Sampling
5/5
()
About this ebook
"This book has never had a competitor. It is the only book that takes a broad approach to sampling . . . any good personal statistics library should include a copy of this book."
—Technometrics
"Well-written . . . an excellent book on an important subject. Highly recommended."
—Choice
"An ideal reference for scientific researchers and other professionals who use sampling."
—Zentralblatt Math
Features new developments in the field combined with all aspects of obtaining, interpreting, and using sample data
Sampling provides an up-to-date treatment of both classical and modern sampling design and estimation methods, along with sampling methods for rare, clustered, and hard-to-detect populations. This Third Edition retains the general organization of the two previous editions, but incorporates extensive new material—sections, exercises, and examples—throughout. Inside, readers will find all-new approaches to explain the various techniques in the book; new figures to assist in better visualizing and comprehending underlying concepts such as the different sampling strategies; computing notes for sample selection, calculation of estimates, and simulations; and more.
Organized into six sections, the book covers basic sampling, from simple random to unequal probability sampling; the use of auxiliary data with ratio and regression estimation; sufficient data, model, and design in practical sampling; useful designs such as stratified, cluster and systematic, multistage, double and network sampling; detectability methods for elusive populations; spatial sampling; and adaptive sampling designs.
Featuring a broad range of topics, Sampling, Third Edition serves as a valuable reference on useful sampling and estimation methods for researchers in various fields of study, including biostatistics, ecology, and the health sciences. The book is also ideal for courses on statistical sampling at the upper-undergraduate and graduate levels.
Related to Sampling
Related ebooks
ANOVA and ANCOVA: A GLM Approach Rating: 0 out of 5 stars0 ratingsStatistics and Causality: Methods for Applied Empirical Research Rating: 0 out of 5 stars0 ratingsSpatio-temporal Design: Advances in Efficient Data Acquisition Rating: 0 out of 5 stars0 ratingsPharmacometrics: The Science of Quantitative Pharmacology Rating: 0 out of 5 stars0 ratingsA Course in Statistics with R Rating: 0 out of 5 stars0 ratingsBayesian Models for Categorical Data Rating: 4 out of 5 stars4/5Case Studies in Bayesian Statistical Modelling and Analysis Rating: 0 out of 5 stars0 ratingsRobust Statistics: Theory and Methods (with R) Rating: 0 out of 5 stars0 ratingsIntroduction to Population Pharmacokinetic / Pharmacodynamic Analysis with Nonlinear Mixed Effects Models Rating: 0 out of 5 stars0 ratingsNonparametric Regression Methods for Longitudinal Data Analysis: Mixed-Effects Modeling Approaches Rating: 0 out of 5 stars0 ratingsStatistical hypothesis testing Second Edition Rating: 0 out of 5 stars0 ratingsMethods of Multivariate Analysis Rating: 2 out of 5 stars2/5Statistics for Research Rating: 0 out of 5 stars0 ratingsBayesian Biostatistics Rating: 0 out of 5 stars0 ratingsJMP for Basic Univariate and Multivariate Statistics: Methods for Researchers and Social Scientists, Second Edition Rating: 0 out of 5 stars0 ratingsModern Experimental Design Rating: 0 out of 5 stars0 ratingsSPSS for Applied Sciences: Basic Statistical Testing Rating: 3 out of 5 stars3/5Regression Models for Time Series Analysis Rating: 2 out of 5 stars2/5Mathematical Methods of Statistics Rating: 3 out of 5 stars3/5Beginner’s Guide to Correlation Analysis: Bite-Size Stats, #4 Rating: 0 out of 5 stars0 ratingsIntroduction To Non Parametric Methods Through R Software Rating: 0 out of 5 stars0 ratingsTime Series Analysis in the Social Sciences: The Fundamentals Rating: 0 out of 5 stars0 ratingsFundamental Statistical Inference: A Computational Approach Rating: 0 out of 5 stars0 ratingsDesign and Analysis of Experiments in the Health Sciences Rating: 0 out of 5 stars0 ratingsTutorials in Biostatistics, Statistical Methods in Clinical Studies Rating: 0 out of 5 stars0 ratingsSampling in Statistics Rating: 0 out of 5 stars0 ratingsAn Introduction to Envelopes: Dimension Reduction for Efficient Estimation in Multivariate Statistics Rating: 0 out of 5 stars0 ratingsBayesian Non- and Semi-parametric Methods and Applications Rating: 3 out of 5 stars3/5Handbook of Regression Analysis Rating: 0 out of 5 stars0 ratingsggplot2 Essentials Rating: 0 out of 5 stars0 ratings
Mathematics For You
The Little Book of Mathematical Principles, Theories & Things Rating: 3 out of 5 stars3/5Basic Math & Pre-Algebra For Dummies Rating: 4 out of 5 stars4/5Quantum Physics for Beginners Rating: 4 out of 5 stars4/5Standard Deviations: Flawed Assumptions, Tortured Data, and Other Ways to Lie with Statistics Rating: 4 out of 5 stars4/5Mental Math Secrets - How To Be a Human Calculator Rating: 5 out of 5 stars5/5What If?: Serious Scientific Answers to Absurd Hypothetical Questions Rating: 5 out of 5 stars5/5My Best Mathematical and Logic Puzzles Rating: 4 out of 5 stars4/5Calculus Made Easy Rating: 4 out of 5 stars4/5Algebra - The Very Basics Rating: 5 out of 5 stars5/5Algebra II For Dummies Rating: 3 out of 5 stars3/5The Golden Ratio: The Divine Beauty of Mathematics Rating: 5 out of 5 stars5/5Real Estate by the Numbers: A Complete Reference Guide to Deal Analysis Rating: 0 out of 5 stars0 ratingsThe Everything Guide to Algebra: A Step-by-Step Guide to the Basics of Algebra - in Plain English! Rating: 4 out of 5 stars4/5Limitless Mind: Learn, Lead, and Live Without Barriers Rating: 4 out of 5 stars4/5Calculus For Dummies Rating: 4 out of 5 stars4/5Algebra I Workbook For Dummies Rating: 3 out of 5 stars3/5Mental Math: Tricks To Become A Human Calculator Rating: 5 out of 5 stars5/5The Math of Life and Death: 7 Mathematical Principles That Shape Our Lives Rating: 4 out of 5 stars4/5The Everything Everyday Math Book: From Tipping to Taxes, All the Real-World, Everyday Math Skills You Need Rating: 5 out of 5 stars5/5ACT Math & Science Prep: Includes 500+ Practice Questions Rating: 3 out of 5 stars3/5The New York Times Book of Mathematics: More Than 100 Years of Writing by the Numbers Rating: 0 out of 5 stars0 ratingsPre-Calculus For Dummies Rating: 5 out of 5 stars5/5How to Calculate Quickly: Full Course in Speed Arithmetic Rating: 4 out of 5 stars4/5Relativity: The special and the general theory Rating: 5 out of 5 stars5/5Painless Algebra Rating: 0 out of 5 stars0 ratingsHow Math Explains the World: A Guide to the Power of Numbers, from Car Repair to Modern Physics Rating: 3 out of 5 stars3/5How to Solve It: A New Aspect of Mathematical Method Rating: 4 out of 5 stars4/5
Reviews for Sampling
1 rating0 reviews
Book preview
Sampling - Steven K. Thompson
For further information visit: the book web page http://www.openmodelica.org, the Modelica Association web page http://www.modelica.org, the authors research page http://www.ida.liu.se/labs/pelab/modelica, or home page http://www.ida.liu.se/~petfr/, or email the author at [email protected]. Certain material from the Modelica Tutorial and the Modelica Language Specification available at http://www.modelica.org has been reproduced in this book with permission from the Modelica Association under the Modelica License 2 Copyright © 1998–2011, Modelica Association, see the license conditions (including the disclaimer of warranty) at http://www.modelica.org/modelica-legal-documents/ModelicaLicense2.html. Licensed by Modelica Association under the Modelica License 2.
Modelica© is a registered trademark of the Modelica Association. MathModelica© is a registered trademark of MathCore Engineering AB. Dymola© is a registered trademark of Dassault Syst`emes. MATLAB© and Simulink© are registered trademarks of MathWorks Inc. Java is a trademark of Sun MicroSystems AB. Mathematica© is a registered trademark of Wolfram Research Inc.
Copyright © 2011 by the Institute of Electrical and Electronics Engineers, Inc.
Published by John Wiley & Sons, Inc., Hoboken, New Jersey. All rights reserved.
Published simultaneously in Canada.
No part of this publication may be reproduced, stored in a retrieval system, or transmitted in any form or by any means, electronic, mechanical, photocopying, recording, scanning, or otherwise, except as permitted under Section 107 or 108 of the 1976 United States Copyright Act, without either the prior written permission of the Publisher, or authorization through payment of the appropriate per-copy fee to the Copyright Clearance Center, Inc., 222 Rosewood Drive, Danvers, MA 01923, (978) 750-8400, fax (978) 750-4744. Requests to the Publisher for permission should be addressed to the Permissions Department, John Wiley & Sons, Inc., 111 River Street, Hoboken, NJ 07030, (201) 748-6011, fax (201) 748-6008, or online at http://www.wiley.com/go/permission.
Limit of Liability/Disclaimer of Warranty: While the publisher and author have used their best efforts in preparing this book, they make no representations or warranties with respect to the accuracy or completeness of the contents of this book and specifically disclaim any implied warranties of merchantability or fitness for a particular purpose. No warranty may be created or extended by sales representatives or written sales materials. The advice and strategies contained herein may not be suitable for your situation. You should consult with a professional where appropriate. Neither the publisher nor author shall be liable for any loss of profit or any other commercial damages, including but not limited to special, incidental, consequential, or other damages.
For general information on our other products and services or for technical support, please contact our Customer Care Department within the United States at (800) 762-2974, outside the United States at (317) 572-3993 or fax (317) 572-4002.
Wiley also publishes its books in a variety of electronic formats. Some content that appears in print may not be available in electronic formats. For more information about Wiley products, visit our web site at www.wiley.com.
Library of Congress Cataloging-in-Publication Data:
Thompson, Steven K., 1945‒
Sampling / Steven K. Thompson. – 3rd ed.
p.cm. – (Wiley series in probability and statistics ; 755)
Includes index.
ISBN 978-0-470-40231-3 (hardback)
1 Sampling (Statistics) I. Title.
QA276.6.T58 2012
519.5'2–dc23
2011028944
Preface
One change with this edition of Sampling is that I have included sections of computing notes for sample selection, calculation of estimates, and simulations. These computations are illustrated using the statistical programming language R. In doing this I have avoided the use of specialized packages for specific complex designs, choosing instead to show simple calculations and sampling procedures from scratch using a few basic functions. The purpose of these sections is as much for understanding of sampling ideas as for easy ways to select samples and calculate estimates. Other software than R can, of course, be used for the same purpose. The advantages of R include: it is a free and open source, is widely supported by the statistical and other research communities, is available to anyone, and is easily installed on a computer with any of the common operating systems, including Windows, Macintosh OS X, Linux, and other types of Unix. The syntax of R tends to read like generic code and conveys the thinking that goes along with calculations rather than serving as a magic box. R is interactive and has very nice graphics.
Once one learns how to select a sample with a given type of design and to produce various types of estimates using the sample data from the design, it is an easy step to wrap that procedure into a simulation of a sampling strategy. Much of the attention of the computing sections is devoted to the simulation of sampling strategies. The idea is to construct a population
in the computer as much as possible like the real one which needs to be sampled. With this artificial but more-or-less realistic population, the sampling strategy is then carried out many times. So on each of the runs a sample is selected using the design, and estimates are calculated from the sample data obtained. The distribution of these estimates over the many runs is the sampling distribution. It depends as much on the sampling design and estimation procedure chosen as upon the characteristics of the population. In this way one prospective sampling strategy can be evaluated in comparison to others before committing to one to use in the field. In addition to providing a practical way to evaluate and improve potential sampling strategies, simulations of this kind can give an understanding that is right at the heart of sampling.
Some new examples have been added to this edition. New figures have been added, in particular illustrating the ideas of sampling distributions and the results of various types of simulations. Numerous incremental improvements and the odd new section have been added.
I would like to thank especially the students in my classes and colleagues at other institutions who have helped with corrections of typographical errors and other improvements. I would like to thank Susanne Steitz-Filler and Stephen Quigley at John Wiley & Sons for encouragement in preparation of this edition. Research support for my work in the area of sampling has been provided by the Natural Sciences and Engineering Research Council, the National Center for Health Statistics, Centers for Disease Control and Prevention, the U.S. Census Bureau, the National Institutes of Health, and the National Science Foundation.
Steven K. Thompson
Simon Fraser University
British Columbia
Preface to the Second Edition
The Second Edition retains the general organization of the first, but incorporates new material interspersed throughout the text. For example, model-based ideas and alternatives are included from the earliest chapters, including those on simple random sampling and stratified sampling, rather than suddenly appearing along with ratio and regression estimation methods as has been traditional. Estimation methods deriving from a combination of design and model considerations receive added attention in this edition. Some useful ideas from the ever-developing theory of sampling are briefly described in the chapters on making the most of survey data.
Among the added sections is an expanded description of methods for adjusting for nonsampling errors. A wider discussion of link-tracing designs for sampling hidden human populations—or the Internet—has been added to the chapter on network sampling. New developments in the rapidly expanding field of adaptive sampling are briefly summarized.
Additional numerical examples, as well as exercises, have been added. A number of additional derivations of results have been tucked into the later parts of chapters.
A brief history of sampling has been added to the introduction.
I would like to express my thanks and appreciation to the many people who have so generously shared with me their views on sampling theory and methods in discussions, collaborations, and visits to field sites. They include my colleagues at The Pennsylvania State University and those in the wider research community of sampling and statistics, as well as researchers in other fields such as ecology, biology, environmental science, computer science, sociology, anthropology, ethnography, and the health sciences. I would like to thank my editor Steve Quigley and editorial program coordinator Heather Haselkorn at John Wiley & Sons for their encouragement and assistance with this project. Research support for my work has been provided by grants from the National Science Foundation (DMS-9626102) and the National Institutes of Health (R01 DA09872).
Steven K. Thompson
University Park, Pennsylvania
Preface to the First Edition
This book covers the basic and standard sampling design and estimation methods and, in addition, gives special attention to methods for populations that are inherently difficult to sample, elusive, rare, clustered, or hard to detect. It is intended as a reference for scientific researchers and others who use sampling and as a textbook for a graduate or upper-level undergraduate course in sampling.
The twenty-six chapters of the book are organized into six parts. Part I covers basic sampling from simple random sampling to unequal probability sampling. Part II treats the use of auxiliary data with ratio and regression estimation and looks at the ideas of sufficient data and of model and design in practical sampling. Part III covers major useful designs including stratified, cluster, systematic, multistage, double, and network sampling. Part IV examines detectability methods for elusive populations: Basic problems in detectability, visibility, and catchability are discussed and specific methods of line transects, variable circular plots, capture–recapture, and line-intercept sampling are covered. Part V concerns spatial sampling, with the prediction or kriging
methods of geostatistics, considerations of efficient spatial designs, and comparisons of different observational methods including plot shapes and detection aspects. Part VI introduces adaptive sampling designs, in which the sampling procedure depends on what is observed during the survey; for example, sampling effort may be increased in the vicinity of high observed abundance. The adaptive cluster sampling designs described can be remarkably effective for sampling rare, clustered populations, which by conventional methods are notoriously difficult to sample.
Researchers faced with such problems as estimating the abundance of an animal population or an elusive human population, predicting the amount of mineral or fossil-fuel resource at a new site, or estimating the prevalence of a rare disease must be aware that the most effective methods go beyond the material traditionally found in sampling books. At the same time, such researchers may not be aware of the potential usefulness of some of the relatively recent developments in sampling theory and methods—such as network sampling, adaptive sampling designs, and generalized ratio and regression estimation with unequal probability designs. For these reasons, the selection of topics covered in this book is wider than has been traditional for sampling texts.
Some important sampling methodologies have developed largely in particular fields—such as ecology, geology, or health sciences—seemingly in isolation from the mainstream of statistical sampling theory. In the chapters on such methods, I have endeavored to bring out the connections with and the advantages to be gained from basic sampling design, estimation, and prediction results. Thus, for instance, in the chapters on detectability methods associated in particular with ecological sampling, sampling design is emphasized. In the chapter on the prediction or kriging methods associated with geostatistics, the connection to regression estimation results is noted. In the chapter on network sampling, originally associated with epidemiological surveys, the notation has been simplified and connections to basic unequal probability sampling estimators are observed.
Although the range of topics in this book is for the above-noted reasons considerably wider than has been traditional for sampling texts, it has been necessary, in order to keep the book of the desired size, to be selective in what to include. To the reader for whom an additional topic would have been particularly helpful, I can only offer the recompense of the references cited throughout the text to give access to the wider literature in sampling.
My immediate purposes in writing this book were to provide a text for graduate and upper-level undergraduate courses in sampling at the University of Alaska Fairbanks and at the University of Auckland and to provide a manual of useful sampling and estimation methods for researchers with whom I had worked on various projects in a variety of scientific fields. No available manual or text covered the range of topics of interest to these people.
In my experience the backgrounds of the researchers and students interested in sampling topics have been extremely diverse: While some are in statistics or mathematics, many others are in the natural and social sciences and other fields. In writing this book I have assumed the same diversity of backgrounds; the only common factor I feel I can take for granted is some previous course in statistics. The chapters are for the most part organized so that the basic methods and worked examples come first, with generalizations and key derivations following for those interested.
A basic one-semester course in sampling can consist of Chapters 1 through 8 and 11 through 13 or 14, with one or more topics from the remainder of the book added, depending on time and interest. For a graduate class in which many of the students are interested in the special topics of the last three parts of the book, the instructor may wish to cover the basic ideas and methods of the first three parts quite quickly, drawing on them for background later, and spend most of the time on the second half of the book.
I would like to give my thanks to the many people who have influenced and enriched the contents of this book through conversations, joint work, and other interactions on sampling and statistics. In particular, I would like to express appreciation to Fred Ramsey, P. X. Quang, Dana Thomas, and Lyle Calvin. Also, I am grateful to Lyman McDonald, David Siegmund, Richard Cormack, Stephen Buckland, Bryan Manly, Scott Overton, and Tore Schweder for enlightening conversations on statistical sampling methods. I would like to thank my colleagues at Auckland—George Seber, Alastair Scott, Chris Wild, Chris Triggs, Alan Lee, Peter Danaher, and Ross Ihaka—for the benefits of our collaborations, discussions, and daily interactions through which my awareness of relevant and interesting issues in sampling has been increased. I thank my sabbatical hosts at the Institute of Mathematical Statistics at the University of Copenhagen, where some of the sampling designs of this book were first seen as sketches on napkins in the lunch room: Søren Johansen, Tue Tjur, Hans Brøns, Martin Jacobsen, Inge Henningsen, Søren Tolver Jensen, and Steen Andersson. Among the many friends and associates around Alaska who have shared their experiences and ideas on sampling to the benefit of this book are Pat Holmes, Peter Jackson, Jerry McCrary, Jack Hodges, Hal Geiger, Dan Reed, Earl Becker, Dave Bernard, Sam Harbo, Linda Brannian, Allen Bingham, Alan Johnson, Terry Quinn, Bob Fagen, Don Marx, and Daniel Hawkins. Questions and comments leading to rethinking and rewriting of sampling topics have been contributed by many students, to each of whom I offer my thanks and among whom I would particularly like to mention Cheang Wai Kwong, Steve Fleischman, Ed Berg, and Heather McIntyre.
I would like to give a special thanks to my editor, Kate Roach, at John Wiley & Sons for her encouragement and enthusiasm. Research support provided by two grants from the National Science Foundation (DMS-8705812, supported by the Probability and Statistics Program and DMS-9016708, jointly supported by the Probability and Statistics Program and the Environmental Biology Division) resulted in a better book than would have otherwise been possible. I wish to thank Mary for, among many other things, her supportive sense of humor; when on a trip through Norway I could not find a certain guide book after ransacking the luggage jumble from one end of our vehicle to the other, she reminded me to use adaptive sampling
and, starting with the location of another book randomly discovered amidst the chaos, soon produced the wanted volume. Finally, I thank Jonathan, Lynn, Daniel, and Christopher for an environment of enthusiasm and innovativeness providing inspiration all along the way.
Steven K. Thompson
Auckland, New Zealand
Chapter 1
Introduction
Sampling consists of selecting some part of a population to observe so that one may estimate something about the whole population. Thus, to estimate the amount of lichen available as food for caribou in Alaska, a biologist collects lichen from selected small plots within the study area. Based on the dry weight of these specimens, the available biomass for the whole region is estimated. Similarly, to estimate the amount of recoverable oil in a region, a few (highly expensive) sample holes are drilled. The situation is similar in a national opinion survey, in which only a sample of the people in the population is contacted, and the opinions in the sample are used to estimate the proportions with the various opinions in the whole population. To estimate the prevalence of a rare disease, the sample might consist of a number of medical institutions, each of which has records of patients treated. To estimate the abundance of a rare and endangered bird species, the abundance of birds in the population is estimated based on the pattern of detections from a sample of sites in the study region. In a study of risk behaviors associated with the transmission of the human immunodeficiency virus (HIV), a sample of injecting drug users is obtained by following social links from one member of the population to another.
Some obvious questions for such studies are how best to obtain the sample and make the observations and, once the sample data are in hand, how best to use them to estimate the characteristic of the whole population. Obtaining the observations involves questions of sample size, how to select the sample, what observational methods to use, and what measurements to record. Getting good estimates with observations means picking out the relevant aspects of the data, deciding whether to use auxiliary information in estimation, and choosing the form of the estimator.
Sampling is usually distinguished from the closely related field of experimental design, in that in experiments one deliberately perturbs some part of a population in order to see what the effect of that action is. In sampling, more often one likes to find out what the population is like without perturbing or disturbing it. Thus, one hopes that the wording of a questionnaire will not influence the respondents' opinions or that observing animals in a population will not significantly affect the distribution or behavior of the population.
Sampling is also usually distinguished from observational studies, in which one has little or no control over how the observations on the population were obtained. In sampling one has the opportunity to deliberately select the sample, thus avoiding many of the factors that make data observed by happenstance, convenience, or other uncontrolled means unrepresentative.
More broadly, the field of sampling concerns every aspect of how data are selected, out of all the possibilities that might have been observed, whether the selection process has been under the control of investigators or has been determined by nature or happenstance, and how to use such data to make inferences about the larger population of interest. Surveys in which there is some control over the procedure by which the sample is selected turn out to have considerable advantages for purposes of inference about the population from which the sample comes.
1.1 Basic Ideas of Sampling and Estimation
In the basic sampling setup, the population consists of a known, finite number N of units—such as people or plots of ground. With each unit is associated a value of a variable of interest, sometimes referred to as the y-value of that unit. The y-value of each unit in the population is viewed as a fixed, if unknown quantity—not a random variable. The units in the population are identifiable and may be labeled with numbers 1, 2, … , N.
Only a sample of the units in the population are selected and observed. The data collected consist of the y-value for each unit in the sample, together with the unit's label. Thus, for each hole drilled in the oil reserve, the data not only record how much oil was found but also identify, through the label, the location of the hole. In addition to the variable of interest, any number of auxiliary variables, such as depth and substrate types, may be recorded. In a lichen survey, auxiliary variables recorded could include elevation, presence of other vegetation, or even eyeball
estimates of the lichen biomass. In an opinion poll, auxiliary variables such as gender, age, or income class may be recorded along with the opinions.
The procedure by which the sample of units is selected from the population is called the sampling design. With most well-known sampling designs, the design is determined by assigning to each possible sample s the probability P(s) of selecting that sample. For example, in a simple random sampling design with sample size n, a possible sample s consists of a set of n distinct units from the population, and the probability P(s) is the same for every possible sample s. In practice, the design may equivalently be described as a step-by-step procedure for selecting units rather than the resulting probabilities for selecting whole samples. In the case of simple random sampling, a step-by-step procedure consists of selecting a unit label at random from {1, 2, … , N}, selecting the next unit label at random from the remaining numbers between 1 and N, and so on until n distinct sample units are selected.
The entire sequence y1, y2, … , yN of y-values in the population is considered a fixed characteristic or parameter of the population in the basic sampling view. The usual inference problem in sampling is to estimate some summary characteristic of the population, such as the mean or the total of the y-values, after observing only the sample. Additionally, in most sampling and estimation situations, one would like to be able to assess the accuracy or confidence associated with estimates; this assessment is most often expressed with a confidence interval.
In the basic sampling view, if the sample size were expanded until all N units of the population were included in the sample, the population characteristic of interest would be known exactly. The uncertainty in estimates obtained by sampling thus stems from the fact that only part of the population is observed. While the population characteristic remains fixed, the estimate of it depends on which sample is selected. If for every possible sample the estimate is quite close to the true value of the population characteristic, there is little uncertainty associated with the sampling strategy; such a strategy is considered desirable. If, on the other hand, the value of the estimate varies greatly from one possible sample to another, uncertainty is associated with the method. A trick performed with many of the most useful sampling designs—cleverer than it may appear at first glance—is that this variability from sample to sample is estimated using only the single sample selected.
With careful attention to the sampling design and using a suitable estimation method, one can obtain estimates that are unbiased for population quantities, such as the population mean or total, without relying on any assumptions about the population itself. The estimate is unbiased in that its expected value over all possible samples that might be selected with the design equals the actual population value. Thus, through the design and estimation procedure, an unbiased estimate of lichen biomass is obtained whether lichens are evenly distributed throughout the study area or are clumped into a few patches. Additionally, the random or probability selection of samples removes recognized and unrecognized human sources of bias, such as conscious or unconscious tendencies to select units with larger (or smaller) than average values of the variable of interest. Such a procedure is especially desirable when survey results are relied on by persons with conflicting sets of interests—a fish population survey that will be used by fishery managers, commercial fishermen, and environmentalists, for instance. In such cases, it is unlikely that all parties concerned could agree on the purposive selection of a representative
sample.
A probability design such as simple random sampling thus can provide unbiased estimates of the population mean or total and also an unbiased estimate of variability, which is used to assess the reliability of the survey result. Unbiased estimates and estimates of variance can also be obtained from unequal probability designs, provided that the probability of inclusion in the sample is known for each unit and for pairs of units.
Along with the goal of unbiased or nearly unbiased estimates from the survey come goals of precise or low-variance estimates and procedures that are convenient or cost-effective to carry out. The desire to satisfy as many of these goals as possible under a variety of circumstances has led to the development of widely used sampling designs and estimation methods, including simple random and unequal probability sampling; the use of auxiliary information; stratified, systematic, cluster, multistage, and double sampling; and other techniques.
1.2 Sampling Units
With many populations of people and institutions, it is straightforward to identify the type of units to be sampled and to conceive of a list or frame of the units in the population, whatever the practical problems of obtaining the frame or observing the selected sample. The units may be people, households, hospitals, or businesses. A complete list of the people, households, medical institutions, or firms in the target population would provide an ideal frame from which the sample units could be selected. In practice, it is often difficult to obtain a list that corresponds exactly to the population of interest. A telephone directory does not list people without telephones or with unlisted numbers. The set of all possible telephone numbers, which may be sampled by random dialing, still does not include households without telephones. A list of public or private institutions may not be up-to-date.
With many other populations, it is not so clear what the units should be. In a survey of a natural resource or agricultural crop in a region, the region may be divided into a set of geographic units (plots or segments) and a sample of units may be selected using a map. However, one is free to choose alternative sizes and shapes of units, and such choices may affect the cost of the survey and the precision of estimators. Further, with a sampling procedure in which a point location is chosen at random in a study region and sample units are then centered around the selected points, the sample units can potentially overlap, and hence the number of units in the population from which the sample is selected is not finite.
For an elusive population with detectability problems, the role of units or plots may be superseded by that of detectability functions, which are associated with the methods by which the population is observed and the locations are selected for making the observations. For example, in selecting the locations of line transects in a bird survey and choosing the speed at which they are traversed, one determines the effective areas
observed within the study area in place of traditional sampling units or plots.
In some sampling situations the variable of interest may vary continuously over a region. For example, in a survey to assess the oil reserves in a region, the variable measured may be the depth or core volume of oil at a location. The value of such a variable is not necessarily associated with any of a finite set of units in the region, but rather, may be measured or estimated either at a point or as a total over a subregion of any size or shape.
Although the foregoing sampling situations go beyond the framework of a population divided uniquely into a finite collection of units from which the sample is selected, basic sampling design considerations regarding random sampling, stratified sampling, and other designs, and estimation results on design-unbiased estimation, ratio estimation, and other methods still apply.
1.3 Sampling and Nonsampling Errors
The basic sampling view assumes that the variable of interest is measured on every unit in the sample without error, so that errors in the estimates occur only because just part of the population is included in the sample. Such errors are referred to as sampling errors. But in real survey situations, nonsampling errors may arise also. Some people in a sample may be away from home when phoned or may refuse to answer a question on a questionnaire, and such nonrespondents may not be typical of the population as a whole, so that the sample tends to be unrepresentative of the population and the estimates are biased. In a fish survey, some selected sites may not be observed due to rough weather conditions; sites farthest from shore, which may not be typical of the study region as a whole, are the most likely to have such weather problems.
The problem of nonresponse is particularly pronounced in a survey with a very low response rate, in which the probability of responding is related to the characteristic to be measured—magazine readership surveys of sexual practices exemplify the problem. The effect of the nonresponse problem may be reduced through additional sampling effort to estimate the characteristics of the nonresponse stratum of the population, by judicious use of auxiliary information available on both responding and nonresponding units, or by modeling of the nonresponse situation. But perhaps the best advice is to strive to keep nonresponse rates as low as possible.
Errors in measuring or recording the variable of interest may also occur. Quality-control effort throughout every stage of a survey is needed to keep errors to a minimum. In some situations, it may be possible to model measurement errors separately from sampling issues in order to relate the observations to population characteristics.
Detectability problems are a type of nonsampling error that occurs with a wide range of elusive populations. On a bird survey, the observer is typically unable to detect every individual of the species in the vicinity of a sampling site. In a trawl survey of fish, not every fish in the path of the net is caught. Nor is every homeless person in a society counted in a census. A number of special techniques, including line transect, capture–recapture, and related methods, have been developed for estimating population quantities when detectability problems are a central issue.
1.4 Models in Sampling
In the basic sampling view the population is a finite set of units, each with a fixed value of the variable of interest, and probability enters only through the design, that is, the procedure by which the sample of units is selected. But for some populations it may be realistic and of practical advantage to consider a probability model for the population itself. The model might be based on knowledge of the natural phenomena influencing the distribution of the type of population or on a pragmatic statistical model summarizing some basic characteristics of such populations.
For example, a regression model may empirically describe a relationship between a variable of interest, the yield of a horticultural crop, say, with an auxiliary variable, such as the median level of an air pollutant. The model relating the variable of interest with the auxiliary variable has implications both for how to design the survey and how to make estimates.
In spatial sampling situations, the existence of correlations between values of the variable of interest at different sites, depending on the distance between the sites, has implications for choices regarding sampling design, estimation or prediction, and observational method. A model-based approach utilizing such correlation patterns has been particularly influential in geological surveys of mineral and fossil-fuel resources. In ecological surveys, such correlation patterns have implications not only for the spatial selection of observational sites, but for the observational methods (including plot shapes) used.
Ideally, one would like to be able to use a model of the population without having all conclusions of the survey depend on the model's being exactly true. A robust
approach to sampling uses models to suggest efficient procedures while using the design to protect against departures from the model.
1.5 Adaptive and Nonadaptive Designs
Surveys of rare, clustered populations motivate a further advance beyond the basic view of a sampling design. In adaptive sampling designs, the procedure for selecting sites or units on which to make observations may depend on observed values of the variable of interest. For example, in a survey for estimating the abundance of a natural resource, additional sites may be added to the sample during the survey in the vicinity of high observed abundance. Such designs have important applications to surveys of animal, plant, mineral, and fossil-fuel resources and may also have applications to other fields such as epidemiology and quality control.
The main purpose of adaptive procedures is to achieve gains in precision or efficiency, compared to conventional designs of equivalent sample size, by taking advantage of observed characteristics of the population. Adaptive procedures include such procedures as sequential stopping rules and sequential allocation among strata—procedures that have been rather heavily studied outside the finite-population context in the field of sequential analysis. With the population units identifiable as in the sampling situation, the possibilities for adaptive procedures are even greater, since it is possible to decide during a survey not just how many units to sample next but exactly which units or group of units to sample next.
In adaptive cluster sampling, whenever an observed value of the variable of interest satisfies a given criterion—for example, high abundance of animals observed at a site—units in the neighborhood of that unit (site) are added to the sample. A number of variations on this type of design are described in the final chapters of this book. For some populations, the designs produce remarkable increases in efficiency and appear to be particularly effective for sampling rare, clustered populations.
The sampling design is given for a conventional or nonadaptive design by a probability P(s) of selecting any particular sample s. For an adaptive design, the probability of selecting a given sample of units is P(s|y), that is, the probability of selecting sample s is conditional on the set y of values of the variable of interest in the population. Of course, in practice, the selection procedure can depend only on those values already observed.
Many natural populations tend to aggregate into fairly small portions of the study region, but the locations of these concentrations cannot be predicted prior to the survey. An effective adaptive design for such a population can result in higher selection probabilities assigned to samples that have a preponderance of units in those concentration areas. While the primary purpose of such a design may be to obtain a more precise estimate of the population total, a secondary benefit can be a dramatic increase in the yield of interesting observations—for example, more animals seen or more of a mineral obtained. Once adaptive designs are considered, the scope and potential of sampling methodology widens considerably.
1.6 Some Sampling History
In the earliest known European nonfiction book, The Histories (ca. 440 B.C.), the author Herodotus describes a sampling method used by a Persian king to estimate the number of his troops during an invasion of Greece. A sample group of a fixed number of soldiers was instructed to stand as close together as possible and the area in which they had stood was enclosed by a short fence. Then the entire army was marched through, filling the enclosure group by group, and the number of groups required was tabulated. Multiplying the number of groups by the number in the sample group gave the estimated size of the whole force. No attempt was made to assess the accuracy of the estimate, and no description is given of how the initial sample group was selected. In fact, historians believe that the estimate reported, 1,700,000, was a gross overestimate based on present knowledge regarding feasible sizes of populations and armies at that time. Even so, the sampling strategy appears to be a fairly sensible use of an expansion estimator, and the recorded overestimate may have more to do with military propagandizing or to Herodotus's enthusiasm for large numbers than to sampling variability or bias.
This place seemed to Xerxes a convenient spot for reviewing and numbering his soldiers; which things accordingly he proceeded to do … .What the exact number of the troops of each nation was I cannot say with certainty—for it is not mentioned by any one—but the whole land army together was found to amount to one million seven hundred thousand men. The manner in which the numbering took place was the following. A body of ten thousand men was brought to a certain place, and the men were made to stand as close together as possible; after which a circle was drawn around them, and the men were let go: then where the circle had been, a fence was built about the height of a man's middle; and the enclosure was filled continually with fresh troops, till the whole army had in this way been numbered. When the numbering was over, the troops were drawn up according to their several nations. (The History of Herodotus, Book VII, translated by George Rawlingson, The Internet Classics Archive by Daniel C. Stevenson, Web Atomics, 1994–2000, http:classics.mit.edu/Herodotus/history.html)
Many of the specific sampling designs and estimation methods in wide use today were developed in the twentieth century. Early in the twentieth century there was considerable debate among survey practitioners on the merits of random sampling versus purposively trying to select the most representative
sample possible. The basic methods and formulas of simple random sampling were worked out in the first two decades of the century. An article by Neyman (1934) compared the two methods and laid out the conceptual basis for probability sampling, in which the sample is selected at random from a known distribution. Most standard sampling designs—stratified sampling, systematic sampling, cluster sampling, multistage sampling, and double or multiphase sampling—had been introduced by the end of the 1930s. The U.S. Census introduced probability sampling methods when it took over the sample survey of unemployment in the early 1940s. Unequal probability designs were introduced in the 1940s and 1950s.
The theory and methods of sampling have continued to develop and expand throughout the second half of the twentieth and the early twenty-first centuries. Studies in the theory of sampling by Godambe and others from the early 1950s forward have helped clarify the inference issues in sampling and have opened the way for subsequent development of new methods. A number of new designs and inference methods have been introduced in response to difficult problems in studies of natural and human populations, with contributing developments coming from many fields. Differences of opinion over design-based versus model-based approaches in sampling have led to the development of methods that combine both approaches. Recent developments in the field of missing data analysis have opened up new analysis methods and underscored the importance of how observed