined by four factors which appeared in the first three survey objectives: (a) the defined survey population; (b) the nature of the survey variable(s), and their distributions in some cases; (c) the methods of observation; and (d) the mathematical expression for deriving the population value from the individual element values. Both population value and true value refer to numerical expressions derived from the entire population. The difference between them arises from errors of observation. The true value would be obtained from all population elements, if the observations were not subject to error. The population value, also a function of all the observations, is subject to the same nonsampling errors as its sample estimate; itis a value that would be obtained if the entire population—rather than just a sample—were designated for observation under the actual survey conditions, Even that value is subject to fluctuation, since repeated measurements would obtain different values, But this variability of the population value is ordinarily small compared to the sampling variance, when the sample is only a small fraction of the population, Hence the population value is usually regarded as an unknown constant, neglecting its measurement error. We avoid the word “parameter” because its use in statistics is not clearly defined; in different contexts it may stand either for the true value or the population value; in mathematics it means something else again The sample value, or statistic, is an estimate computed from the n elements in the sample. An important example is the mean of the sample elements j == © y,|n, where y, is the observed value of the jth element in the sample. [t isa variate, or random variable, that depends on the sample design and on the particular combination of elements which happen to be10 Introduction Ch. 1 selected. Hence the particular estimate is only one among the many possible estimates which could have been obtained by the same sample design. On the contrary, the population value depends on all N values in the population. This isa constant independent of the vagaries of selection, although its value is usually unknown, Statistics are subject to both nonsampling errors and sampling errors. ‘The latter arise because only part of the total population is designated for observation in the sample. The nonsampling errors occur because the procedures of observation are imperfect. Their contribution to the total error of the survey should be considered jointly with the sample design, and they are treated in Chapter 13. Most of this book, and sampling design in general, deals chiefly with the variability of statistics around the population value. ‘The information that the mean J of a specific sample is, say, 14.7 units, has no intrinsic practical worth; its value lies in what it may tell us about, the population mean Y, We know that the unknown ¥ should differ from the known g, but we do not know by how much, The sample mean depends on which sample of # elements happened to be selected, since different samples, though of the same size n, and taken with the same design from the same population, would result in different sample means. The deviation of any single mean from the population mean is unknown; it may be plus or minus, large or small. The only statistical way of regarding sampling variability is in probability terms: what size deviations are likely to occur in the long run? We can ask: Given a sample design and size, what values of the sample mean 7 are possible, and what is the probability of occurrence of each of those values? This array of possible values of g, each with its probability of occurrence, is the sampling distribution of the possible sample means J for a fixed population, sample design, and size. “Sample mean” here denotes the sample estimate of the population mean; it is not necessarily the simple mean of the sample cases. We shall see examples where other estimators are preferred; for example, a weighted mean or a ratio mean. Furthermore, the discussions in this section are generally relevant not only to the mean but also to other statistics, Imagine that a sample design has been specified, including the sample size, and the selection and the estimation procedures; that we apply it toa fixed population, drawing sample after sample; and that we compute the mean for each sample and then tabulate and plot these values as the distribution of their relative frequencies. As the number of plotted values increases, the shape of the distribution becomes more stable and gradually approaches the true sampling distribution. This occurs as the “relative frequencies” for the different values of J approach the true probabilitiesSec. 1.3 Population Values and Statistics 11 of their occurrence: the theoretical result of an infinitely large number of drawings. Thus the sampling distribution of an estimate (mean) is the theoretical distribution of all possible values of the estimate (J,), each with its probability of occurrence (P.). The possible values and their probabilities depend on the sample design (size, selection, and estimation) applied to a fixed population of characteristics. The mean of the sampling distribution is the expected value of the estimate: thus, EG) = LPG. (13.0) This mean value E(y) may or may not be equal to the population value Y. The difference between the two we call the sampling bias = E(j) — Y. A sample design is called unbiased if E(j) = ¥. Note that this is not a property of a single sample, but of the entire sampling distribution, and that it belongs neither to the selection nor the estimation procedure alone, but to both jointly. Examples of unbiased estimates are the mean and the variance obiained from simple random samples. Many estimates are not, unbiased: the standard deviation of a simple random sample, the simple mean of a systematic sample of elements, and the ratio mean based on a random selection of clusters of unequal sizes; but in all these cases the sampling bias becomes negligibly small with increasing sample size. The standard deviation of the sampling distribution is called the standard error; it is the square root of the variance of the sampling distribution, which is the mean squared deviation around the mean E(J): Var (9) = > RU. — EC and SE (a) The sampling distribution represents the random fluctuation of g, due to the specific sample design, and this variability is measured by the standard error. This definition is fundamental and general. But it cannot be used directly to obtain the standard error because in practical situations we know only a single point in the whole sampling distribution. However, using this concept, statisticians have developed specific formulas for the variances of sample means for many practical sample designs. Different specific expressions appear for the different sample designs, For example, for the mean of an unrestricted sample (a simple random sample with replacement), the standard error is equal to the familiar o,/v/n. But this formula does not apply for other sample designs. However, knowing the population value of the standard error, denoted SE (9), is still of no practical use, because its value depends on the entire population, hence cannot be computed from the sample. For example, (1.3.2)n Introduction Ch. 1 x on 6 =>(%;— YPN cannot be computed from the sample. But statisticians have also developed practical formulas for computing variance estimates from the sample data. For example, se (J) = s,/V/ is the estimated standard error of an unrestricted sample where 5,2 = ¥(y; — D/(n — Formulas for computing appropriate sample estimates of the standard ~2SE(9) ~SEG) EG) +SE(5) ++2SE(5) be s1"Standard error of } = SEG) = VES - ET? ¥ You JL | sampling bias = £9) - ¥ Nonsampling bias = ¥ — Fie — Total error of 2 spectc sample 3” ~ Figg FIGURE 1.3.1 Schematic View of a Sampling Distribution We can think of each dot as a possible sample value, or as a group of equal numbers of values with similar values. The horizontal scale represents the possible values j. of the statistics, and the vertical scale represents the probability P. of occurrence of those values. The normal bell-shape is approached with moderate sized samples for most statistics, and for most variables ordinarily encountered, although these are typically not normal, The standard deviation of the sampling distribution is the standard error of the statistic, Arbitrary points were chosen to illustrate the population value Y, the true value Fyeuqs and a specific sample value g’. The standard error may more properly be called “variable error” if it includes variable errors from nonsampling sources. errors form important parts of the sections devoted to specific designs. These are denoted as var (7) and se (j) for sample estimates of the variance and the standard error, respectively, of the estimator j. Subscripts under 7 often distinguish the kind of sample designs for which specific formulas are appropriate. To be precise, an estimator should be distinguished from a particular estimate (or statistic, or sample value) for a specific sample. A sample design specifies the selection and estimation methods; for example, the mean of a simple random sample. A sample design with specified sample size, applied to a population of characteristics defined by survey operationsSec. 1.3 Population Values and Statistics 13 defines an estimator and gives rise to the sampling distribution of an estimator. The mean and the standard deviation of the distribution are the expected value and standard error of the estimator. Both of these are properties of the estimator (7), and not of any specific sample estimate (G,). Strictly speaking, then, se (J) is the “estimated standard error of the estimator 7.” But we shall often say briefly, “standard error of the esti- mate” or “computed standard error.” The mean square error of the sampling distribution is related to its variance, but with the deviations taken around the population value Y: MSE (7) = > PU. — ¥, ar (7) + [E@) — YF. (1.3.3) The added term is the square of the sampling bias, and it vanishes for unbiased sample designs. When not zero, the sampling bias is small in most well-designed samples, and it tends to diminish with increasing sample size, This property resembles the consistency property in mathematical statistics. In many practical situations the sampling distribution of the estimated mean is approximately normally distributed. Just how good that approxi- mation is depends on the underlying distribution of the characteristics in the population and on the sample design; and the approximation improves with increasing sample size. This approach to normality of the sampling distribution of large samples is not based on the normality of distribution of the elements in the population. On the contrary, the distributions of survey characteristics in the population are usually far from normal. They are more often of the kinds pictured in Fig. 1.3.11, and the distributions of elements in the samples selected with equal probability tend to resemble that in the population. Nevertheless, the sampling distributions of means of even moderate size samples, perhaps 100 or perhaps 1000 cases, will often be close enough to normal for the practical purpose of making statistical inference from sample estimates to population values. Furthermore, the approach to normality of sampling distributions for large samples occurs not only for the mean but for most estimators commonly used to present survey results. This problem is the concern of Central Limit Theorems. The term bias covers three concepts that have diverse origins and differ- ent practical consequences. Although, formally, each denotes a difference [E() — Yerue] between the expected value of a statistic and its “true” population value, they differ in important ways. First, the mathematical statistician thinks of biases due to using # for 2 — 1, or s to estimate o, or the median to estimate the median of a skewed distribution, or y/x to estimate ¥/¥. These may have academic interest, but they are of little14 Introduction Ch. 1 concern for survey samples and will be treated only in a few specific remarks. Second, large selection biases frequently result from the improper use of imperfect selection frames, These errors can mostly be avoided by any skilled sampler, yet they are amazingly common. This kind of common and avoidable selection bias will be our chief concern throughout the book, and especially in Sections 2.8 and 11.1 to 11.5. 956 Percentage of US 30 families with 5 or more children Distribution of percentage of US. families with various numbers of children 44 ° 0 1 o1234567 89 illia FIGURE 1.3.11 Two Illustrations of Population Distributions The cases in the sample will tend to have similar nonnormal distributions. But the sampling distribution of statistics based on moderate or large samples (for example, sample means based on a few hundred cases) will be approximately normal. Third, many surveys have large biases of observation that are difficult either to reduce or to assess, The sampler cannot remain indifferent to this central problem of research. However, the problem does not lie strictly within his competence and is treated separately in Chapter 13. 1.4 STATISTICAL INFERENCE IN SURVEYS The theory of sample surveys has concentrated on standard errors of estimates, especially of means and aggregates, for complex sample designs That is also a central concern of this book. This concentration is justified because statistical inference is based on standard errors. Statistical inference from surveys typically takes the form [7 + ¢, se (g)]. Briefly, this denotes the statement that the population mean Y is within the interval (7 — ¢, se (y)] to (7 +6, se(a)]; and the probability P of that statement is a function (usually and approximately normal or Student) of the chosen constant (,.Sec. 14 Statistical Inference 15 In Example 3.3a the mean is 7 = 0.20, and the standard error se (g) = 0.017. Then, choosing 1, = 2.6, we may say, with about P = 0.99 confidence, that the population mean is between the limits 0.20 + 2.6(0.017) = 0.20 + 0.044 = 0.156 to 0.244, If we are satisfied with P = 0.95 confidence and use ¢, = 2, then we can use the narrower limits 0.20 + 2(0.017) = 0.20 + 0.034 = 0.166 to 0.234. 1 P= 2P’—1; twosides 0.50 0,68 0.80 0.90 0.95 0.99 0.9973 0.999 2 Pi=1/2+P/2; oneside 0,75 0.84 0,90 0.95 0.975 0,995 0,9986 0.9995 3. tynormal (df = =) 0.67 1.00 1.28 1.64 1.96 2.58 3.00 3.29 4 t, Student (df = 30) 0.67 1,03 1.31 1,70 2.04 2.75 3.27 3.65 5. t, Student (df = 10) 0.70 1.05 137 1.81 2.23 3.17 3.96 4.59 TABLE 1.4.1 Selected Values of Normal and “Student” Deviates Probability levels of two-sided (1) and one-sided (2) intervals. Normal deviates (3) land “Student” deviates with 30 degrees (4) and with 10 degrees (5) of freedom, Frequently, we may prefer to construct one-sided intervals. Thus we may state that Y is greater than 7 — 1, se (9); or, alternatively, that Y is less than 7 + , se (J). For these one-sided intervals the probabilities are P’ =0.5 + 0.5P (or P = 2P’ — 1), where P denotes the probability for two-sided intervals corresponding to f,. In the above example, we may say with P’ = 0.5 + 0.5(0.95) = 0.975 confidence either that the popula- tion mean is greater than 0.166, or that it is less than 0.234, ‘The length of the interval depends on the standard error and, through 1,, on the probability level (P or P’). Choosing any two factors determines the third. The reader is assumed to have acquired the necessary statistical knowledge to use and to choose proper values of the standard error. Statisticians agree that it is good survey practice to publish standard errors together with survey results. This permits the reader to construct intervals and to make inferences according to his needs. Statisticians are well agreed also on the approximate amount of information conveyed by standard errors computed from survey samples, and on the practical procedures for using them to construct intervals. There is considerable disagreement about the meaning of those intervals, L avoid any adjective for the intervals 1, se (7), and the reader will utilize them according to his knowledge. If he wants confidence intervals, he should know that for most statistics based on complex samples the standard errors yield only approximate, not exact, confidence intervals,416 Introduction Ch. 1 Others will prefer to think of credible or likelihood or fiducial intervals (14.3), The interval [g + 1, se (J)] includes symbols for the sample mean, but, its place can be taken by many other statistics. We could write [£1,860 where d stands for some unspecified statistic and se (d) is its computed standard error. We should also note that se (6) is the estimate of the standard error of the entire sampling distribution, and not of the specific é obtained from the sample. Typically, se (3) is computed from the same data as 6; but this is not a mathematical requirement, and exceptions do arise in practice. For example, se (6) may be better known or estimated from other sources; or it may be estimated from only part of the entire sample. Using the normal deviates ¢, for constructing the interval [7 + 1, se (9)] involves the assumption that the normal distribution is a good approxima- tion for the sampling distribution of the statistic g. The normality of the sampling distribution depends on Central Limit Theorems for statistics based on large samples. (The approach to normality for distri- butions of statistics based on large samples does not require normality for the variables in the population distributions.) To the degree that the sampling distribution departs from the normal, the interval fails to possess the intended probabilities. I believe that for most sample estimates encountered in practical survey research, assumptions of normality lead to errors that are small compared to other sources of inaccuracy. For two reasons complex probability samples are typically large, from several hundred to thousands of cases, t, the organization and operations needed to obtain complex samples of large populations are so ponderous and expensive that only large samples can justify their economi- cal utilization, Second, most complex survey samples are aimed at measuring effects that are too small for detection with small samples. Complex probability samples are essentially tools for large samples. Small samples from small populations can be selected with simple random sampling, and then the vast available theory, parametric and nonpara- metric, can be resorted to, Smal] samples from widespread populations may be too difficult to select with either complex or simple random sam- pling. Then purposive selection can replace probability sampling, which is essentially a tool for large samples (1.7 and 14.4). Nevertheless, serious departures from normality can occur. When normality cannot be assumed, remedies may be sought in the statistical literature. Perhaps a transformation of the variables will bring theSec. 15 Selection Procedures 17 statistics closer to normal, Or a nonnormal distribution can be found and applied to the sampling distribution of the statistics. Ora distribution- free, or nonparametric, method can be applied. But most of these are not easily applicable, because they assume simple random selection (14.2) Criteria for judging the seriousness of departures from the normal would be difficult to define. Searching for clues of serious departures is also difficult, but several factors may be mentioned: 1. Departures from normality are proportionately greater at the higher probability levels, corresponding to the tails of the sampling distribution, An artificial example of a typical situation should clarify our meaning. At /, = 3.29 instead of the expected P = 0.999, nonnormality may give P ='0.995 for an increase factor of 0.005/0.001 = 5 in the error rate. At 1, = 1.96, a change from P = 0.95 to 0.945 would mean an increase factor of 0.055/0.05 = 1.1 in the error rate, 2. The approach to normality of the sampling distribution of the statistic faster for some variables than for others. Faster approach to normality means a better approximation of the sampling distribution for a given sample size, or a desired approximation reached with a smaller sample size. Even for small samples, the approximation is good for many com- mon variables, For example, 30 to 200 element selections suffice for proportions between 20 and 80 percent (Cochran, 1963, p. $7]. Contrari- wise, skewed distributions can result in serious departures from normality even for moderate size samples. For example, special methods are needed for analyzing income and wealth, and for most statistics of business and other establishments and organizations (11,4B). 3. Subclasses even from large samples can present problems. First, suppose that the subclass consists of a few primary clusters, each of which contains a fair number of elements, The analysis then depends on rough normality for the sample values of the individual clusters. Then the values of “Student's” 1, are used if the number of primary clusters is small, Degrees of freedom refer to numbers of primary clusters (less the number of strata) if the sample sizes within the clusters do not differ greatly (8.6D). Second, suppose that the subclass consists of a few elements spread individually over the entire breadth and width of the sample. Then, one can conjecture that the sample resembles a random choice, and refer to the vast literature for random selection. i.5 A TAXONOMY OF SELECTION PROCEDURES Any 100 percent census can be regarded as a sample for two reasons. To the degree that it is subject to errors of observation, the population18 Introduction Ch. value of a census is only one of many that could have resulted from essentially the same operations. Second, the particular population is arbitrarily specified from a universe of interest that is usually greater as to time, space, and perhaps other dimensions. For example, the decennial census singles one day out of the 10 x 365 days of a 10-year period. Nevertheless, we should separate these problems from the province of sampling proper, and consider the population from which a sample is drawn as a specified, entire population Survey sampling, ot population sampling, deals with methods for selecting and observing a part (sample) of the population in order to make inferences about the whole population, A sample can have several advantages over a complete census: (1) economy; (2) speed and time- liness; (3) feasibility (if the observation is “destructive,” a census is not practical); (4) quality and accuracy (in some situations money simply cannot buy the trained personnel and supervisors for a good census, or even a large sample) On the other hand, complete censuses possess special advantages in some situations. (1) Data for small units can be obtained. (2) Public acceptance is easier to secure for complete data, (3) Public compliance and response may be better secured, (4) Bias of coverage may be easier to check and reduce. (5) Sampling statisticians are not required [Yates, 1960, p. 2; Zarkovich, 1961, Chs. 1 and 2]. An important field of sampling that lies outside the scope of survey sampling is quality control. This deals with acceptance sampling by lot inspection; that is, sampling each lot to judge if it conforms to the quality specified for the entire universe. For this, frequent use is made of sequential sampling, where the sample size depends on the results of successive selections. But the collecting, processing, and analyzing procedures of surveys are generally lengthy and cumbersome, ill-suited to sequential sampling (8.4D). Survey sampling concentrates on the study of one-step probability samples for estimating population values. We frequently make inferences about populations from arbitrary and informal samples: we judge a basket of grapes by tasting one of them; a buyer accepts a shipment after inspecting a few items he picks hap- hazardly. Much research in the physical and biological sciences is based on items picked in a haphazard manner. The researchers assume, vaguely and implicitly, that “typical items” were chosen, They hope that the important characteristics are distributed either uniformly or randomly in the population. These are simple examples of model sampling, which we define generally as sampling based on broad assumptions about the distribution of the survey variables in the population. Several forms of model sampling can be distinguished.Sec. 1.5 Selection Procedures 19 1. Haphazard or fortuitous samples form the bases of most research in many fields. Samples of volunteer subjects should be included here. Archaeology, history, and often medicine draw conclusions from whatever items come to hand, Astronomy, experimental physics, and chemistry are leading sciences which display little care about the representativeness of their specimens, 2. Expert choice is a form of purposive or judgment sampling used by experts to pick “typical” or “representative” specimens, units, or portions. For example, consider the judgment sampling of “representative” indi- viduals or specimens for experiments; or the accounting practice of choosing typical weeks for auditing ledgers; or the common practice of picking a typical city or village to represent a national urban or rural population. Experts often hold differing views on the best way to choose representative specimens, or to decide which are the most representative units, Sometimes the researcher asks that, instead of a real population, a hypothetical universe be postulated as the parent of the sample. Some believe that a hypothetical universe suffices for theoretical models. But I think that inference from empirical data to a hypothetical universe does not lead to useful results if the gap between that universe and any real population is too great (14.4). 3. Quota sampling is a form of purposive sampling widely used in opinion, market, and similar surveys, The enumerators are instructed to obtain specified quotas from which to build a sample roughly propor- tional to the population, on a few demographic variables. Within the quotas, the enumerators are supposed to obtain representative individuals. The nature of the controls and instructions depends on the expert judgment of the practitioner (13.7). 4, Sampling of mobile populations often depends on “capture-tag- recapture” methods. The total population is estimated from the pro- portion in the recapture of individuals (insects, fish, deer) which have been previously captured and tagged. Ingenious theoretical models are used to state explicitly the assumptions of the method (Remark 11.1.1). These different types of model sampling vary greatly as to degree and area of justification. However, they have in common a heavy dependence on the validity of broad assumptions about the distributions of the survey variables in the population. On the contrary, from the results of ideal probability sampling, the inferences to the population can be made entirely by statistical methods, without assumptions regarding the population distributions, The need for assumptions of randomization of the popula- tion is by-passed by introducing randomization into the selection pro- cedures. Similarly, in card playing and in lotteries, instead of relying on20 Introduction Ch. 1 the dealer's judgment for a fair distribution, we insist on open and thorough shuffling. In probability sampling, every element in the population has a known nonzero probability of being selected, This probability is attained through some mechanical operation of randomization (1,7). ts value is determined in accord with the demands of the sample design. Probability samples are usually designed to be measurable; that is, so designed that statistical inference to population values can be based on measures of variability, usually standard errors, computed from the sample data. The desired and idealized properties of probability samples can only be approached, because many imperfections creep into the actual Epsem: equal probability for all elements (@) Equal probabilities at all stages (b) Equal overall probabilities for all elements obtained through compensating unequal proba- bilities at several stages Unequal probabilities for different elements; ordinarily compensated with inverse weights (a) Caused by irregularities in selection framesand procedures (b) Disproportionate allocation de- signed for optimum allocation Cluster Sampling: sampling units Il. Element Sampling: single stage, sampling unit contains only one | are clusters of elements element (a) One-stage cluster sampling (b) Subsampling oF multistage sampling (c) Equal clusters (a) Unequal clusters LL. Unstratified Selection: sampling | Stratified Sampling: separated se- units selected from entire popu- | lections from partitions, or strata, lation of population IV. Random Selection of individual | Systematic Selection of sampling sampling units from entire stratum | units with selection interval applied or population to list, V. One-Phase Sampling: final sample | Two-Phase (or Double) Sampling: TABLE 1.5.1 A Taxonomy of Probability Selection Methods. Five Alterna selected directly from entire popu- lation That May Be Combined. final sample selected from. first- phase sample, which obtains in- formation for stratification or estimationSec. 1.5 Selection Procedures 21 execution of practical samples. Hence, inference from sample values to population values also involves assumptions regarding the possible effects of imperfections; but often we can investigate, successively reduce, and put some limits on the possible or likely effects of imperfections. Therefore, I believe that separating probability sampling from the different types of model sampling is justified and useful. Simple random sampling (srs) is the basic selection process, and all other procedures can be viewed as modifications of it, introduced to provide more practical, economical, or precise designs, There are five major types of modifications and, since any of these can occur with any other, a large variety of possible designs appear immediately. This variety increases rapidly in multistage samples, where the choices must be made at each stage of selecting sampling units; for example, counties, blocks, dwellings. To these varieties must be added the possible diversity in estimating procedures. The great flexibility provided by different sampling methods is further i cased with the choice of different kinds of frames and practical procedures: 1. Epsem (equal probability of sclection method) sampling describes any sample in which the population elements have equal probabilities of selection. Note that this is a special type of _beobability sampling and, in tum, srsis.a special type oF epsem. Epsemt is used widely Because Tiswally leads io self-weighting samples; Where the simple mean of the sample cases is a good estimate of the population mean (2.8.3). Epsem sampling can result either from equal probability selection throughout, or from variable probabilities that compensate each other through the several stages of multistage selection. In contrast to epsem selection, some designs—we call them /oaded—call for variable probabilities of selection for different kinds of elements, 2. In element sampling (Chapters 2 to 4) the elements are also the only sampling units, Cluster sampling, on the contrary, involves the selection of groups, called clusters of elements, as sampling units, Subsampling of the Clusters results in multistage sampling: where the selection of the elements results from selection of sampling units in two or more stages (Chapters 5 to 10). 3. Stratification denotes selection from several subpopulations, called strata, into which the population is divided. 4, Systematic selection, an alternative to random choice, denotes the selection of sampling units in sequences separated on lists by the interval of selection. Briefly, we select every kth sampling unit. 5. Twro-phase sampling, or double sampling, refers to the subselection of the final sample from a preselected larger sample, that provides informa- tion for improving the final selection, Multiphase sampling refers to the possibility of more than two phases of selection1 2 3 4 s | 6 1 8 es y Ly? | vary) |y —fY|y —f¥ 013 4 io 35 | -8 o1s 6 2% | 105 | -6 016 7 37 15.5 -s 019 10 82 | 365 | -2 035 8 34 9.5 4 036 9 45 135 | 3 -3 039 12 90 31S oO 056 et 6L 13.3 -1 059 14 | 106 | 305 2 069 15 M7 315 2 135 9 35 6 —3 136 10 46 95 | ~2 139 13 OL 26 I 156 12 2 | 105 0 159 IS 107 24 3 3 169 16 | 1s | 245 4 356 4 70 35 2 359 17 1s 14 5 369 is | 126 | 13.5 6 569 20 142 6.5 8 8 Mean | 12 7 | 168 0 0 0 Var (y) = Ey -— f YP 16.8 3.5 9 64 TABLE 1.5.11 Numerical Illustration of a Sampling Distribution 1, The population of N = 6 elements have values ¥, = (0, 1, 3, 5, 6,9}. The population total is Y= = ¥,=24 and the population mean is Y/N = 24/6=4. From 3B ¥2 = 152, the element variance is S,? = (2 ¥¢— Y¥YN)MN = 1) = (152 = 2AYO)/S = 56/5 = 11.2. 2. A simple random sample of n = 3 elements results in sample total y = © y, is the estimator of fY Var (y) = (1 — njN)nS,* = (1 — 0.5)3(11.2) nIN = 3/6 = 0.5. The n¥. It has the variance 16.8. The sampling distribution con- iN’ 6" tains ( ) = () = 6 X 5 x 4)/(1 x 2 x 3) = 20 possible samples, in column 1. Each of the elements appears in f = 0.5 of the samples. ‘The sample estimates Of y in column 2 are seen to vary. Their average or expected alue is 12, illustrating that E@) = /Y, that y is an unbiased estimate of fY. Then ifs an unbiased estimate of Y, and y/n of ¥. Note in column 3 that the expected value of E y,? is 76, equal to fE ¥,%, hence also an unbiased estimate, ‘The computed variances in column 4 vary greatly, but their expected value is 16.8; thus E[var (y)} = Var (y) is another unbiased estimate. The deviations (y — fY)