Bayesian Biostatistics E-Book

Emmanuel Lesaffre

• Herausgeber: John Wiley & Sons
• Kategorie: Wissenschaft und neue Technologien
• Serie: Statistics in Practice
• Sprache: Englisch
• Veröffentlichungsjahr: 2012

Seitenzahl: 1005

Beschreibung

The growth of biostatistics has been phenomenal in recent years and has been marked by considerable technical innovation in both methodology and computational practicality. One area that has experienced significant growth is Bayesian methods. The growing use of Bayesian methodology has taken place partly due to an increasing number of practitioners valuing the Bayesian paradigm as matching that of scientific discovery. In addition, computational advances have allowed for more complex models to be fitted routinely to realistic data sets. Through examples, exercises and a combination of introductory and more advanced chapters, this book provides an invaluable understanding of the complex world of biomedical statistics illustrated via a diverse range of applications taken from epidemiology, exploratory clinical studies, health promotion studies, image analysis and clinical trials. Key Features: * Provides an authoritative account of Bayesian methodology, from its most basic elements to its practical implementation, with an emphasis on healthcare techniques. * Contains introductory explanations of Bayesian principles common to all areas of application. * Presents clear and concise examples in biostatistics applications such as clinical trials, longitudinal studies, bioassay, survival, image analysis and bioinformatics. * Illustrated throughout with examples using software including WinBUGS, OpenBUGS, SAS and various dedicated R programs. * Highlights the differences between the Bayesian and classical approaches. * Supported by an accompanying website hosting free software and case study guides. Bayesian Biostatistics introduces the reader smoothly into the Bayesian statistical methods with chapters that gradually increase in level of complexity. Master students in biostatistics, applied statisticians and all researchers with a good background in classical statistics who have interest in Bayesian methods will find this book useful.

Leseprobe

Contents

Cover

Series

Title Page

Copyright

Preface

Acknowledgments

Notation, terminology and some guidance for reading the book

Notation and terminology

Guidance for reading the book

Part 1: BASIC CONCEPTS IN BAYESIAN METHODS

Chapter 1: Modes of statistical inference

1.1 The frequentist approach: A critical reflection

1.2 Statistical inference based on the likelihood function

1.3 The Bayesian approach: Some basic ideas

1.4 Outlook

Chapter 2: Bayes theorem: Computing the posterior distribution

2.1 Introduction

2.2 Bayes theorem – the binary version

2.3 Probability in a Bayesian context

2.4 Bayes theorem – the categorical version

2.5 Bayes theorem – the continuous version

2.6 The binomial case

2.7 The Gaussian case

2.8 The Poisson case

2.9 The prior and posterior distribution of (θ)

2.10 Bayesian versus likelihood approach

2.11 Bayesian versus frequentist approach

2.12 The different modes of the Bayesian approach

2.13 An historical note on the Bayesian approach

2.14 Closing remarks

Chapter 3: Introduction to Bayesian inference

3.1 Introduction

3.2 Summarizing the posterior by probabilities

3.3 Posterior summary measures

3.4 Predictive distributions

3.5 Exchangeability

3.6 A normal approximation to the posterior

3.7 Numerical techniques to determine the posterior

3.8 Bayesian hypothesis testing

3.9 Closing remarks

Chapter 4: More than one parameter

4.1 Introduction

4.2 Joint versus marginal posterior inference

4.3 The normal distribution with μ and unknown

4.4 Multivariate distributions

4.5 Frequentist properties of Bayesian inference

4.6 Sampling from the posterior distribution: The Method of Composition

4.7 Bayesian linear regression models

4.8 Bayesian generalized linear models

4.9 More complex regression models

4.10 Closing remarks

Chapter 5: Choosing the prior distribution

5.1 Introduction

5.2 The sequential use of Bayes theorem

5.3 Conjugate prior distributions

5.4 Noninformative prior distributions

5.5 Informative prior distributions

5.6 Prior distributions for regression models

5.7 Modeling priors

5.8 Other regression models

5.9 Closing remarks

Chapter 6: Markov chain Monte Carlo sampling

6.1 Introduction

6.2 The Gibbs sampler

6.3 The Metropolis(–Hastings) algorithm

6.4 Justification of the MCMC approaches*

6.5 Choice of the sampler

6.6 The Reversible Jump MCMC algorithm*

6.7 Closing remarks

Chapter 7: Assessing and improving convergence of the Markov chain

7.1 Introduction

7.2 Assessing convergence of a Markov chain

7.3 Accelerating convergence

7.4 Practical guidelines for assessing and accelerating convergence

7.5 Data augmentation

7.6 Closing remarks

Chapter 8: Software

8.1 WinBUGS and related software

8.2 Bayesian analysis using SAS

8.3 Additional Bayesian software and comparisons

8.4 Closing remarks

Part II: BAYESIAN TOOLS FOR STATISTICAL MODELING

Chapter 9: Hierarchical models

9.1 Introduction

9.2 The Poisson-gamma hierarchical model

9.3 Full versus empirical Bayesian approach

9.4 Gaussian hierarchical models

9.5 Mixed models

9.6 Propriety of the posterior

9.7 Assessing and accelerating convergence

9.8 Comparison of Bayesian and frequentist hierarchical models

9.9 Closing remarks

Chapter 10: Model building and assessment

10.1 Introduction

10.2 Measures for model selection

10.3 Model checking

10.4 Closing remarks

Chapter 11: Variable selection

11.1 Introduction

11.2 Classical variable selection

11.3 Bayesian variable selection: Concepts and questions

11.4 Introduction to Bayesian variable selection

11.5 Variable selection based on Zellner's -prior

11.6 Variable selection based on Reversible Jump Markov chain Monte Carlo

11.7 Spike and slab priors

11.8 Bayesian regularization

11.9 The many regressors case

11.10 Bayesian model selection

11.11 Bayesian model averaging

11.12 Closing remarks

Part III: BAYESIAN METHODS IN PRACTICAL APPLICATIONS

Chapter 12: Bioassay

12.1 Bioassay essentials

12.2 A generic in vitro example

12.3 Ames/Salmonella mutagenic assay

12.4 Mouse lymphoma assay (L5178Y TK+/−)

12.5 Closing remarks

Chapter 13: Measurement error

13.1 Continuous measurement error

13.2 Discrete measurement error

13.3 Closing remarks

Chapter 14: Survival analysis

14.1 Basic terminology

14.2 The Bayesian model formulation

14.3 Examples

14.4 Closing remarks

Chapter 15: Longitudinal analysis

15.1 Fixed time periods

15.2 Random event times

15.3 Dealing with missing data

15.4 Joint modeling of longitudinal and survival responses

15.5 Closing remarks

Chapter 16: Spatial applications: Disease mapping and image analysis

16.1 Introduction

16.2 Disease mapping

16.3 Image analysis

Chapter 17: Final chapter

17.1 What this book covered

17.2 Additional Bayesian developments

17.3 Alternative reading

Appendix: Distributions

A.1 Introduction

A.2 Continuous univariate distributions

A.3 Discrete univariate distributions

A.4 Multivariate distributions

References

Index

Index

End User License Agreement

Statistics in Practice

Series Advisors

Human and Biological SciencesStephen Senn CRP-Santé, Luxembourg

Earth and Environmental SciencesMarian Scott University of Glasgow, UK

Industry, Commerce and FinanceWolfgang Jank University of Maryland, USA

Statistics in Practice is an important international series of texts which provide detailed coverage of statistical concepts, methods and worked case studies in specific fields of investigation and study.

With sound motivation and many worked practical examples, the books show in down-to-earth terms how to select and use an appropriate range of statistical techniques in a particular practical field within each title's special topic area.

The books provide statistical support for professionals and research workers across a range of employment fields and research environments. Subject areas covered include medicine and pharmaceutics; industry, finance and commerce; public services; the earth and environmental sciences, and so on.

The books also provide support to students studying statistical courses applied to the above areas. The demand for graduates to be equipped for the work environment has led to such courses becoming increasingly prevalent at universities and colleges.

It is our aim to present judiciously chosen and well-written workbooks to meet everyday practical needs. Feedback of views from readers will be most valuable to monitor the success of this aim.

A complete list of titles in this series can be found at www.wiley.com/go/statisticsinpractice

This edition first published 2012 © 2012 John Wiley & Sons, Ltd

Registered officeJohn Wiley & Sons Ltd, The Atrium, Southern Gate, Chichester, West Sussex, PO19 8SQ, United Kingdom

For details of our global editorial offices, for customer services and for information about how to apply for permission to reuse the copyright material in this book please see our website at www.wiley.com.

The right of the author to be identified as the author of this work has been asserted in accordance with the Copyright, Designs and Patents Act 1988.

All rights reserved. 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 or otherwise, except as permitted by the UK Copyright, Designs and Patents Act 1988, without the prior permission of the publisher.

Wiley also publishes its books in a variety of electronic formats. Some content that appears in print may not be available in electronic books.

Designations used by companies to distinguish their products are often claimed as trademarks. All brand names and product names used in this book are trade names, service marks, trademarks or registered trademarks of their respective owners. The publisher is not associated with any product or vendor mentioned in this book. This publication is designed to provide accurate and authoritative information in regard to the subject matter covered. It is sold on the understanding that the publisher is not engaged in rendering professional services. If professional advice or other expert assistance is required, the services of a competent professional should be sought.

Library of Congress Cataloging-in-Publication Data

Lesaffre, Emmanuel. Bayesian biostatistics / Emmanuel Lesaffre, Andrew Lawson. p. ; cm. Includes bibliographical references and index. ISBN 978-0-470-01823-1 (cloth) 1. Biometry–Methodology. 2. Bayesian statistical decision theory. I. Lawson, Andrew (Andrew B.) II. Title. [DNLM: 1. Biostatistics–methods. 2. Bayes Theorem. QH 323.5] QH323.5.L45 2012 570.1′5195–dc23 2012004237

A catalogue record for this book is available from the British Library.

ISBN: 978-0-470-01823-1

Preface

The growth of biostatistics as a subject has been phenomenal in recent years and has been marked by a considerable technical innovation in methodology and computational practicality. One area that has a significant growth is the class of Bayesian methods. This growth has taken place partly because a growing number of practitioners value the Bayesian paradigm as matching that of scientific discovery. But the computational advances in the last decade that have allowed for more complex models to be fitted routinely to realistic data sets have also led to this growth.

In this book, we explore the Bayesian approach via a great variety of medical application areas. In effect, from the elementary concepts to the more advanced modeling exercises, the Bayesian tools will be exemplified using a diversity of applications taken from epidemiology, exploratory clinical studies, health promotion studies and clinical trials.

This book grew out of a course that the first author has taught for many years (especially) in the Master programs in (bio)statistics at the universities of Hasselt and Leuven, both in Belgium. The course material was the inspiration for two out of three parts in the book. Therefore, the intended readership of this book are Master program students in (bio)statistics, but we hope that applied researchers with a good statistical background will also find the book useful. The structure of the book allows it to be used as course material for a course in Bayesian methods at an undergraduate or early stage postgraduate level. The aim of the book is to introduce the reader smoothly into Bayesian statistical methods with chapters that gradually increase in the level of complexity. The book consists of three parts. The first two parts of this work were the chapters primarily covered by the first author, while the last five chapters were primarily covered by the second author.

In Part I, we first review the fundamental concepts of the significance test and the associated P-value and note that frequentist methods, although proved to be quite useful over many years, are not without conceptual flaws. We also note that there are other methods on the market, such as the likelihood approach, but more importantly, the Bayesian approach. In addition, we introduce (an embryonic version of) the Bayes theorem. In Chapter 2, we derive the general expression of Bayes theorem and illustrate extensively the analytical computations to arrive at the posterior distribution on the binomial, the Gaussian and the Poisson case. For this, simple textbook examples are used. In Chapter 3, the reader is introduced to various posterior summary measures and the predictive distributions. Since sampling is fundamental to contemporary Bayesian approaches, sampling algorithms are introduced and exemplified in this chapter. While these sampling procedures will not yet prove their usefulness in practice, we believe that the early introduction of relatively simple sampling techniques will prepare the reader better for the advanced algorithms seen in later chapters. In this chapter, approaches to Bayesian hypothesis testing are treated and we introduce the Bayes factor. In Chapter 4, we extend all the concepts and computations seen in the first three chapters for univariate problems to the multivariate case, introducing also Bayesian regression models. It is then seen that, in general, no analytical methods to derive the posterior distribution are available, neither are the classical numerical approaches to integration sufficient. A new approach is then needed. Before addressing the solution to the problem, we treat in Chapter 5 the choice of the prior distribution. The prior distribution is the keystone to the Bayesian methodology. Yet, the appropriate choice of the prior distribution has been the topic of extensive discussions between non-Bayesians and Bayesians, but also among Bayesians. In this chapter, we extensively treat the various ways of specifying prior knowledge. In Chapters 6 and 7, we treat the basics of the Markov chain Monte Carlo methodologies. In Chapter 6, the Gibbs and the Metropolis–Hastings samplers are introduced again illustrated using a variety of medical examples. Chapter 7 is devoted to assessing and accelerating the convergence of the Markov chains. In addition, we cover the extension of the EM-algorithm to the Bayesian context, i.e. the data augmentation approach is exemplified. It is then time to see how Bayesian analyses can be done in practice. For this reason, we review in Chapter 8 the Bayesian software. We focus on two software packages: the most popular WinBUGS and the recently released Bayesian SAS® procedures. In both cases, a simple regression analysis serves as a guiding example helping the readers in their first analyzes with these packages. We end this chapter with a review of other Bayesian software, such as the packages OpenBUGS and JAGS, and also various R packages written to perform specific analyses.

In Part II, we develop Bayesian tools for statistical modeling. We start in Chapter 9 with reviewing hierarchical models. To fix ideas, we focus first on two simple two-level hierarchical models. The first is the Poisson-gamma model applied to a spatial data set on lip cancer cases in former East Germany. This example serves to introduce the concepts of hierarchical modeling. Then, we turn to the Gaussian hierarchical model as an introduction to the more general mixed models. A variety of mixed models are explored and amply exemplified. Also in this chapter comparisons between frequentist and Bayesian solutions aim to help the reader to see the differences between the two approaches. Model building and assessment are the topics of Chapter 10. The aim of this chapter is to see how statistical modeling could be performed entirely within the Bayesian framework. To this end, we reintroduce the Bayes factor (and its variants) to select between two statistical models. The Bayes factor is an important tool in model selection but is also fraught with serious computational difficulties. We then move to the Deviance Information Criterion (DIC). For a better understanding of this popular model selection criterion, we introduce at length the classical model selection criteria, i.e. AIC and BIC and relate them to DIC. The part on model checking describes all classical actions one would take to construct and evaluate a model, such as checking the residuals for outliers and influential observations, finding the correct scale of the response and the covariates, choosing the correct link function, etc. In this chapter, we also elaborate on the posterior predictive check as a general tool for goodness-of-fit testing. The final chapter, Chapter 11, in Part II handles Bayesian variable selection. This is a rapidly evolving topic that received a great impetus from the developments in bioinformatics. A broad overview of possible variable and model selection approaches and the associated software is given. While in the previous two chapters, the WinBUGS software and to a lesser extent the Bayesian SAS procedures were dominant, in this chapter we focus on software packages in R.

In Part III, we address particular application areas for Bayesian modeling. We include the most important areas from a practical biostatistical standpoint. In Chapter 12, we examine bioassay, where we consider preclinical testing methods: both Ames and Mouse Lymphoma in vitro assays and the famous Beetles LD50 toxicity assay are considered. In Chapter 13, we consider the important and pervasive problem of measurement error and also the misclassification in biostatistical studies. We discuss Berkson and classical joint models, bias such as attenuation and the problem of discrete error in the form of misclassification. In Chapter 14, we examine survival analysis from a Bayesian perspective. In this chapter, we cover basic survival time models and risk-set-based approaches and extend models to consider contextual effects within hazards. In Chapter 15, longitudinal analysis is considered in greater depth. Correlated prior distributions for parameters and also temporally correlated errors are considered. Missingness mechanisms are discussed and a nonrandom missingness example is explored. In Chapter 16, two important spatial biostatistical application areas are then considered: disease mapping and image analysis. In disease mapping, basic Poisson convolution models that include spatially structured random effects are examined for risk estimation, while in image analysis a focus on Bayesian fMRI analysis with correlated prior distributions is presented.

In Chapter 17, we end the book with a brief review of the topics that we did not cover in this book and give some key references for further reading. Finally, in the appendix, we provide an overview of the characteristics of most popular distributions used in Bayesian analyses.

Throughout the book there are numerous examples. In Parts I and II, explicit reference is made to the programs associated with the examples. These programs can be found at the website www.wiley.com/go/bayesian_methods_biostatistics. The programs used in Part III can also be found at this website.

Acknowledgments

When writing the early chapters, the first author benefitted from discussions with master students at Leuven, Hasselt and Leiden who pointed out various typos and ambiguities in earlier versions of the book. In addition, thanks go to the colleagues and former/current PhD students at L-Biostat at KU Leuven and at the Department of Biostatistics, Erasmus MC, Rotterdam, for illustrative discussions, critical remarks and help with software. In this respect, special thanks go to Susan Bryan, Silvia Cecere, Luwis Diya, Alejandro Jara, Arnošt Komárek, Marek Molas, Mukendi Mbuyi, Timothy Mutsvari, Veronika Rockova, Robin Van Oirbeek and Sten Willemsen. Software support was received from Sabanés Bové on the R package glmBfp, Fang Chen on the Bayesian SAS programs, Robert Gramacy on the R program monomvn, David Hastie and Peter Green on an R program for RJMCMC, David Lunn on the Jump interface in WinBUGS, Elizabeth Slate and Karen Vines. Thanks also go to those who provided data for the book or who gave permission to use their data, i.e. Steven Boonen, Elly Den Hondt, Jolanda Luime, the Signal-Tandmobiel® team, Bako Topal and Vincent van Weel. The authors also wish to thank, for interesting discussions, their colleagues in Leuven, Rotterdam and Charleston especially Dipankar Bandyopadhyay, Paul Eilers, Steffen Fieuws, Mulugeta Gebregziabher, Dimitris Rizopoulos and Elizabeth Slate. Finally, insightful conversations with George Casella, James Hodges, Helmut Küchenhoff and Paul Schmitt were much appreciated.

Last, the first author especially wishes to thank his wife Lieve Sels for her patience not only during the preparation of the ‘book’ but also during his whole professional career. To his children, Annemie and Kristof, he apologizes for the many times that their father was present but ‘absent’.

Emmanuel Lesaffre (Rotterdam and Leuven) Andrew B. Lawson (Charleston)

December 2011

Notation, terminology and some guidance for reading the book

Notation and terminology

In this section, we review some notation used in the book. We limit ourselves to outline some general principles; for precise definitions, we refer to the text.

First, both the random variable and its realization are denoted in this book as y. The vector of covariates is most often denoted as . The distribution of a discrete y as well as the density of a continuous variable will be denoted as p(y), unless otherwise stated. In the text, we make clear which of the two meanings applies. A sample of observations is denoted as , but also, a d-dimensional vector will be denoted in bold, i.e. y. We make it clear from the context what is implied. Further, independent, identically distributed random variables are indicated as i.i.d. A distribution (density) depending on a parameter vector is denoted as . The joint distribution of y and z is denoted as and the conditional distribution of y, given z will be denoted as . Alternatively, we use the notation . The probability that an event happens will occasionally be denoted as P for reasons of clarity.

A particular distribution will be addressed in two ways. For example, indicates that the random variable y has a gamma distribution with parameters α and β, but to indicate that the distribution is evaluated in y we will use the notation . When parameters have been given almost the same notation, say , then the notation is used where * is a place holder for 1, 2, 3.

In the normal case, some Bayesian textbooks use the precision notation while others use the variance notation. More specifically, if y has a normal distribution with mean μ and variance , then instead of (classical notation) the alternative notation (or even ) with the precision is used by some. This alternative notation is inspired by the fact that in Bayesian statistics some key results are better expressed in terms of the precision rather than the variance. This is also the notation used by WinBUGS. In this book, we frequently refer to classical, frequentist statistics. The use of precision would then be more confusing, rather than illuminating. In addition, when it comes to summarizing the results of a statistical analysis, the standard deviation is a better tool than the precision. For these reasons, we primarily used in this book the variance notation. But, throughout the book (especially in the later chapters) we regularly make the transition from one notation to the other.

Finally, we use some generally accepted standard notations, such as always denotes a column vector, |A| denotes the determinant of a matrix A, tr(A) is the trace of a matrix and AT denotes the transpose of a matrix A. The sample mean of is denoted as and their standard deviation as s, sy or simply SD.

Guidance for reading the book

No particular guidance is needed to read this book. The flow in the book is natural: starting from elementary Bayesian concepts we gradually introduce the more complex topics. This book deals, basically, only with parametric Bayesian methods. This means that all our random variables are assumed to have a particular distribution with a finite number of parameters. In the Bayesian world, many more distributions are used than in classical statistics. So for the classical reader (whatever meaning this may have), many of the distributions that pop-up in the book will be new. A brief characterization of these distributions, together with a graphical display, can be found in the appendix of the book.

Finally, some of the sections indicated by ‘*’ are technical and may be skipped at first reading.

Part 1

BASIC CONCEPTS IN BAYESIAN METHODS