This Bayesian modeling book provides a self-contained entry to computational Bayesian statistics. Focusing on the most s

*350*
*96*
*8MB*

*English*
*Pages 296
[305]*
*Year 2013*

Springer Texts in Statistics

Jean-Michel Marin Christian Robert

Bayesian Essentials with R Second Edition

Springer Texts in Statistics Series Editors: George Casella Richard DeVeaux Stephen E. Fienberg Ingram Olkin

For further volumes: http://www.springer.com/series/417

Jean-Michel Marin • Christian P. Robert

Bayesian Essentials with R Second Edition

123

Jean-Michel Marin Universit´e Montpellier 2 Montpellier, France

Christian P. Robert Universit´e Paris-Dauphine Paris, France

ISSN 1431-875X ISSN 2197-4136 (electronic) ISBN 978-1-4614-8686-2 ISBN 978-1-4614-8687-9 (eBook) DOI 10.1007/978-1-4614-8687-9 Springer New York Heidelberg Dordrecht London Library of Congress Control Number: 2013950378 © Springer Science+Business Media New York 2014 This work is subject to copyright. All rights are reserved by the Publisher, whether the whole or part of the material is concerned, specifically the rights of translation, reprinting, reuse of illustrations, recitation, broadcasting, reproduction on microfilms or in any other physical way, and transmission or information storage and retrieval, electronic adaptation, computer software, or by similar or dissimilar methodology now known or hereafter developed. Exempted from this legal reservation are brief excerpts in connection with reviews or scholarly analysis or material supplied specifically for the purpose of being entered and executed on a computer system, for exclusive use by the purchaser of the work. Duplication of this publication or parts thereof is permitted only under the provisions of the Copyright Law of the Publisher’s location, in its current version, and permission for use must always be obtained from Springer. Permissions for use may be obtained through RightsLink at the Copyright Clearance Center. Violations are liable to prosecution under the respective Copyright Law. The use of general descriptive names, registered names, trademarks, service marks, etc. in this publication does not imply, even in the absence of a specific statement, that such names are exempt from the relevant protective laws and regulations and therefore free for general use. While the advice and information in this book are believed to be true and accurate at the date of publication, neither the authors nor the editors nor the publisher can accept any legal responsibility for any errors or omissions that may be made. The publisher makes no warranty, express or implied, with respect to the material contained herein. Printed on acid-free paper Springer is part of Springer Science+Business Media (www.springer.com)

To our most rewarding case studies, Chlo´e & Lucas, Joachim & Rachel

Preface

After that, it was down to attitude. —Ian Rankin, Black & Blue.— The purpose of this book is to provide a self-contained entry into practical and computational Bayesian statistics using generic examples from the most common models for a class duration of about seven blocks that roughly correspond to 13–15 weeks of teaching (with three hours of lectures per week), depending on the intended level and the prerequisites imposed on the students. (That estimate does not include practice—i.e., R programming labs, writing data reports—since those may have a variable duration, also depending on the students’ involvement and their programming abilities.) The emphasis on practice is a strong commitment of this book in that its primary audience consists of graduate students who need to use (Bayesian) statistics as a tool to analyze their experiments and/or datasets. The book should also appeal to scientists in all ﬁelds who want to engage into Bayesian statistics, given the versatility of the Bayesian tools. Bayesian essentials can also be used for a more classical statistics audience when aimed at teaching a quick entry to Bayesian statistics at the end of an undergraduate program, for instance. (Obviously, it can supplement another textbook on data analysis at the graduate level.) This book is an extensive revision of our previous book, Bayesian Core, which appeared in 2007, aiming at the same goals. (Glancing at this earlier version will show the ﬁliation to most readers.) However, after publishing Bayesian Core and teaching from it to diﬀerent audiences, we soon realized that the level of mathematics therein was actually more involved than the one expected by those audiences. Students were also asking for more advice and vii

viii

Preface

more R code than what was then available. We thus decided upon a major revision, producing a manual that cut the mathematics and expanded the R code, changing as well some chapters and replacing some datasets. We had at ﬁrst even larger ambitions in terms of contents, but had eventually to sacriﬁce new chapters for the sake of completing the book before we came to blows! To stress further the changes from the 2007 version, we also decided on a new title, Bayesian Essentials, that was actually suggested by Andrew Gelman during a visit to Paris. The current format of the book is one of a quick coverage of the topics, always backed by a motivated problem and a corresponding dataset (available in the associated R package, bayess), and a detailed resolution of the inference procedures pertaining to this problem, always including commented R programs or relevant parts of R programs. Special attention is paid to the derivation of prior distributions, and operational reference solutions are proposed for each model under study. Additional cases are proposed as exercises. The spirit is not unrelated to that of Nolan and Speed (2000), with more emphasis on the methodological backgrounds. While the datasets are inspired by real cases, we also cut on their description and the motivations for their analysis. The current format thus serves as a unique textbook for a service course for scientists aimed at analyzing data the Bayesian way or as an introductory course on Bayesian statistics. Note that we have not included any BUGS-oriented hierarchical analysis in this edition. This choice is deliberate: We have instead focussed on the Bayesian processing of mostly standard statistical models, notably in terms of prior speciﬁcation and of the stochastic algorithms that are required to handle Bayesian estimation and model choice questions. We plainly expect that the readers of our book will have no diﬃculty in assimilating the BUGS philosophy, relying, for instance, on the highly relevant books by Lunn et al. (2012) and Gelman et al. (2013). A course corresponding to the book has now been taught by both of us for several years in a second year master’s program for students aiming at a professional degree in data processing and statistics (at Universit´e Paris Dauphine, France) as well as in several US and Canadian universities. In Paris Dauphine the ﬁrst half of the book was used in a 6-week (intensive) program, and students were tested on both the exercises (meaning all exercises) and their (practical) mastery of the datasets, the stated expectation being that they should go beyond a mere reproduction of the R outputs presented in the book. While the students found that the amount of work required by this course was rather beyond their usual standards (!), we observed that their understanding and mastery of Bayesian techniques were much deeper and more ingrained than in the more formal courses their counterparts had in the years before. In short, they started to think about the purpose of a Bayesian statistical analysis rather than on the contents of the ﬁnal test and they ended up building a true intuition about what the results should look like, intuition

Preface

ix

that, for instance, helped them to detect modeling and programming errors! In most subjects, working on Bayesian statistics from this perspective created a genuine interest in the approach and several students continued to use this approach in later courses or, even better, on the job. Exercises are now focussed on solving problems rather than addressing ﬁner theoretical points. Solutions to about half of the exercises are freely available on our webpages. We insist upon the point that the developments contained in those exercises are often relevant for fully understanding in the chapter. Thanks We are immensely grateful to colleagues and friends for their help with this book and its previous version, Bayesian Core, in particular, to the following people: Fran¸cois Perron somehow started thinking about this book and did a thorough editing of it during a second visit to Dauphine, helping us to adapt it more closely to North American audiences. He also adopted Bayesian Core as a textbook in Montr´eal as soon as it appeared. George Casella made helpful suggestions on the format of the book. J´erˆome Dupuis provided capture–recapture slides that have been recycled in Chap. 5. Arnaud Doucet taught from the book at the University of British Columbia, Vancouver. Jean-Dominique Lebreton provided the European dipper dataset of Chap. 5. Gaelle Lefol pointed out the Eurostoxx series as a versatile dataset for Chap. 7. Kerrie Mengersen collaborated with both of us on a review paper about mixtures that is related to Chap. 6, Jim Kay introduced us to the Lake of Menteith dataset. Mike Titterington is thanked for collaborative friendship over the years and for a detailed set of comments on the book (quite in tune with his dedicated editorship of Biometrika). Jean-Louis Foulley provided us with some dataset and with extensive comments on their Bayesian processing. Even though we did not use those examples in the end, in connection with the strategy not to include BUGS-oriented materials, we are indebted to Jean-Louis for this help. Gilles Celeux carefully read the manuscript of the ﬁrst edition and made numerous suggestions on both content and style. Darren Wraith, Julyan Arbel, Marco Banterle, Robin Ryder, and Sophie Donnet all reviewed some chapters or some R code and provided highly relevant comments, which clearly contributed to the ﬁnal output. The picture of the caterpillar nest at the beginning of Chapter 3 was taken by Brigitte Plessis, Christian P. Robert’s spouse, near his great-grand-mother’s house in Brittany. We are also grateful to the numerous readers who sent us queries about potential typos, as there were indeed many typos and if not unclear statements. Thanks in particular to Jarrett Barber, Hossein Gholami, we thus encourage all new readers of Bayesian Essentials to do the same! The second edition of Bayesian Core was started, thanks to the support of the Centre International de Rencontres Math´ematiques (CIRM), sponsored

x

Preface

by both the Centre National de la Recherche Scientiﬁque (CNRS) and the Soci´et´e Math´ematique de France (SMF), located on the Luminy campus near Marseille. Being able to work “in pair” in the center for 2 weeks was an invaluable opportunity, boosted by the lovely surroundings of the Calanques, where mountain and sea meet! The help provided by the CIRM staﬀ during the stay is also most gratefully acknowledged. Montpellier, France Paris, France September 19, 2013

Jean-Michel Marin Christian P. Robert

Contents

1

User’s Manual . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.1 Expectations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.2 Prerequisites and Further Reading . . . . . . . . . . . . . . . . . . . . . . . . . 1.3 Styles and Fonts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.4 An Introduction to R . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.4.1 Getting Started . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.4.2 R Objects . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.4.3 Probability Distributions in R . . . . . . . . . . . . . . . . . . . . . . . 1.4.4 Graphical Facilities . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.4.5 Writing New R Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.4.6 Input and Output in R . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.4.7 Administration of R Objects . . . . . . . . . . . . . . . . . . . . . . . . 1.5 The bayess Package . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

1 2 3 4 5 6 8 15 16 19 21 21 22

2

Normal Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.1 Normal Modeling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.2 The Bayesian Toolkit . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.2.1 Posterior Distribution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.2.2 Bayesian Estimates . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.2.3 Conjugate Prior Distributions . . . . . . . . . . . . . . . . . . . . . . . 2.2.4 Noninformative Priors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.2.5 Bayesian Credible Intervals . . . . . . . . . . . . . . . . . . . . . . . . . 2.3 Bayesian Model Choice . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.3.1 The Model Index as a Parameter . . . . . . . . . . . . . . . . . . . . 2.3.2 The Bayes Factor . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.3.3 The Ban on Improper Priors . . . . . . . . . . . . . . . . . . . . . . . . 2.4 Monte Carlo Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.4.1 An Approximation Based on Simulations . . . . . . . . . . . . . 2.4.2 Importance Sampling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.4.3 Approximation of Bayes Factors . . . . . . . . . . . . . . . . . . . . .

25 26 28 28 33 34 35 37 38 39 41 43 46 47 49 52

xi

xii

Contents

2.5 Outlier Detection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58 2.6 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61 3

Regression and Variable Selection . . . . . . . . . . . . . . . . . . . . . . . . . 3.1 Linear Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.2 Classical Least Squares Estimator . . . . . . . . . . . . . . . . . . . . . . . . . 3.3 The Jeﬀreys Prior Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.4 Zellner’s G-Prior Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.4.1 A Semi-noninformative Solution . . . . . . . . . . . . . . . . . . . . . 3.4.2 The BayesReg R Function . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.4.3 Bayes Factors and Model Comparison . . . . . . . . . . . . . . . . 3.4.4 Prediction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.5 Markov Chain Monte Carlo Methods . . . . . . . . . . . . . . . . . . . . . . . 3.5.1 Conditionals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.5.2 Two-Stage Gibbs Sampler . . . . . . . . . . . . . . . . . . . . . . . . . . 3.5.3 The General Gibbs Sampler . . . . . . . . . . . . . . . . . . . . . . . . 3.6 Variable Selection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.6.1 Deciding on Explanatory Variables . . . . . . . . . . . . . . . . . . 3.6.2 G-Prior Distributions for Model Choice . . . . . . . . . . . . . . 3.6.3 A Stochastic Search for the Most Likely Model . . . . . . . . 3.7 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

65 66 69 73 74 75 80 81 84 85 86 87 90 91 91 93 96 98

4

Generalized Linear Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103 4.1 A Generalization of the Linear Model . . . . . . . . . . . . . . . . . . . . . . 104 4.1.1 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104 4.1.2 Link Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106 4.2 Metropolis–Hastings Algorithms . . . . . . . . . . . . . . . . . . . . . . . . . . . 108 4.2.1 Deﬁnition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109 4.2.2 The Independence Sampler . . . . . . . . . . . . . . . . . . . . . . . . . 110 4.2.3 The Random Walk Sampler . . . . . . . . . . . . . . . . . . . . . . . . 111 4.2.4 Output Analysis and Proposal Design . . . . . . . . . . . . . . . . 111 4.3 The Probit Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 115 4.3.1 Flat Prior . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 115 4.3.2 Noninformative G-Priors . . . . . . . . . . . . . . . . . . . . . . . . . . . 117 4.3.3 About Informative Prior Analyses . . . . . . . . . . . . . . . . . . . 122 4.4 The Logit Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 124 4.5 Log-Linear Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127 4.5.1 Contingency Tables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127 4.5.2 Inference Under a Flat Prior . . . . . . . . . . . . . . . . . . . . . . . . 131 4.5.3 Model Choice and Signiﬁcance of the Parameters . . . . . . 133 4.6 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 137

Contents

xiii

5

Capture–Recapture Experiments . . . . . . . . . . . . . . . . . . . . . . . . . . 139 5.1 Inference in a Finite Population . . . . . . . . . . . . . . . . . . . . . . . . . . . 140 5.2 Sampling Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 142 5.2.1 The Binomial Capture Model . . . . . . . . . . . . . . . . . . . . . . . 142 5.2.2 The Two-Stage Capture–Recapture Model . . . . . . . . . . . . 143 5.2.3 The T -Stage Capture–Recapture Model . . . . . . . . . . . . . . 148 5.3 Open Populations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 152 5.4 Accept–Reject Algorithms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 156 5.5 The Arnason–Schwarz Capture–Recapture Model . . . . . . . . . . . . 160 5.5.1 Modeling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 161 5.5.2 Gibbs Sampler . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 165 5.6 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 168

6

Mixture Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 173 6.1 Missing Variable Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 174 6.2 Finite Mixture Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 176 6.3 Mixture Likelihoods and Posteriors . . . . . . . . . . . . . . . . . . . . . . . . 177 6.4 MCMC Solutions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 182 6.5 Label Switching Diﬃculty . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 192 6.6 Prior Selection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 198 6.7 Tempering . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 199 6.8 Mixtures with an Unknown Number of Components . . . . . . . . . 201 6.9 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 206

7

Time Series . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 209 7.1 Time-Indexed Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 210 7.1.1 Setting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 210 7.1.2 Stability of Time Series . . . . . . . . . . . . . . . . . . . . . . . . . . . . 212 7.2 Autoregressive (AR) Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 214 7.2.1 The Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 215 7.2.2 Exploring the Parameter Space by MCMC Algorithms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 219 7.3 Moving Average (MA) Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . 226 7.4 ARMA Models and Other Extensions . . . . . . . . . . . . . . . . . . . . . . 232 7.5 Hidden Markov Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 236 7.5.1 Basics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 237 7.5.2 Forward–Backward Representation . . . . . . . . . . . . . . . . . . 241 7.6 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 248

8

Image Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 251 8.1 Image Analysis as a Statistical Problem . . . . . . . . . . . . . . . . . . . . 252 8.2 Spatial Dependence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 252 8.2.1 Grids and Lattices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 252 8.2.2 Markov Random Fields . . . . . . . . . . . . . . . . . . . . . . . . . . . . 254 8.2.3 The Ising Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 256 8.2.4 The Potts Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 260

xiv

Contents

8.3 Handling the Normalizing Constant . . . . . . . . . . . . . . . . . . . . . . . . 262 8.3.1 Path Sampling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 264 8.3.2 The ABC Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 267 8.3.3 Inference on Potts Models . . . . . . . . . . . . . . . . . . . . . . . . . . 270 8.4 Image Segmentation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 273 8.5 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 281 About the Authors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 285 References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 287 Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 291

1 User’s Manual

The bare essentials, in other words. —Ian Rankin, Tooth & Nail.—

Roadmap The Roadmap is a section that will start each chapter by providing a commented table of contents. It also usually contains indications on the purpose of the chapter. For instance, in this initial chapter, we explain the typographical notations that we adopted to distinguish between the diﬀerent semantic levels of the course. We also try to detail how one should work with this book and how one could best beneﬁt from this work. This chapter is to be understood as a user’s (or instructor’s) manual that details our pedagogical choices. It also seems the right place to introduce the programming language R, which we use to illustrate all the introduced concepts. In each chapter, both Ian Rankin’s quotation and the ﬁgure on top of the title page are (at best) vaguely related to the topic of the chapter, and one should not waste too much time pondering their implications and multiple meanings. The similarity with the introductory chapter of Introducing Monte Carlo Methods with R is not coincidental, as Robert and Casella (2009) used the same skeleton as in Bayesian Core and as we restarted from their version.

J.-M. Marin and C.P. Robert, Bayesian Essentials with R, Springer Texts in Statistics, DOI 10.1007/978-1-4614-8687-9 1, © Springer Science+Business Media New York 2014

1

2

1 User’s Manual

1.1 Expectations The key word associated with this book is modeling, that is, the ability to build up a probabilistic interpretation of an observed phenomenon and the “story” that goes with it. The “grand scheme” is to get anyone involved in analyzing data to process a dataset within this coherent methodology. This means picking a parameterized probability distribution, denoted by fθ , and extracting information about (shortened in “estimating”) the unknown parameter θ of this probability distribution in order to provide a convincing interpretation of the reasons that led to the phenomenon at the basis of the dataset (and/or to be able to draw predictions about upcoming phenomena of the same nature). Before starting the description of the probability distributions, we want to impose on the reader the essential feature that a model is an interpretation of a real phenomenon that ﬁts its characteristics up to some degree of approximation rather than an explanation that would require the model to be “true”. In short, there is no such thing as a “true model”, even though some models are more appropriate than others! In this book, we chose to describe the use of “classical” probability models for several reasons: First, it is often better to start a trip on well-traveled paths because they are less likely to give rise to unexpected surprises and misinterpretations. Second, they can serve as references for more advanced modelings: Quantities that appear in both simple and advanced modelings should get comparable estimators or, if not, the more advanced modeling should account for that diﬀerence. At last, the deliberate choice of an artiﬁcial model should give a clearer meaning to the motto that all models are false in that it illustrates the fact that a model is not necessarily justiﬁed by the theory beyond the modeled phenomenon but that its corresponding inference can nonetheless be exploited as if it were a true model. By the end of the book, the reader should also be in a position to assess the relevance of a particular model for a given dataset. Working with this book should not appear as a major endeavor: The datasets are described along with the methods that are relevant for the corresponding model, and the statistical analysis is provided with detailed comments. The R code that backs up this analysis is included and commented throughout the text. If there is a diﬃculty with this scheme, it actually starts at this point: Once the reader has seen the analysis, it should be possible for her or him to repeat this analysis or a similar analysis with no further assistance. Even better, the reader should try to read as little as possible of the analysis proposed in this book and on the opposite hand should try to conduct the following stage of the analysis before reading the proposed (but not unique) solution. The ultimate lesson here is that there are indeed many ways to analyze a dataset and to propose modeling scenarios and inferential schemes. It is beyond the purpose of this book to provide all of those analyses, and the reader (or the instructor) is supposed to look for alternatives on her or his own.

1.2

Prerequisites and Further Reading

3

We thus expect readers to place themselves in a realistic situation to conduct this analysis in life-threatening (or job-threatening) situations. As detailed in the preface, the course was originally intended for students in the last year of study toward a professional degree, and it seems quite reasonable to insist that they face similar situations before entering their incoming job!

1.2 Prerequisites and Further Reading This being a textbook about statistical modeling, the students are supposed to have a background in both probability and statistics, at the level, for instance, of Casella and Berger (2001). In particular, a knowledge of standard sampling distributions and their properties is desirable. Lab work in the spirit of Nolan and Speed (2000) is also a plus. (One should read in particular their Appendix A on “How to write lab reports?”) Further knowledge about Bayesian statistics is not a requirement, although using Robert (2007) or Hoﬀ (2009) as further references would bring a better insight into the topics treated here. Similarly, we expect students to be able to understand the bits of R programs provided in the analysis, mostly because the syntax of R is very simple. We include an introduction to this language in this chapter and we refer to Dalgaard (2002) for a deeper entry and also to Venables and Ripley (2002). Besides Robert (2007), the philosophy of which is obviously reﬂected in this book, other reference books pertaining to applied Bayesian statistics include Gelman et al. (2013), Carlin and Louis (1996), and Congdon (2001, 2003). More speciﬁc books that cover parts of the topics of a given chapter are mentioned (with moderation) in the corresponding chapter, but we can quote here the relevant books of Holmes et al. (2002), Pole et al. (1994), and Gill (2002). We want to stress that the citations are limited for eﬃciency purposes: There is no extensive coverage of the literature as in, e.g., Robert (2007) or Gelman et al. (2013), because the prime purpose of the book is to provide a working methodology, for which incremental improvements and historical perspectives are not directly relevant. While we also cover simulation-based techniques in a self-contained perspective, and thus do not assume prior knowledge of Monte Carlo methods, detailed references are Robert and Casella (2004, 2009) and Chen et al. (2000). Although we had at some stage intended to write a new chapter about hierarchical Bayes analysis, we ended up not including this chapter in the current edition and this for several reasons. First, we were not completely convinced about the relevance of a speciﬁc hierarchical chapter, given that the hierarchical theme is somehow transversal to the book and pops in the mixture (Chap. 6), dynamic (Chap. 7) and image (Chap. 8) chapters. Second, the revision took already too long and creating a brand new chapter did not sound a manageable goal. Third, managing realistic hierarchical models meant relying on codes written in JAGS and BUGS, which clashed with the philosophy of backing the whole book on R codes. This was subsumed by the recent and highly relevant publication of The BUGS Book (Lunn et al., 2012) and by the incoming new edition of Bayesian Data Analysis (Gelman et al., 2013).

4

1 User’s Manual

1.3 Styles and Fonts Presentation often matters almost as much as content towards a better understanding, and this is particularly true for data analyzes, since they aim to reproduce a realistic situation of a consultancy job where the consultant must report to a customer the results of an analysis. An equilibrated use of graphics, tables, itemized comments, and short paragraphs is, for instance, quite important for providing an analysis that stresses the diﬀerent conclusions of the work, as well as the points that are yet unclear and those that could be expanded. In particular, because this book is doing several things at once (that is, to introduce theoretical and computational concepts and to implement them in realistic situations), it needs to diﬀerentiate between the purposes and the levels of the parts of the text so that it is as obvious as possible to the reader. To this eﬀect, we take advantage of the many possibilities of modern computer editing, and in particular of LATEX, as follows. First, a minimal amount of theoretical bases is required for dealing with the model introduced in each chapter, either for Bayesian statistics or for Monte Carlo theory. This aspect of the material is necessarily part of the main text, but it is also kept to a minimum—just enough for the book to be self-contained—and therefore occasional references to more detailed books such as Robert (2007) and Robert and Casella (2004) are necessary. These sections need be well-understood before handling the following applications or realistic cases. This book is primarily intended for those without a strong background in the theory of Bayesian statistics or computational methods, and “theoretical” sections are essential for them, hence the need to keep those sections within the main text. Statistics is as much about data processing as about mathematical and probabilistic modeling. To enforce this principle, we center each chapter around one or two speciﬁc realistic datasets that are described early enough in the chapter to be used extensively throughout the chapter. These datasets are available on the book’s Website (http://www.ceremade.dauphine.fr/~xian/BCS/) and are part of the corresponding R package bayess, as normaldata, capturedata, and so on, the name being chosen in reference to the case/chapter heading. (Some of these datasets are already available as datasets in the R language.) In particular, we explain the “how and why” of the corresponding dataset in a separate paragraph in this shaded format. This style is also used for illustrating theoretical developments for the corresponding dataset and for speciﬁc computations related to this dataset. For typographical convenience, large graphs and tables may appear outside these sections, in subsequent pages, but are obviously mentioned and identiﬁed within them.

1.4

An Introduction to R

5

Example 1.1. There may also be a need for detailed examples in addition to the main datasets, although we strived to keep them to a minimum and only for very speciﬁc issues where the reference dataset was not appropriate. They follow this numbered style, the sideways triangle indicating the end of the example.

The

last style used in the book is the warning, represented by a lightning symbol in the margin: This entry is intended to signal major warnings about things that can (and do) go wrong “otherwise”; that is, if the warning is not taken into account. Needless to say, these paragraphs must be given the utmost attention!

A diverse collection of exercises is proposed at the end of each chapter, with solutions to all those exercises freely available on Springer-Verlag webpage.

1.4 An Introduction to R This section attempts at introducing R to newcomers in a few pages and, as such, it should not be considered as a proper introduction to R. Entire volumes, such as the monumental R Book by Crawley (2007), and the introduction by Dalgaard (2002), are dedicated to the practice of this language, and therefore additional eﬀorts (besides reading this chapter) will be required from the reader to suﬃciently master the language.1 However, before discouraging anyone, let us comfort you with the fact that: (a) The syntax of R is simple and logical enough to quickly allow for a basic understanding of simple R programs, as should become obvious in a few paragraphs. (b) The best, and in a sense the only, way to learn R is through trial-anderror on simple and then more complex examples. Reading the book with a computer available nearby is therefore the best way of implementing this recommendation. In particular, the embedded help commands help() and help.search() are very good starting points to gather information about a speciﬁc function or a general issue, even though more detailed manuals are available both locally and on-line. Note that help.start() opens a Web browser linked to the local manual pages. One may ﬁrst wonder why we support using R as the programming interface for this introduction to Monte Carlo methods, since there exist other 1 If you decide to skip this chapter, be sure to at least print the handy R Reference Card available at http://cran.r-project.org/doc/contrib/Short-refcard.pdf that summarizes, in four pages, the major commands of R.

6

1 User’s Manual

languages, most (all?) of them faster than R, like Matlab, and some even free, like C or Python. We obviously have no partisan or commercial involvement in this language.2 Rather, besides the ease of presentation, our main reason for this choice is that the language combines a suﬃciently high power (for an interpreted language) with a very clear syntax both for statistical computation and graphics. R is a ﬂexible language that is object-oriented and thus allows the manipulation of complex data structures in a condensed and eﬃcient manner. Its graphical abilities are also remarkable. R provides a powerful interface that can integrate programs written in other languages such as C, C++, Fortran, Perl, Python, and Java. At last, it is increasingly common to see people who develop new methodology simultaneously producing an R package in support of their approach and to back up introductory statistics courses with illustrations in R. One choice we have not addressed above is “why R and not BUGS?” BUGS (which stands for Bayesian inference Using Gibbs Sampling) is a Bayesian analysis software developed since the early 1990s, mostly by researchers from the Medical Research Council (MRC) at Cambridge University. The most common version is WinBugs, working under Windows, but there also exists an open-source version called OpenBugs. So, to return to the initial question, we are not addressing the possible links and advantages of BUGS simply because the purpose is diﬀerent. While access to Monte Carlo speciﬁcations is possible in BUGS, most computing operations are handled by the software itself, with the possible outcome that the user does not bother about this side of the problem and instead concentrates on Bayesian modeling. Thus, while R can be easily linked with BUGS and simulation can be done via BUGS, we think that a lower-level language such as R is more eﬀective in bringing you in touch. However, more advanced models like the hierarchical models cannot be easily handled by basic R programming and packages are not necessarily available to handle the variety of those models and call for other programming languages like JAGS. (JAGS standing for Just Another Gibbs Sampler and being dedicated to the study of Bayesian hierarchical models. This program is also freely available and distributed under the GNU Licence, the current version being JAGS 3.3.0.) 1.4.1 Getting Started The R language is straightforward to install: it can be downloaded (obviously free) from one of the numerous CRAN (Comprehensive R Archive Network) mirror Websites around the world.3 At this stage, we refrain from covering the installation of the R package and thus assume that (a) R is installed on the machine you want to work with and (b) that you have managed to launch it (in most cases, you simply have 2 3

Once again, R is a freely distributed and open-source language. The main CRAN Website is http://cran.r-project.org/.

1.4

An Introduction to R

7

to click on the proper icon). In the event you use a friendly (GUI) interface like RKWard, the interface opens several windows whose use should be selfexplanatory (along with a proper on-line help). Otherwise, you should then obtain a terminal window whose ﬁrst lines resemble the following, most likely with a more recent version: R version 2.14.1 (2011-12-22) Copyright (C) 2011 The R Foundation for Statistical Computing ISBN 3-900051-07-0 Platform: i686-pc-linux-gnu (32-bit) R is free software and comes with ABSOLUTELY NO WARRANTY. You are welcome to redistribute it under certain conditions. Type ’license()’ or ’licence()’ for distribution details. R is a collaborative project with many contributors. Type ’contributors()’ for more information and ’citation()’ on how to cite R or R packages in publications. Type ’demo()’ for some demos, ’help()’ for on-line help, or ’help.start()’ for an HTML browser interface to help. Type ’q()’ to quit R. > Neither this austere beginning nor the prospect of using a line editor should put you oﬀ, though, as there are many other ways of inputting and outputting commands and data, as we shall soon see! The ﬁnal line above with the symbol > means that the R software is waiting for a command from the user. This character > at the beginning of each line in the executable window is called the prompt and precedes the line command, which is terminated by pressing the RETURN key. At this early stage, all commands will be passed as line commands, and you should thus spot commands thanks to this symbol. Commands and programs that need to be stopped during their execution, for instance because they take too long or too much memory to complete or because they involve a programming mistake such as an inﬁnite loop, can be stopped by the Control-C double-key action without exiting the R session. For memory and eﬃciency reasons, R does not install all the available functions and programs when launched but only the basic packages that it requires to run properly. Additional packages can be loaded via the library command, as in > library(mnormt) # Multivariate Normal and t Distributions and the entire list of available packages is provided by library(). (The symbol # in the prompt lines above indicates a comment: All characters following # until the end of the command line are ignored. Comments are recommended to

8

1 User’s Manual

improve the readability of your programs.) There exist hundreds of packages available on the Web.4 Installing a new package such as the package mnormt is done by downloading the ﬁle from the Web depository and calling > install.package("mnormt") For a given package, the install.package command obviously needs to be executed only once, while the library call is required each time R is launched (as the corresponding package is not kept as part of the .RData ﬁle). Thus, it is good practice to include calls to required libraries within your R programs in order to avoid error messages when launching them. 1.4.2 R Objects As with many advanced programming languages, R distinguishes between several types of objects. Those types include scalar, vector, matrix, time series, data frames, functions, or graphics. An R object is mostly characterized by a mode that describes its contents and a class that describes its structure. The R function str applied to any R object, including R functions, will show its structure. For instance, > str(log) function (x, base = exp(1)) The diﬀerent modes are -

null (empty object), logical (TRUE or FALSE), numeric (such as 3, 0.14159, or 2+sqrt(3)), complex, (such as 3-2i or complex(1,4,-2)), and character (such as ‘‘Blue’’, ‘‘binomial’’, ‘‘male’’, or ‘‘y=a+bx’’),

and the main classes are vector, matrix, array, factor, time-series, data.frame, and list. Heterogeneous objects such as those of the list class can include elements with various modes. Manual entries about those classes can be obtained via the help commands help(data.frame) or ?matrix for instance. R can operate on most of those types as a regular function would operate on a scalar, which is a feature that should be exploited as much as possible for compact and eﬃcient programming. The fact that R is interpreted rather than compiled involves many subtle diﬀerences, but a major issue is that all variables in the system are evaluated and stored at every step of R programs. This means that loops in R are enormously time-consuming and should be avoided at all costs! Therefore, using the shortcuts oﬀered by R in the manipulation of vectors, matrices, and other structures is a must. 4

Packages that have been validated and tested by the R core team are listed at http://cran.r-project.org/src/contrib/PACKAGES.html.

1.4

An Introduction to R

9

The vector class As indicated logically by its name, the vector object corresponds to a mathematical vector of elements of the same type, such as (TRUE,TRUE,FALSE) or (1,2,3,5,7,11). Creating small vectors can be done using the R command c() as in > a=c(2,6,-4,9,18) This fundamental function combines or concatenates terms together. For instance, > d=c(a,b) concatenates the two vectors a and b into a new vector d. Note that decimal numbers should be encoded with a dot, character strings in quotes " ", and logical values with the character strings TRUE and FALSE or with their respective abbreviations T and F. Missing values are encoded with the character string NA. In Fig. 1.1, we give a few illustrations of the use of vectors in R. The character + indicates that the console is waiting for a supplementary instruction, which is useful when typing long expressions. The assignment operator is =, not to be confused with ==, which is the Boolean operator for equality. An older assignment operator is x a=c(5,5.6,1,4,-5)

build the object a containing a numeric vector of dimension 5 with elements 5, 5.6, 1, 4, –5

> a[1]

display the ﬁrst element of a

> b=a[2:4]

build the numeric vector b of dimension 3 with elements 5.6, 1, 4

> d=a[c(1,3,5)]

build the numeric vector d of dimension 3 with elements 5, 1, –5 multiply each element of a by 2

> 2*a

and display the result > b%%3

provides each element of b modulo 3

> d%/%2.4 > e=3/d

build the numeric vector5 e of dimension 3

computes the integer division of each element of d by 2.4 and elements 3/5, 3, –3/5 multiply the vectors d and e term by term

> log(d*e)

and transform each term into its natural logarithm > sum(d)

calculate the sum of d

> length(d)

display the length of d

> t(d)

transpose d, the result is a row vector

> t(d)%*%e

scalar product between the row vector t(b) and

> t(d)*e

element-wise product between two vectors

the column vector e with identical length with identical lengths > g=c(sqrt(2),log(10))

build the numeric vector g of dimension 2 √ and elements 2, log(10)

> e[d==5]

build the subvector of e that contains the components e[i] such that d[i]=5 create the subvector of a that contains

> a[-3]

all components of a but the third. > is.vector(d)

display the logical expression TRUE if a vector and FALSE else

Fig. 1.1. Illustrations of the processing of vectors in R

Note6 in Table 1.1 the convenient use of Boolean expressions to extract subvectors from a vector without having to resort to a component-bycomponent test (and hence a loop). The quantity d==5 is itself a vector of Booleans, while the number of components satisfying the constraint can be computed by sum(d==5). The ability to apply scalar functions to vectors as a whole is also a major advantage of R. In the event the function depends on a parameter or an option, this quantity can be entered as in 5 6

The variable e is not predeﬁned in R as exp(1). Positive and negative indices cannot be used simultaneously.

1.4

An Introduction to R

11

> e=lgamma(e^2) #warning: this is not the exponential basis, exp(1) which returns the vector with components log Γ (e2i ). Functions that are specially designed for vectors include, for instance, sample, order, sort and rank, which all have to do with manipulating the order in which the components of the vector occur. Besides their numeric and logical indexes, the components of a vector can also be identiﬁed by names. For a given vector x, names(x) is a vector of characters of the same length as x. This additional attribute is most useful when dealing with real data, where the components have a meaning such as "unemployed" or "democrat". Those names can also be erased by the command > names(x)=NULL

The : operator found in Fig. 1.1 is a very useful device that deﬁnes a consecutive sequence, but it is also fragile in that sequences do not always produce what is expected. For instance, 1:2*n corresponds to (1:2)*n rather than 1:(2*n).

The matrix, array, and factor classes The matrix class provides the R representation of matrices. A typical entry is, for instance, > x=matrix(vec,nrow=n,ncol=p) which creates an n × p matrix whose elements are those of the vector vec, assuming this vector is of dimension np. An important feature of this entry is that, in a somewhat unusual way, the components of vec are stored by column, which means that x[1,1] is equal to vec[1], x[2,1] is equal to vec[2], and so on, except if the option byrow=T is used in matrix. (Because of this choice of storage convention, working on R matrices column-wise is faster then working row-wise.) Note also that, if vec is of dimension n × p, it is not necessary to specify both the nrow=n and ncol=p options in matrix. One of those two parameters is suﬃcient to deﬁne the matrix. On the other hand, if vec is not of dimension n × p, matrix(vec,nrow=n,ncol=p) will create an n × p matrix with the components of vec repeated the appropriate number of times. For instance, > matrix(1:4,ncol=3) [,1] [,2] [,3] [1,] 1 3 1 [2,] 2 4 2 Warning message: data length [4] is not a submultiple or multiple of the number of columns [3] in matrix in: matrix(1:4, ncol = 3)

12

1 User’s Manual

produces a 2 × 3 matrix along with a warning message that something may be missing in the call to matrix. Note again that 1, 2, 3, 4 are entered consecutively when following the column (or lexicographic) order. Names can be given to the rows and columns of a matrix using the rownames and colnames functions. Note that, in some situations, it is useful to remember that an R matrix can also be used as a vector. If x is an n × p matrix, x[i+p*(j-1)] is equal to x[i,j], i.e., x can also be manipulated as a vector made of the columns of vec piled on top of one another. For instance, x[x>5] is a vector, while x[x>5]=0 modiﬁes the right entries in the matrix x. Conversely, vectors can be turned into p × 1 matrices by the command as.matrix. Note that x[1,] produces the ﬁrst row of x as a vector rather than as a p × 1 matrix. R allows for a wide range of manipulations on matrices, both termwise and in the classical matrix algebra perspective. For instance, the standard matrix product is denoted by %*%, while * represents the term-by-term product. (Note that taking the product a%*%b when the number of columns of a diﬀers from the number of rows of b produces an error message.) Figure 1.2 gives a few examples of matrix-related commands. The apply function is particularly easy to use for functions operating on matrices by row or column. > x1=matrix(1:20,nrow=5)

build the numeric matrix x1 of dimension 5 × 4 with ﬁrst row 1, 6, 11, 16

> x2=matrix(1:20,nrow=5,byrow=T)

build the numeric matrix x2 of dimension

> a=x3%*%x2

matrix summation of x2 and x3

> x3=t(x2)

transpose the matrix x2

5 × 4 with ﬁrst row 1, 2, 3, 4

matrix product between x2 and x3,

> b=x3%*%x2

with a check of the dimension compatibility > c=x1*x2

term-by-term product between x1 and x2

> dim(x1)

display the dimensions of x1

> b[,2]

select the second column of b

> b[c(3,4),]

select the third and fourth rows of b

> b[-2,]

delete the second row of b

> rbind(x1,x2)

vertical merging of x1 and x2

> cbind(x1,x2)

horizontal merging of x1 and x2

> apply(x1,1,sum)

calculate the sum of each row of x1

> as.matrix(1:10)

turn the vector 1:10 into a 10 × 1 matrix

Fig. 1.2. Illustrations of the processing of matrices in R

The function diag can be used to extract the vector of the diagonal elements of a matrix, as in diag(a), or to create a diagonal matrix with a given diagonal, as in diag(1:10). Since matrix algebra is central to good

1.4

An Introduction to R

13

programming in R, as matrix programming allows for the elimination of timeconsuming loops, it is important to be familiar with matrix manipulation. For instance, the function crossprod replaces the product t(x)%*%y on either vectors or matrices by crossprod(x,y) more eﬃciently: > system.time(crossprod(1:10^6,1:10^6)) user system elapsed 0.016 0.048 0.066 > system.time(t(1:10^6)%*%(1:10^6)) user system elapsed 0.084 0.036 0.121 Eigen-analysis of square matrices is also included in the base package. For instance, chol(m) returns the upper triangular factor of the Choleski decomposition of m; that is, the matrix R such that RT R is equal to m. Similarly, eigen(m) returns a list that contains the eigenvalues of m (some of which can be complex numbers) as well as the corresponding eigenvectors (some of which are complex if there are complex eigenvalues). Related functions are svd and qr, which provide the singular values and the QR decomposition of their argument, respectively. Note that the inverse M−1 of a matrix M can be found either by solve(M) (recommended) or ginv(M), which requires downloading the library MASS and also produces generalized inverses (which may be a mixed blessing since the fact that a matrix is not invertible is not signaled by ginv). Special versions of solve are backsolve and forwardsolve, which are restricted to upper and lower diagonal triangular systems, respectively. Note also the alternative of using chol2inv which returns the inverse of a matrix m when provided by the Choleski decomposition chol(m). Structures with more than two indices are represented by arrays and can also be processed by R commands, for instance x=array(1:50,c(2,5,5)), which gives a three-entry table of 50 terms. Once again, they can also be interpreted as vectors. The apply function used in Fig. 1.2 is a very powerful device that operates on arrays and, in particular, matrices. Since it can return arrays, it bypasses calls to multiple loops and makes for (sometimes) quicker and (always) cleaner programs. It should not be considered as a panacea, however, as apply hides calls to loops inside a single command. For instance, a comparison of apply(A, 1,mean) with rowMeans(A) shows the second version is about 200 times faster. Using linear algebra whenever possible is therefore a more eﬃcient solution. Spector (2009, Sect. 8.7) gives a detailed analysis of the limitations of apply and the advantages of vectorization in R. A factor is a vector of characters or integers used to specify a discrete classiﬁcation of the components of other vectors with the same length. Its main diﬀerence from a standard vector is that it comes with a level attribute used to specify the possible values of the factor. This structure is therefore appropriate to represent qualitative variables. R provides both ordered and

14

1 User’s Manual

unordered factors, whose major appeal lies within model formulas, as illustrated in Fig. 1.3. Note the subtle diﬀerence between apply and tapply. > state=c("tas","tas","sa","sa","wa") create a vector with ﬁve values > statef=factor(state)

distinguish entries by group

> levels(statef)

give the groups

> incomes=c(60,59,40,42,23)

create a vector of incomes

> tapply(incomes,statef,mean)

average the incomes for each group

> statef=factor(state,

deﬁne a new level with one more

+ levels=c("tas","sa","wa","yo"))

group than observed

> table(statef)

return statistics for all levels

Fig. 1.3. Illustrations of the factor class

The list and data.frame classes A list in R is a rather loose object made of a collection of other arbitrary objects known as its components.7 For instance, a list can be derived from n existing objects using the function list: a=list(name_1=object_1,...,name_n=object_n) This command creates a list with n arguments using object_1,...,object_n for the components, each being associated with the argument’s name, name_i. For instance, a$name_1 will be equal to object_1. (It can also be represented as a[[1]], but this is less practical, as it requires some bookkeeping of the order of the objects contained in the list.) Lists are very useful in preserving information about the values of variables used within R functions in the sense that all relevant values can be put within a list that is the output of the corresponding function (see Sect. 1.4.5 for details about the construction of functions in R). Most standard functions in R, for instance eigen in Fig. 1.4, return a list as their output. Note the use of the abbreviations vec and val in the last line of Fig. 1.4. Such abbreviations are acceptable as long as they do not induce confusion. (Using res$v would not work!) The local version of apply is lapply, which computes a function for each argument of the list > x = list(a = 1:10, beta = exp(-3:3), + logic = c(TRUE,FALSE,FALSE,TRUE)) > lapply(x,mean) #compute the empirical means $a [1] 5.5 7

Lists can contain lists as elements.

1.4

An Introduction to R

> li=list(num=1:5,y="color",a=T)

create a list with three arguments

> a=matrix(c(6,2,0,2,6,0,0,0,36),nrow=3)

create a (3,3) matrix

> res=eigen(a,symmetric=T)

diagonalize a and

> names(res)

produce a list with two

> res$vectors

vectors arguments of res

> diag(res$values)

create the diagonal matrix

15

arguments: vectors and values

of eigenvalues > res$vec%*%diag(res$val)%*%t(res$vec)

recover a

Fig. 1.4. Chosen features of the list class

$beta [1] 4.535125 $logic [1] 0.5 provided each argument is of a mode that is compatible with the function argument (i.e., is numeric in this case). A “user-friendly” version of lapply is sapply, as in > sapply(x,mean) a beta logic 5.500000 4.535125 0.500000 The last class we brieﬂy mention here is the data.frame. A data frame is a list whose elements are possibly made of diﬀering modes and attributes but have the same length, as in the example provided in Fig. 1.5. A data frame can be displayed in matrix form, and its rows and columns can be extracted using matrix indexing conventions. A list whose components satisfy the restrictions imposed on a data frame can be coerced into a data frame using the function as.data.frame. The main purpose of this object is to import data from an external ﬁle by using the read.table function. 1.4.3 Probability Distributions in R R is primarily a statistical language. It is therefore well-equipped with probability distributions. As described in Table 1.1, all standard distributions are available, with a clever programming shortcut: A “core” name, such as norm, is associated with each distribution and the four basic associated functions, namely the cdf, the pdf, the quantile function, and the simulation procedure, are deﬁned by appending the preﬁxes d, p, q, r to the core name, such as dnorm(), pnorm(), qnorm(), and rnorm(). Obviously, each function requires

16

1 User’s Manual > v1=sample(1:12,30,rep=T) > v2=sample(LETTERS[1:10],30,rep=T) > v3=runif(30) > v4=rnorm(30) > xx=data.frame(v1,v2,v3,v4)

simulate 30 independent uniform random variables on {1, 2, ..., 12} simulate 30 independent uniform random variables on {a, b, ...., j} simulate 30 independent uniform random variables on [0, 1] simulate 30 independent realizations from a standard normal distribution create a data frame

Fig. 1.5. Deﬁnition of a data.frame

additional entries, as in pnorm(1.96) or rnorm(10,mean=3,sd=3). Recall that pnorm() and qnorm() are inverses of one another. Table 1.1. Standard distributions with R core name Distribution Beta Binomial Cauchy Chi-square Exponential Fisher Gamma Geometric Hypergeometric Log-Normal Logistic Normal Poisson Student Uniform Weibull

Core beta binom cauchy chisq exp f gamma geom hyper lnorm logis norm pois t unif weibull

Parameters shape1, shape2 size, prob location, scale df 1/mean df1, df2 shape,1/scale prob m, n, k mean, sd location, scale mean, sd lambda df min, max shape

Default values

0, 1 1 NA, 1

0, 1 0, 1 0, 1

0, 1

In addition to these probability functions, R also provides a battery of (classical) statistical tools, ranging from descriptive statistics to nonparametric tests and generalized linear models. A description of these abilities is not possible in this section but we refer the reader to, e.g., Dalgaard (2002) or Venables and Ripley (2002) for a complete entry. 1.4.4 Graphical Facilities Another clear advantage of using the R language is that it allows a very rich range of graphical possibilities. Functions such as plot and image can be customized to a large extent, as described in Venables and Ripley (2002) or Murrell (2005) (the latter being entirely dedicated to the R graphic abilities). Even though the default output of plot as for instance in

1.4

An Introduction to R

17

> plot(faithful) is not highly most enticing, plot is incredibly ﬂexible: To see the number of parameters involved, you can type par() that delivers the default values of all those parameters.

The wealth of graphical possibilities oﬀered by R should be taken advantage of cautiously! That is, good design avoids clutter, small fonts, unreadable scale, etc. The recommendations found in Tufte (2001) are thus worth following to avoid horrid outputs like those often found in some periodicals! In addition, graphs produced by R usually tend to look nicer on the current device than when printed or included in a slide presentation. Colors may change, font sizes may turn awkward, separate curves may end up overlapping, and so on.

Before covering the most standard graphic commands, we start by describing the notion of device that is at the core of those graphic commands. Each graphical operation sends its outcome to a device, which can be a graphical window (like the one that automatically appears when calling a graphical command for the ﬁrst time as in the example above) or a ﬁle where the graphical outcome is stored for printing or other uses. Under Unix, Linux and mac OS, launching a new graphical window can be done via X11(), with many possibilities for customization (such as size, positions, color, etc.). Once a graphical window is created, it is given a device number and can be managed by functions that start with dev., such as dev.list, dev.set, and others. An important command is dev.off, which closes the current graphical window. When the device is a ﬁle, it is created by a function that is named after its driver. There are therefore a postscript, a pdf, a jpeg, and a png function. When printing to a ﬁle, as in the following example, > > > + > + >

pdf(file="faith.pdf") par(mfrow=c(1,2),mar=c(4,2,2,1)) hist(faithful[,1],nclass=21,col="grey",main="", xlab=names(faithful)[1]) hist(faithful[,2],nclass=21,col="wheat",main="", xlab=names(faithful)[2]) dev.off()

closing the sequence with dev.off() is compulsory since it completes the ﬁle, which is then saved. If the command pdf(file="faith.pdf") is repeated, the earlier version of the pdf ﬁle is erased. Of course, using a line command interface for controlling graphics may seem antiquated, but this is the consequence of the R object-oriented philosophy. In addition, current graphs can be saved to a postscript ﬁle using the dev.copy and dev.print functions. Note that R-produced graphs tend to be large objects, in part because the graphs are not pictures of the current state but instead preserve every action ever taken.

18

1 User’s Manual

As already stressed above, plot is a highly versatile tool that can be used to represent functional curves and two-dimensional datasets. Colors (chosen by colors() or colours() out of 650 hues), widths, and types can be calibrated at will and LATEX-like formulas can be included within the graphs using expression. Text and legends can be included at a speciﬁc point with locator (see also identify) and legend. An example of (relatively simple) output is > + > > > + > > +

plot(as.vector(time(mdeaths)),as.vector(mdeaths),cex=.6, pch=19,xlab="",ylab="Monthly deaths from bronchitis") lines(spline(mdeaths),lwd=2,col="red",lty=3) ar=arima(mdeaths,order=c(1,0,0))$coef lines(as.vector(time(mdeaths))[-1], ar[2]+ar[1]* (mdeaths[-length(mdeaths)]-ar[2]),col="blue",lwd=2,lty=2) title("Splines versus AR(1) predictor") legend(1974,2800,legend=c("spline","AR(1)"),col=c("red", "blue"),lty=c(3,2),lwd=c(2,2),cex=.5)

represented in Fig. 1.6, which compares spline ﬁtting to an AR(1) predictor and to an SAR(1,12) predictor. Note that the seasonal model is doing worse.

1500

2000

2500

spline AR(1) SAR(1,12)

1000

Monthly deaths from bronchitis

Splines versus AR(1) predictor and SAR(1,12) predictor

1974

1975

1976

1977

1978

1979

1980

Fig. 1.6. Monthly deaths from bronchitis in the UK over the period 1974–1980 and ﬁts by a spline approximation and an AR predictor

1.4

An Introduction to R

19

Useful graphical functions include – – – – – –

hist for constructing and optionally plotting histograms of datasets; points for adding points on an existing graph; lines for linking points together on an existing graph; polygon for ﬁlling the area between two sets of points; barplot for creating barplots; boxplot for creating boxplots.

The two-dimensional representations oﬀered by image and contour are quite handy when providing likelihood or posterior surfaces. Figure 1.7 gives some of the most usual graphical commands. > x=rnorm(100) > hist(x,nclass=10, prob=T) > > > > >

curve(dnorm(x),add=T) y=2*x+rnorm(100,0,2) plot(x,y,xlim=c(-5,5),ylim=c(-10,10)) lines(c(0,0),c(1,2),col="sienna3") boxplot(x)

compute and plot an histogram of x draw the normal density on top draw a scatterplot of x against y compute and plot a box-and-whiskers plot of x

> state=c("tas","tas","sa","sa","wa","sa") > statef=factor(state) > barplot(table(statef)) draw a bar diagram of x

Fig. 1.7. Some standard plotting commands

1.4.5 Writing New R Functions One of the strengths of R is that new functions and libraries can be created by anyone and then added to Web depositories to continuously enrich the language. These new functions are not distinguishable from the core functions of R, such as median() or var(), because those are also written in R. This means their code can be accessed and potentially modiﬁed, although it is safer to deﬁne new functions. (A few functions are written in C, though, for eﬃciency.) Learning how to write functions designed for one’s own problems is paramount for their resolution, even though the huge collection of available R functions may often contain a function already written for that purpose. A function is deﬁned in R by an assignment of the form name=function(arg1[=expr1],arg2[=expr2],...) { expression ... expression value }

20

1 User’s Manual

where expression denotes an R command that uses some of the arguments arg1, arg2, ... to calculate a value, value, that is the outcome of the function. The braces indicate the beginning and the end of the function and the brackets some possible default values for the arguments. Note that producing a value at the end of a function is essential because anything done within a function is local and temporary, and therefore lost once the function has been exited, unless saved in value (hence, again, the appeal of list()). For instance, the following function, named sqrnt, implements a version of Newton’s method for calculating the square root of y: sqrnt=function(y) { x=y/2 while (abs(x*x-y) > 1e-10) x=(x+y/x)/2 x } When designing a new R function, it is more convenient to use an external text editor and to store the function under development in an external ﬁle, say myfunction.R, which can be executed in R as source("myfunction.R"). Note also that some external commands can be launched within an R function via the very handy command system(). This is, for instance, the easiest way to incorporate programs written in other languages (e.g., Fortran, C, Matlab) within R programs. Without getting deeply into R programming, let us note a distinction between global and local variables: the former are deﬁned in the core of the R code and are recognized everywhere, while the later are only deﬁned within a speciﬁc function. This means in particular that a local variable, locax say, initialized within a function, myfunc say, will not be recognized outside myfunc. (It will not even be recognized in a function deﬁned within myfunc.) The expressions used in a function rely on a syntax that is quite similar to those of other programming languages, with conditional statements such as if (expres1) expres2 else expres3 where expres1 is a logical value, and loops such as for (name in expres1) expres2 and while (name in expres1) expres2 where expres1 is a collection of values, as illustrated in Fig. 1.8. In particular, Boolean operators can be used within those expressions, including == for testing equality, != for testing inequality, & for the logical and, | for the logical or, and ! for the logical contradiction. Since R is an interpreted language, avoiding loops is generally a good idea, but this may render programs much harder to read. It is therefore extremely useful to include comments within the programs by using the symbol #.

1.4

An Introduction to R

> bool=T;i=0

separate commands by semicolons

> while(bool==T) {i=i+1; bool=(i s=0;x=rnorm(10000) > system.time(for (i in 1:length(x)){

output sum(x) and

+ s=s+x[i]})[3]

provide computing time

> system.time(t(rep(1,10000))%*%x)[3]

compare with vector product

> system.time(sum(x))[3]

compare with sum() eﬃciency

Fig. 1.8. Some artiﬁcial loops in R

1.4.6 Input and Output in R Large data objects need to be read as values from external ﬁles rather than entered during an R session at the keyboard (or by cut-and-paste). Input facilities are simple, but their requirements are fairly strict. In fact, there is a clear presumption that it is possible to modify input ﬁles using other tools outside R. An entire data frame can be read directly with the read.table() function. Plain ﬁles containing rows of values with a single mode can be downloaded using the scan() function, as in > a=matrix(scan("myfile"),nrow=5,byrow=T) When data frames have been produced by another statistical software, the library foreign can be used to input those frames in R. For example, the function read.spss() allows ones to read SPSS data frames. Conversely, the generic function save() can be used to store all R objects in a given ﬁle, either in binary or ASCII format. (The alternative function dump() is more rudimentary but also useful.) The function write.table() is used to export R data frames as ASCII ﬁles. 1.4.7 Administration of R Objects During an R session, objects are created and stored by name. The command objects() (or, alternatively, ls()) can be used to display, within a directory called the workspace, the names of the objects that are currently stored. Individual objects can be deleted with the function rm(). Removing all objects created so far is done by rm(list=ls()). All objects created during an R session (including functions) can be stored permanently in a ﬁle in provision of future R sessions. At the end of each R session, obtained by the command quit() (which can be abbreviated as q()), the user is given the opportunity to save all the currently available objects, as in >q() Save workspace image? [y/n/c]:

22

1 User’s Manual

If the user answers y, the object created during the current session and those saved from earlier sessions are saved in a ﬁle called .RData and located in the working directory. When R is called again, it reloads the workspace from this ﬁle, which means that the user starts the new session exactly where the old one had stopped. In addition, the entire past command history is stored in the ﬁle .Rhistory and can be used in the current or in later sessions by using the command history().

1.5 The bayess Package Since this is originally a paper book, copying by hand the R code represented on the following pages to your computer terminal would be both tedious and time-wasting. We have therefore gathered all the programs and codes of this book within an R package called bayess that you should download from CRAN before proceeding to the next chapter. Once downloaded on your computer following the instructions provided on the CRAN Webpage, the package bayess is loaded into your current R session by library(bayess). All the functions deﬁned inside the package are then available, and so is a step-by-step reproduction of the examples provided in the book, using the demo command: > demo(Chapter.1) demo(Chapter.1) ---- ~~~~~~~~~ Type

to start :

> # Chapter 1 R commands > > # Section 1.4.2 > > str(log) function (x, base = exp(1)) > a=c(2,6,-4,9,18) > x d=a[c(1,3,5)] > e=3/d > e=lgamma(e^2)

1.5

> S=readline(prompt="Type Type to continue :

The bayess Package

23

to continue : ")

and similarly for the following chapters. Obviously, all commands contained in the demonstrations and all functions deﬁned in the package can be accessed and modiﬁed.

Although most steps of the demonstrations are short, some may require longer execution times. If you need to interrupt the demonstration, recall that Ctrl-C is an interruption command.

2 Normal Models

This was where the work really took place. —Ian Rankin, Knots & Crosses.—

Roadmap This chapter uses the standard normal N (μ, σ 2 ) distribution as an easy entry to generic Bayesian inferential methods. As in every subsequent chapter, we start with a description of the data used as a chapter benchmark for illustrating new methods and for testing assimilation of the techniques. We then propose a corresponding statistical model centered on the normal distribution and consider speciﬁc inferential questions to address at this level, namely parameter estimation, model choice, and outlier detection, once set the description of the Bayesian resolution of inferential problems. As beﬁts a ﬁrst chapter, we also introduce here general computational techniques known as Monte Carlo methods.

J.-M. Marin and C.P. Robert, Bayesian Essentials with R, Springer Texts in Statistics, DOI 10.1007/978-1-4614-8687-9 2, © Springer Science+Business Media New York 2014

25

26

2 Normal Models

2.1 Normal Modeling The normal (or Gaussian) distribution N (μ, σ 2 ), with density on R, 1 1 f (x|μ, σ) = √ exp − 2 (x − μ)2 , 2σ 2πσ is certainly one of the most studied and one of the most used distributions because of its “normality”: It appears both as the limit of additive small eﬀects and as a representation of symmetric phenomena without long tails, and it oﬀers many openings in terms of analytical properties and closed-form computations. As such, it is thus the natural opening to a modeling course, even more than discrete and apparently simpler models such as the binomial and Poisson models we will discuss in the following chapters. Note, however, that we do not advocate at this stage the use of the normal distribution as a one-ﬁts-all model: There exist many continuous situations where a normal model is inappropriate for many possible reasons (e.g., skewness, fat tails, dependence, multimodality).

Fig. 2.1. Dataset normaldata: Histogram of the observed fringe shifts in Illingworth’s 1927 experiment

2.1

Normal Modeling

27

Our normal dataset, normaldata, is linked with the famous Michelson– Morlay experiment that opened the way to Einstein’s relativity theory in 1887. The experiment was intended to detect the “æther ﬂow” and hence the existence of æther, this theoretical medium physicists postulated at this epoch was necessary to the transmission of light. Michelson’s measuring device consisted in measuring the diﬀerence in the speeds of two light beams travelling the same distance in two orthogonal directions. As often in physics, the measurement was done by interferometry and diﬀerences in the travelling time inferred from shift in the fringes of the light spectrum. However, the experiment produced very small measurements that were not conclusive for the detection of the æther. Later experiments tried to achieve higher precision, as the one by Illingworth in 1927 used here as normaldata, only to obtain smaller and smaller upper bounds on the æther windspeed. While the original dataset is available in R as morley, the entries are approximated to the nearest multiple of ten and are therefore diﬃcult to analyze as normal observations. The 64 data points in normaldata are associated with session numbers (ﬁrst column), corresponding to diﬀerent times of the day, and the values in the second column represent the averaged fringe displacement due to orientation taken over ten measurements made by Illingworth, who assumed a normal error model. Figure 2.1 produces an histogram of the data by the simple R commands > data(normaldata) > shift=normaldata[,2] > hist(shift,nclass=10,col="steelblue",prob=TRUE,main="") This histogram seems compatible with a symmetric unimodal distribution such as the normal distribution. As shown in Fig. 2.2 by a qq-plot obtained by the commands > qqnorm((shift-mean(shift))/sd(shift),pch=19,col="gold2") > abline(a=0,b=1,lty=2,col="indianred",lwd=2) which compare the empirical cdf with a pluggin normal estimate, The N (μ, σ 2 ) ﬁt may not be perfect, though, because of (a) a possible bimodality of the histogram and (b) potential outliers. As mentioned above, the use of a normal distribution for modeling a given dataset is a convenient device that does not need to correspond to a perfect ﬁt. With some degree of approximation, the normal distribution may agree with the data suﬃciently to be used in place of the true distribution (if any). There exist, however, some setups where the normal distribution is thought to be the exact distribution behind the dataset (or where departure from normality has a signiﬁcance for the theory behind the observations). In Marin and Robert (2007), we introduced a huge dataset related to the astronomical concept of

28

2 Normal Models

Fig. 2.2. Dataset normaldata: qq-plot of the observed fringe shifts against the normal quantiles

the cosmological background noise that illustrated this point, but chose not to reproduce the set in this edition due to the diﬃculty in handling it.

2.2 The Bayesian Toolkit 2.2.1 Posterior Distribution Given an independent and identically distributed (later abbreviated as iid) sample Dn = (x1 , . . . , xn ) from a density f (x|θ), depending upon an unknown parameter θ ∈ Θ, for instance the mean μ of the benchmark normal distribution, the associated likelihood function is (θ|Dn ) =

n

f (xi |θ) .

(2.1)

i=1

This function of θ is a fundamental entity for the analysis of the information provided about θ by the sample Dn , and Bayesian analysis relies on (2.1) to draw its inference on θ. For instance, when Dn is a normal N (μ, σ 2 ) sample of size n and θ = (μ, σ 2 ), we get (θ|Dn ) =

n i=1

√ exp{−(xi − μ)2 /2σ 2 }/ 2πσ

2.2

∝ exp −

The Bayesian Toolkit

29

(xi − μ)2 /2σ 2

/σ n

i=1

∝ exp − nμ2 − 2n¯ xμ +

x2i

2σ 2

/σ n

i=1

∝ exp − n(μ − x¯)2 + s2 2σ 2 /σ n , n where x ¯ denotes the empirical mean and where s2 is the sum i=1 (xi − x¯)2 . This shows in particular that x ¯ and s2 are suﬃcient statistics.

In

the above display of equations, the sign ∝ means proportional to. This proportionality is understood for functions of θ, meaning that the discarded constants do not depend on θ but may well depend on the data Dn . This shortcut is both handy in complex Bayesian derivations and fraught with danger when considering several levels of parameters.

The major input of the Bayesian approach, compared with a traditional likelihood approach, is that it modiﬁes the likelihood function into a posterior distribution, which is a valid probability distribution on Θ deﬁned by the classical Bayes’ formula (or theorem) π(θ|Dn ) =

(θ|Dn )π(θ) . (θ|Dn )π(θ) dθ

(2.2)

The factor π(θ) in (2.2) is called the prior and it obviously has to be chosen to start the analysis.

The posterior density is a probability density on the parameter, which does not mean the parameter θ need be a genuine random variable. This density is used as an inferential tool, not as a truthful representation.

A ﬁrst motivation for this approach is that the prior distribution summarizes the prior information on θ; that is, the knowledge that is available on θ prior to the observation of the sample Dn . However, the choice of π(θ) is often decided on practical grounds rather than strong subjective beliefs or overwhelming prior information. A second motivation for the Bayesian construct is therefore to provide a fully probabilistic framework for the inferential analysis, with respect to a reference measure π(θ). As an illustration, consider the simplest case of the normal distribution with known variance, N (μ, σ 2 ). If the prior distribution on μ, π(μ), is the 2 normal N 0, σ , the posterior distribution is easily derived via Bayes’ theorem π(μ|Dn ) ∝ π(μ) (θ|Dn )

x − μ)2 /2σ 2 ∝ exp{−μ2 /2σ 2 } exp −n(¯

30

2 Normal Models

∝ exp −(n + 1)μ2 /2σ 2 + 2nμ¯ x/2σ 2 ∝ exp −(n + 1)[μ − n¯ x/(n + 1)]2 /2σ 2 , which means that this posterior distribution in μ is a normal distribution with mean n¯ x/(n + 1) and variance σ 2 /(n + 1). The mean (and mode) of the posterior is therefore diﬀerent from the classical estimator x ¯, which may seem as a paradoxical feature of this Bayesian analysis. The reason for the diﬀerence is that the prior information that μ is close enough to zero is taken into account by the posterior distribution, which thus shrinks the original estimate towards zero. If we were given an alternative information that μ was close to ten, the posterior distribution would similarly shrink μ towards ten. The change from a factor n to a factor (n + 1) in the (posterior) variance is similarly explained by the prior information, in that accounting for this information reduces the variability of our answer. For normaldata, we can ﬁrst assume that the value of σ is the variability of the Michelson–Morley apparatus, namely 0.75. In that case, the posterior distribution on the fringe shift average μ is a normal N (n¯ x/(n+1), σ 2 /(n+1)) distribution, hence with mean and variance > n=length(shift) > mmu=sum(shift)/(n+1); mmu [1] -0.01461538 > vmu=0.75^2/(n+1); vmu [1] 0.008653846 represented on Fig. 2.3 as a dotted curve. The case of a normal distribution with a known variance being quite unrealistic, we now consider the general case of an iid sample Dn = (x1 , . . . , xn ) 2 2 from the normal distribution N (μ, σ ) and θ = (μ, σ ). Keeping the same 2 prior distribution N 0, σ on μ, which then appears as a conditional distribution of μ given σ 2 , i.e., relies on the generic decomposition π(μ, σ 2 ) = π(μ|σ 2 )π(σ 2 ) , we have to introduce a further prior distribution on σ 2 . To make computations simple at this early stage, we choose an exponential E (1) distribution on σ −2 . This means that the random variable ω = σ −2 is distributed from an exponential E (1) distribution, the distribution on σ 2 being derived by the usual change of variable technique, −2 dσ = exp(−σ −2 ) (σ 2 )−2 . π(σ 2 ) = exp(−σ −2 ) dσ 2 (This distribution is a special case of an inverse gamma distribution, namely IG(1, 1).) The corresponding posterior density on θ is then given by

2.2

The Bayesian Toolkit

31

π((μ, σ 2 )|Dn ) ∝ π(σ 2 ) × π(μ|σ 2 ) × ((μ, σ 2 )|Dn ) ∝ (σ −2 )1/2+2 exp −(μ2 +2)/2σ 2 ×(σ −2 )n/2 exp − n(μ − x)2 +s2 /2σ 2

∝ (σ 2 )−(n+5)/2 exp − (n+1)(μ − n¯ x/(n+1))2 +(2+s2 ) /2σ 2 ∝ (σ 2 )−1/2 exp −(n+1)[μ − n¯ x/(n+1)]2 /2σ 2 . ×(σ 2 )−(n+2)/2−1 exp −(2+s2 )/2σ 2 . Therefore, the posterior on θ can be decomposed as the product of an inverse gamma distribution on σ 2 , I G ((n + 2)/2, [2 + s2 ]/2)—which is the distribution of the inverse of a gamma G ((n + 2)/2, [2 + s2 ]/2) random variable—and, conditionally on σ 2 , a normal distribution on μ, N (n¯ x/(n + 1), σ 2 /(n + 1)). The interpretation of this posterior is quite similar to the case when σ is known, with the diﬀerence that the variability in σ induces more variability in μ, the marginal posterior in μ being then a Student’s t distribution1 (Exercise 2.1) μ|Dn ∼ T n + 2, n¯ x/(n + 1), (2 + s2 )/(n + 1)(n + 2) , with n + 2 degrees of freedom, a location parameter proportional to x ¯ and a scale parameter (almost) proportional to s. For normaldata, an E xp(1) prior on σ −2 being compatible with the value observed on the Michelson–Morley experiment, the parameters of the t distribution on μ are therefore n = 64, > mtmu=sum(shift)/(n+1);mtmu [1] -0.01461538 > stmu=(2+(n-1)*var(shift))/((n+2)*(n+1));stmu [1] 0.0010841496 We compare the resulting posterior with the one based on the assumption σ = 0.75 on Fig. 2.3, using the curve commands (note that the mnormt library may require the preliminary installation of the corresponding package by install.packages("mnormt")): > > + > +

1

library(mnormt) curve(dmt(x,mean=mmu,S=stmu,df=n+2),col="chocolate2",lwd=2, xlab="x",ylab="",xlim=c(-.5,.5)) curve(dnorm(x,mean=mmu,sd=sqrt(vmu)),col="steelblue2", lwd=2,add=TRUE,lty=2)

We will omit the reference to Student in the subsequent uses of this distribution, as is the rule in anglo-saxon textbooks.

32

2 Normal Models

Fig. 2.3. Dataset normaldata: Two posterior distributions on μ corresponding to an hypothetical σ = 0.75 (dashed lines) and to an unknown σ 2 under the prior σ −2 ∼ E (1) (plain lines)

Although this may sound counterintuitive, in this very case, estimating the variance produces a reduction in the variability of the posterior distribution on μ. This is because the postulated value of σ 2 is actually inappropriate for Illingworth’s experiment, being far too large. Since the posterior distribution on σ 2 is an I G (33, 1.82) distribution for normaldata, the probability that σ is as large as 0.75 can be evaluated as > digmma=function(x,shape,scale){dgamma(1/x,shape,scale)/x^2} > curve(digmma(x,shape=33,scale=(1+(n+1)*var(shift))/2), + xlim=c(0,.2),lwd=2) > pgamma(1/(.75)^2,shape=33,scale=(1+(n+1)*var(shift))/2) [1] 8.99453e-39 which shows that 0.75 is quite unrealistic, being ten times as large as the mode of the posterior density on σ 2 . The above R command library(mnormt) calls the mnormt library, which contains useful additional functions related with multivariate normal and t distributions. In particular, dmt allows for location and scale parameters in the t distribution. Note also that s2 is computed as (n-1)*var(shift) because R implicitly adopts a classical approach in using the “best unbiased estimator” of σ 2 .

2.2

The Bayesian Toolkit

33

2.2.2 Bayesian Estimates A concept that is at the core of Bayesian analysis is that one should provide an inferential assessment conditional on the realized value of Dn . Bayesian analysis gives a proper probabilistic meaning to this conditioning by allocating to θ a probability distribution. Once the prior distribution is selected, Bayesian inference formally is “over”; that is, it is completely determined since the estimation, testing, and evaluation procedures are automatically provided by the prior and the way procedures are compared (or penalized). For instance, if estimations θˆ of θ are compared via the sum of squared errors, ˆ = θ − θ ˆ 2, L(θ, θ) the corresponding Bayes optimum is the expected value of θ under the posterior distribution,2 θ (θ|Dn ) π(θ) dθ ˆ , (2.3) θ = θ π(θ|Dn ) dθ = (θ|Dn ) π(θ) dθ for a given sample Dn . When no speciﬁc penalty criterion is available, the estimator (2.3) is often used as a default estimator, although alternatives are also available. For instance, the maximum a posteriori estimator (MAP) is deﬁned as θˆ = arg maxθ π(θ|Dn ) = arg maxθ π(θ)(θ|Dn ),

(2.4)

where the function to maximize is usually provided in closed form. However, numerical problems often make the optimization involved in ﬁnding the MAP far from trivial. Note also here the similarity of (2.4) with the maximum likelihood estimator (MLE): The inﬂuence of the prior distribution π(θ) on the estimate progressively disappears as the number of observations n increases, and the MAP estimator often recovers the asymptotic properties of the MLE. For normaldata, since the posterior distribution on σ −2 is a G (32, 1.82) distribution, the posterior expectation of σ −2 given Illingworth’s experimental data is 32/1.82 = 17.53. The posterior expectation of σ 2 requires a supplementary eﬀort in order to derive the mean of an inverse gamma distribution (see Exercise 2.2), namely Eπ [σ 2 |Dn ] = 1.82/(33 − 1) = 0.057 .

2 Estimators are functions of the data Dn , while estimates are values taken by those functions. In most cases, we will denote them with a “hat” symbol, the dependence on Dn being implicit.

34

2 Normal Models

Similarly, the MAP estimate is given here by arg maxθ π(σ 2 |Dn ) = 1.82/(33 + 1) = 0.054 (see also Exercise 2.2). These values therefore reinforce our observation that the Michelson–Morley precision is not appropriate for the Illingworth experiment, which is much more precise indeed. 2.2.3 Conjugate Prior Distributions The selection of the prior distribution is an important issue in Bayesian statistics. When prior information is available about the data or the model, it can (and must) be used in building the prior, and we will see some implementations of this recommendation in the following chapters. In many situations, however, the selection of the prior distribution is quite delicate, due to the absence of reliable prior information, and default solutions must be chosen instead. Since the choice of the prior distribution has a considerable inﬂuence on the resulting inference, this inferential step must be conducted with the utmost care. From a computational viewpoint, the most convenient choice of prior distributions is to mimic the likelihood structure within the prior. In the most advantageous cases, priors and posteriors remain within the same parameterized family. Such priors are called conjugate. While the foundations of this principle are too advanced to be processed here (see, e.g., Robert, 2007, Chap. 3), such priors exist for most usual families, including the normal distribution. Indeed, as seen in Sect. 2.2.1, when the prior on a normal mean is normal, the corresponding posterior is also normal. Since conjugate priors are such that the prior and posterior densities belong to the same parametric family, using the observations boils down to an update of the parameters of the prior. To avoid confusion, the parameters involved in the prior distribution on the model parameter are usually called hyperparameters. (They can themselves be associated with prior distributions, then called hyperpriors.) For most practical purposes, it is suﬃcient to consider the conjugate priors described in Table 2.1. The derivation of each row is straightforward if painful and proceeds from the same application of Bayes’ formula as for the normal case above (Exercise 2.5). For distributions that are not within this table, a conjugate prior may or may not be available (Exercise 2.6). An important feature of conjugate priors is that one has a priori to select two hyperparameters, e.g., a mean and a variance in the normal case. On the one hand, this is an advantage when using a conjugate prior, namely that one has to select only a few parameters to determine the prior distribution. On the other hand, this is a drawback in that the information known a priori on μ may be either insuﬃcient to determine both parameters or incompatible with the structure imposed by conjugacy.

2.2

The Bayesian Toolkit

35

Table 2.1. Conjugate priors for the most common statistical families f (x|θ) Normal N (θ, σ 2 )

π(θ) Normal N (μ, τ 2 )

Poisson P(θ) Gamma G (ν, θ) Binomial B(n, θ) Negative Binomial N eg(m, θ) Multinomial Mk (θ1 , . . . , θk ) Normal N (μ, 1/θ)

Gamma G (α, β) Gamma G (α, β) Beta Be(α, β) Beta Be(α, β) Dirichlet D(α1 , . . . , αk ) Gamma G (α, β)

π(θ|x) N (ρ(σ 2 μ + τ 2 x), ρσ 2 τ 2 ) ρ−1 = σ 2 + τ 2 G (α + x, β + 1) G (α + ν, β + x) Be(α + x, β + n − x) Be(α + m, β + x) D(α1 + x1 , . . . , αk + xk ) G (α + 0.5, β + (μ − x)2 /2)

2.2.4 Noninformative Priors There is no compelling reason to choose conjugate priors as our priors, except for their simplicity, but the restrictive aspect of conjugate priors can be attenuated by using hyperpriors on the hyperparameters themselves, although we will not deal with this additional level of complexity in the current chapter. The core message is therefore that conjugate priors are nice to work with, but require a hyperparameter determination that may prove awkward in some settings and that may moreover have a lasting impact on the resulting inference. Instead of using conjugate priors, one can opt for a completely diﬀerent perspective and rely on so-called noninformative priors that aim at attenuating the impact of the prior on the resulting inference. These priors are fundamentally deﬁned as coherent extensions of the uniform distribution. Their purpose is to provide a reference measure that has as little as possible bearing on the inference (relative to the information brought by the likelihood). We ﬁrst warn the reader that, for unbounded parameter spaces, the densities of noninformative priors actually fail to integrate to a ﬁnite number and they are deﬁned instead as positive measures. While this sounds like an invalid extension of the probabilistic framework, it is quite correct to define the corresponding posterior distributions by (2.2), as long as the integral in the denominator is ﬁnite (almost surely). A more detailed account is for instance provided in Robert (2007, Sect. 1.5) about this possibility of using σ-ﬁnite measures (sometimes called improper priors) in settings where true probability prior distributions are too diﬃcult to come by or too subjective to be accepted by all. For instance, location models

36

2 Normal Models

x ∼ p(x − θ) are usually associated with ﬂat priors π(θ) = 1 (note that these models include the normal N (θ, 1) as a special case), while scale models 1 x x∼ f θ θ are usually associated with the log-transform of a ﬂat prior, that is, π(θ) = 1/θ . In a more general setting, the (noninformative) prior favored by most Bayesians is the so-called Jeﬀreys prior,3 which is related to the Fisher information matrix ∂ log f (X|θ) I F (θ) = varθ ∂θ by

1/2 π J (θ) = I F (θ) ,

where |I| denotes the determinant of the matrix I. Since the mean μ of a normal model is a location parameter, when the variance σ 2 is known, the standard choice of noninformative parameter is an arbitrary constant π(μ) (taken to be 1 by default). Given that this ﬂat prior formally corresponds to the limiting case τ = ∞ in the conjugate normal prior, it is easy to verify that this noninformative prior is associated with the posterior distribution N (x, 1), which happens to be the likelihood function in that case. An interesting consequence of this observation is that the MAP estimator is also the maximum likelihood estimator in that (special) case. For the general case when θ = (μ, σ 2 ), the Fisher information matrix leads to the Jeﬀreys prior π J (θ) = 1/σ 3 (Exercise 2.4). The corresponding posterior distribution on (μ, σ 2 ) is then π((μ, σ 2 )|Dn ) ∝ (σ −2 )(3+n)/2 exp − n(μ − x)2 + s2 /2σ 2 2 −s ∝ σ −1 exp −n(μ − x , ¯)2 /2σ 2 × (σ 2 )−(n+2)/2 exp 2σ 2 that is,

θ ∼ N x¯, σ 2 /n × I G n/2, s2 /2 .

a product of a conditional normal on μ by an inverse gamma on σ 2 . Therefore the marginal posterior distribution on μ is a t distribution (Exercise 2.1) ¯, s2 /n2 ) . μ|Dn ∼ T n, x 3 Harold Jeﬀreys was an English geophysicist who developed and formalized Bayesian methods in the 1930s in order to analyze geophysical data. He ended up writing an inﬂuential treatise on Bayesian statistics entitled Theory of Probability.

2.2

The Bayesian Toolkit

37

For normaldata, the diﬀerence in Fig. 2.3 between the noninformative solution and the conjugate posterior is minor, but it expresses that the prior distribution E (1) on σ −2 is not very appropriate for the Illingworth experiment, since it does not put enough prior weight on the region of importance, i.e. near 0.05. As a result, the most concentrated posterior is (seemingly paradoxically) the one associated with the noninformative prior!

A major

(and potentially dangerous) diﬀerence between proper and improper priors is that the posterior distribution associated with an improper prior is not necessarily deﬁned, that is, it may happen that (2.5) π(θ) (θ|Dn ) dθ < ∞

does not hold. In some cases, this diﬃculty disappears when the sample size is large enough. In others (see Chap. 6), it may remain whatever the sample size. But the main thing is that, when using improper priors, condition (2.5) must always be checked.

2.2.5 Bayesian Credible Intervals One point that must be clear from the beginning is that the Bayesian approach is a complete inferential approach. Therefore, it covers conﬁdence evaluation, testing, prediction, model checking, and point estimation. We will progressively cover the diﬀerent facets of Bayesian analysis in other chapters of this book, but we address here the issue of conﬁdence intervals because it is rather a straightforward step from point estimation. As with everything else, the derivation of the conﬁdence intervals (or conﬁdence regions in more general settings) is based on the posterior distribution π(θ|Dn ). Since the Bayesian approach processes θ as a random variable, a natural deﬁnition of a conﬁdence region on θ is to determine C(Dn ) such that π(θ ∈ C(Dn )|Dn ) = 1 − α

(2.6)

where α is a predetermined level such as 0.05.4 The important diﬀerence with a traditional perspective in (2.6) is that the integration is done over the parameter space, rather than over the observation space. The quantity 1 − α thus corresponds to the probability that a random θ belongs to this set C(Dn ), rather than to the probability that the random set contains the “true” value of θ. Given this drift in the interpretation of a 4

There is nothing special about 0.05 when compared with, say, 0.87 or 0.12. It is just that the famous 5 % level is accepted by most as an acceptable level of error. If the context of the analysis tells a diﬀerent story, another value for α (including one that may even depend on the data) should be chosen!

38

2 Normal Models

conﬁdence set (rather called a credible set by Bayesians), the determination of the best5 credible set turns out to be easier than in the classical sense: indeed, this set simply corresponds to the values of θ with the highest posterior values, C(Dn ) = {θ; π(θ|Dn ) ≥ kα } , where kα is determined by the coverage constraint (2.6). This region is called the highest posterior density (HPD) region. For normaldata, since the marginal posterior distribution on μ associated with the Jeﬀreys prior is the t distribution, T (n, x ¯, s2 /n2 ),

−(n+1)/2 π(μ|Dn ) ∝ n(μ − x ¯)2 + s2 with n = 64 degrees of freedom. Therefore, due to the symmetry properties of the t distribution, the 95 % credible interval on μ is centered at x¯ and its range is derived from the 0.975 quantile of the t distribution with n degrees of freedom, > qt(.975,df=n)*sqrt((n-1)*var(shift)/n^2) [1] 0.05082314 since the mnormt package does not compute quantiles. The resulting conﬁdence interval is therefore given by > qt(.975,df=n)*sqrt((n-1)*var(shift)/n^2)+mean(shift) [1] 0.03597939 > -qt(.975,df=n)*sqrt((n-1)*var(shift)/n^2)+mean(shift) [1] -0.06566689 i.e. equal to [−0.066, 0.036]. In conclusion, the value 0 belongs to this credible interval on μ and this (noninformative) Bayesian analysis of normaldata shows that, indeed, the absence of æther wind is not inﬁrmed by Illingworth’s experiment.

While

the shape of an optimal Bayesian conﬁdence set is easily derived, the computation of either the bound kα or the set C(Dn ) may be too challenging to allow an analytic construction outside conjugate setups (see Exercise 2.11).

2.3 Bayesian Model Choice Deciding the validity of some assumptions or restrictions on the parameter θ is a major part of the statistician’s job. In classical statistics, this type of 5

In the sense of producing the smallest possible volume with a given coverage.

2.3

Bayesian Model Choice

39

problems goes under the name of hypothesis testing, following the framework set by Fisher, Neyman and Pearson in the 1930s. Hypothesis testing considers a decision problem where an hypothesis is either true or false and where the answer provided by the statistician is also a statement whether or not the hypothesis is true. However, we deem this approach to be too formalized—even though it can be directly reproduced from a Bayesian perspective, as shown in Robert (2007, Chap. 5)—, we strongly favour a model choice philosophy, namely that two or more models are proposed in parallel and assessed in terms of their respective ﬁts of the data. This view acknowledges the fact that models are at best approximations of reality and it does not aim at ﬁnding a “true model”, as hypothesis testing may do. In this book, we will thus follow the later approach and take the stand that inference problems expressed as hypothesis testing by the classical statisticians are in fact comparisons of different models. In terms of numerical outcomes, both perspectives—Bayesian hypothesis testing vs. Bayesian model choice—are exchangeable but we already warn the reader that, while the Bayesian solution is formally very close to a likelihood (ratio) statistic, its numerical values often strongly diﬀer from the classical solutions. 2.3.1 The Model Index as a Parameter The essential novelty when dealing with the comparison of models is that this issue makes the model itself an unknown quantity of interest. Therefore, if we are comparing two or more models with indices k = 1, 2, . . . , J, we introduce a model indicator M taking values in {1, 2, . . . , J} and representing the index of the “true” model. If M = k, then the data Dn are generated from a statistical model Mk with likelihood (θk |Dn ) and parameter θk taking its value in a parameter space Θk . An obvious illustration is when opposing two standard parametric families, e.g., a normal family against a t family, in which case J = 2, Θ1 = R × R∗+ —for mean and variance—and Θ2 = R∗+ × R × R∗+ —for degree of freedom, mean and variance—, but this framework also includes soft or hard constraints on the parameters, as for instance imposing that a mean μ is positive. In this setting, a natural Bayes procedure associated with a prior distribution π is to consider the posterior probability δ π (Dn ) = Pπ (M = k|Dn ) , i.e., the posterior probability that the model index is k, and select the index of the model with the highest posterior probability as the model preferred by the data Dn . This representation implies that the prior π is deﬁned over the collection of model indices, {1, 2, . . . , J}, and, conditionally on the model index M, on the corresponding parameter space, Θk . This construction may sound both artiﬁcial and incomplete, as there is no prior on the parameter θk unless M = k, but it nonetheless perfectly translates the problem at hand:

40

2 Normal Models

inference on θk is meaningless unless this is the parameter of the correct model. Furthermore, the quantity of interest integrates out the parameter, since Pπ (M = k) (θk |Dn )πk (θk ) dθk π P (M = k|Dn ) = . J π j=1 P (M = j)πj (θj ) dθj

We believe it is worth emphasizing the above point: A parameter θk associated with a model does not have a statistical meaning outside this model. This means in particular that the notion of parameters “common to all models” often found in the literature, including the Bayesian literature, is not acceptable within a model choice perspective. Two models must have distinct parameters, if only because the purpose of the analysis is to end up with a single model.

The choice of the prior π is highly dependent on the value of the prior model probabilities Pπ (M = k). In some cases, there is experimental or subjective evidence about those probabilities, but in others, we are forced to settle for equal weights Pπ (M = k) = 1/J. For instance, given a single observation x ∼ N (μ, σ 2 ) from a normal model where σ 2 is known, assuming μ ∼ N (ξ, τ 2 ), the posterior distribution π(μ|x) is the normal distribution N (ξ(x), ω 2 ) with ξ(x) =

σ2 ξ + τ 2 x σ2 + τ 2

and

ω2 =

σ2 τ 2 . σ2 + τ 2

If the question of interest is to decide whether μ is negative or positive, we can directly compute μ − ξ(x) −ξ(x) π π < P (μ < 0|x) = P ω ω = Φ (−ξ(x)/ω) , (2.7) where Φ is the normal cdf. This computation does not seem to follow from the principles we just stated but it is only a matter of perspective as we can derive the priors on both models from the original prior. Deriving this posterior probability indeed means that, a priori, μ is negative with probability Pπ (μ < 0) = Φ(−ξ/τ ) and that, in this model, the prior on μ is the truncated normal exp{−(μ − ξ)2 /2τ 2 } √ π1 (μ) = Iμ0 . 2πτ Φ(ξ/τ )

The posterior probability of Pπ (M = k|Dn ) is the core object in Bayesian model choice and, as indicated above, the default procedure is to select the

2.3

Bayesian Model Choice

41

model with the highest posterior probability. However, in decisional settings where the choice between two models has diﬀerent consequences depending on the value of k, the boundary in Pπ (M = k|Dn ) between choosing one model and the other may be far from 0.5. For instance, in a pharmaceutical trial, deciding to start production of a new drug does not have the same ﬁnancial impact as deciding to run more preliminary tests. Changing the bound away from 0.5 is in fact equivalent to changing the prior probabilities of both models. 2.3.2 The Bayes Factor A notion central to Bayesian model choice is the Bayes factor π (Dn ) = B21

Pπ (M = 2|Dn )/Pπ (M = 1|Dn ) , Pπ (M = 2)/Pπ (M = 1)

which corresponds to the classical odds or likelihood ratio, the diﬀerence being that the parameters are integrated rather than maximized under each model. While this quantity is a simple one-to-one transform of the posterior probability, it can be used for Bayesian model choice without ﬁrst resorting to a determination of the prior weights of both models. Obviously, the Bayes factor depends on prior information through the choice of the model priors π1 and π2 , 2 (θ2 |Dn )π2 (θ2 ) dθ2 m2 (Dn ) π , = B21 (Dn ) = Θ2 m1 (Dn ) (θ |D )π (θ ) dθ 1 1 n 1 1 1 Θ1 and thus it can clearly be perceived as a Bayesian likelihood ratio which replaces the likelihoods with the marginals under both models. The evidence brought by the data Dn can be calibrated using for instance Jeﬀreys’ scale of evidence: π ) is between 0 and 0.5, the evidence against model M1 – if log21 (B21 is weak, – if it is between 0.5 and 1, it is substantial, – if it is between 1 and 2, it is strong, and – if it is above 2, it is decisive.

While this scale is purely arbitrary, it provides a reference for model assessment in a generic setting. Consider now the special case when we want to assess whether or not a speciﬁc value of one of the parameters is appropriate, for instance μ = 0 in the normaldata example. While the classical literature presents this problem as a point null hypothesis, we simply interpret it as the comparison of two models, N (0, σ 2 ) and N (μ, σ 2 ), for Illingworth’s data. In a more general framework, when the sample Dn is distributed as Dn ∼ f (Dn |θ), if we decompose θ as θ = (δ, ω) and if the restricted model corresponds to the ﬁxed value δ = δ0 , we deﬁne π1 (ω) as the prior under the restricted model (labelled M1 ) and π2 (θ)

42

2 Normal Models

as the prior under the unrestricted model (labelled M2 ). The corresponding Bayes factor is then (θ|Dn )π2 (θ) dθ π B21 (Dn ) = Θ ((δ0 , ω)|Dn )π1 (ω) dω Ω Note that, as hypotheses, point null problems often are criticized as artiﬁcial and impossible to test (in the sense of how often can one distinguish θ = 0 from θ = 0.0001?! ), but, from a model choice perspective, they simply correspond to more parsimonious models whose ﬁt to the data can be checked against the ﬁt produced by an unconstrained model. While the unconstrained model obviously contains values that produce a better ﬁt, averaging over the whole parameter space Θ may still result in a small integrated likelihood m2 (Dn ). The Bayes factor thus contains an automated penalization for complexity, a feature missed by the classical likelihood ratio statistic.

In

the very special case when the whole parameter is constrained to a ﬁxed value, θ = θ0 , the marginal likelihood under model M1 coincides with the likelihood (θ0 |Dn ) = f (Dn |θ0 ) and the Bayes factor simpliﬁes in f (Dn |θ)π2 (θ) dθ π B21 . (Dn ) = Θ f (Dn |θ0 )

For x ∼ N (μ, σ 2 ) and σ 2 known, consider assessing μ = 0 when μ ∼ N (0, τ 2 ) under the alternative model (labelled M2 ). The Bayes factor is the ratio π (Dn ) = B21

m2 (x) f (x|(0, σ 2 )) 2

2

2

σ e−x /2(σ +τ ) = √ 2 2 σ 2 + τ 2 e−x /2σ τ 2 x2 σ2 = exp . σ2 + τ 2 2σ 2 (σ 2 + τ 2 ) Table 2.2 gives a sample of the values of the Bayes factor when the normalized quantity x/σ varies. They obviously depend on the choice of the prior variance τ 2 and the dependence is actually quite severe, as we will see below with the Jeﬀreys–Lindley paradox. For normaldata, since we saw that setting σ to the Michelson–Morley value of 0.75 was producing a poor outcome compared with the noninformative solution, the comparison between the constrained and the unconstrained models is not very trustworthy, but as an illustration, it gives the following values:

2.3

Bayesian Model Choice

43

Table 2.2. Bayes factor B21 (z) against the null hypothesis μ = 0 for diﬀerent values of z = x/σ and τ z τ 2 = σ2 τ 2 = 10σ 2

0 0.707 0.302

0.68 0.794 0.372

1.28 1.065 0.635

1.96 1.847 1.728

> BaFa=function(z,rat){ #rat denotes the ratio tau^2/sigma^2 sqrt(1/(1+rat))*exp(z^2/(2*(1+1/rat)))} > BaFa(mean(shift),1) [1] 0.7071767 > BaFa(mean(shift),10) [1] 0.3015650 which supports the constraint μ = 0 for those two values of τ , since the Bayes factor is less than 1. (For this dataset, the Bayes factor is always less than one, see Exercise 2.13.) 2.3.3 The Ban on Improper Priors We introduced noninformative priors in Sect. 2.2.4 as a way to handle situations when the prior information was not suﬃcient to build proper priors. We also saw that, for normaldata, a noninformative prior was able to exhibit conﬂicts between the prior information (based on the Michelson–Morley experiment) and the data (resulting from Illingworth’s experiment). Unfortunately, the use of noninformative priors is very much restricted in model choice settings because the fact that they usually are improper leads to the impossibility of comparing the resulting marginal likelihoods. Looking at the expression of the Bayes factor, 2 (θ2 |Dn )π2 (θ2 ) dθ2 π , B21 (Dn ) = Θ2 Θ1 1 (θ1 |Dn )π1 (θ1 ) dθ1 it is clear that, when either π1 or π2 are improper, it is impossible to normalize the improper measures in a unique manner. Therefore, the Bayes factor becomes completely arbitrary since it can be multiplied by one or two arbitrary constants. For instance, when comparing x ∼ N (μ, 1) (model M1 ) with x ∼ N (0, 1) (model M2 ), the improper Jeﬀreys prior on model M1 is π1 (μ) = 1. The Bayes factor corresponding to this choice is π B12 (x) = +∞ −∞

e−x

2

/2

2

e−x /2 . = √ 2 2π e−(x−θ) /2 dθ

44

2 Normal Models

If, instead, we use the prior π1 (μ) = 100, the Bayes factor becomes π B12 (x)

e−x

2

/2

2

e−x /2 √ = = +∞ 2 100 2π 100 −∞ e−(x−θ) /2 dθ

and is thus one-hundredth of the previous value! Since there is no mathematical way to discriminate between π1 (μ) = 1 and π1 (μ) = 100, the answer clearly is non-sensical. Note that, if we are instead comparing model M1 where μ ≤ 0 and model M2 where μ > 0, then the posterior probability of model M1 under the ﬂat prior is 0 2 1 π e−(x−θ) /2 dθ = Φ(−x) , P (μ ≤ 0|x) = √ 2π −∞ which is uniquely deﬁned. The diﬃculty in using an improper prior also relates to what is called the Jeﬀreys–Lindley paradox, a phenomenon that shows that limiting arguments are not valid in testing settings. In contrast with estimation settings, the noninformative prior no longer corresponds to the limit of conjugate inferences. For instance, for the comparison of the normal x ∼ N (μ, σ 2 ) (model M1 ) and of the normal x ∼ N (μ, σ 2 ) (model M2 ) models when σ 2 is known, using a conjugate prior μ ∼ N (0, τ 2 ), the Bayes factor τ 2 x2 σ2 π exp B21 (x) = σ2 + τ 2 2σ 2 (σ 2 + τ 2 ) converges to 0 when τ goes to +∞, for every value of x, again a non-sensical procedure. Since improper priors are an essential part of the Bayesian approach, there are many proposals found in the literature to overcome this ban. Most of those proposals rely on a device that transforms the improper prior into a proper probability distribution by exploiting a fraction of the data Dn and then restricts itself to the remaining part of the data to run the test as in a standard situation. The variety of available solutions is due to the many possibilities of removing the dependence on the choice of the portion of the data used in the ﬁrst step. The resulting procedures are called pseudo-Bayes factors, although some may actually correspond to true Bayes factors. See Robert (2007, Chap. 5) for more details, although we do not advocate using those procedures. There is a major exception to this ban on improper priors that we can exploit. If both models under comparison have parameters that have similar enough meanings to share the same prior distribution, as for instance a measurement error σ 2 , then the normalization issue vanishes. Note that we are not assuming that parameters are common to both models and thus that we do not contradict the earlier warning about diﬀerent parameters to diﬀerent models. An illustration is provided by the above remark on the comparison

2.3

Bayesian Model Choice

45

of μ < 0 with μ > 0. This partial opening in the use of improper priors represents an opportunity but it does not apply to parameters of interest, i.e. to parameters on which restrictions are assessed. Example 2.1. When comparing two id normal samples, (x1 , . . . , xn ) and (y1 , . . . , yn ), with respective distributions N (μx , σ 2 ) and N (μy , σ 2 ), we can examine whether or not the two means are identical, i.e. μx = μy (corresponding to model M1 ). To take advantage of the structure of this model, we can assume that σ 2 is a measurement error with a similar meaning under both models and thus that the same prior πσ (σ 2 ) can be used under both models. This means that the Bayes factor 2 (μx , μy , σ|Dn )π(μx , μy )πσ (σ 2 ) dσ 2 dμx dμy π B21 (Dn ) = 1 (μ, σ|Dn )πμ (μ)πσ (σ 2 ) dσ 2 dμ does not depend on the normalizing constant used for πσ (σ 2 ) and thus that we can still use an improper prior such as πσ (σ 2 ) = 1/σ 2 in that case. Furthermore, we can rewrite μx and μy as μx = μ−ξ and μy = μ+ξ, respectively, and use a prior of the form π(μ, ξ) = πμ (μ)πξ (ξ) on the new parameterization so that, again, the same prior πμ can be used under both models. The same cancellation of the normalizing constant occurs for πμ , which means a Jeﬀreys prior πμ (μ) = 1 can be used. However, we need a proper and well-deﬁned prior on ξ, for instance ξ ∼ N (0, τ 2 ), which leads to √ π B21 (Dn )

e−n[(μ−ξ−¯x)

=

2

+(μ+ξ−¯ y )2 +s2xy ]/2σ2

e−n[(μ−¯x)

2

σ −2n−2 e−ξ

+(μ−¯ y )2 +s2xy ]/2σ2

(μ − ξ − x ¯)2 + (μ + ξ − y¯)2 + s2xy

=

−n

/2τ 2

/τ 2π dσ 2 dμ dξ

σ −2n−2 dσ 2 dμ

e−ξ

(μ − x ¯)2 + (μ − y¯)2 + s2xy

2

2

/2τ 2

−n

√ /τ 2π dμ dξ ,

dμ

where s2xy denotes the average s2xy =

n n 1 1 (xi − x¯)2 + (yi − y¯)2 . n i=1 n i=1

While the denominator can be completely integrated out, the numerator canπ is thus necessary. (This issue is adnot. A numerical approximation to B21 dressed in Sect. 2.4.) We conclude this section by a full processing of the assessment of μ = 0 for the single sample normal problem. Comparing models M1 : N (0, σ 2 ) under the prior π1 (σ 2 ) = 1/σ 2 and M2 : N (μ, σ 2 ) under the prior made of π2 (σ 2 ) = 1/σ 2 and π2 (μ|σ 2 ) equal to the normal N (0, σ 2 ) density, the Bayes factor is

46

2 Normal Models

π B21 (Dn ) =

2 2 2 2 2 dμdσ 2 e−[n(¯x−μ) +s ]/2σ e−μ /2σ σ −n−1−2 √ 2π 2 2 2 e−[n¯x +s ]/2σ σ −n−2 dσ 2

2

e−(n+1)[μ−n¯x/(n+1)] e−[n¯x

=

n¯ x2 + s2 2

(n + 1)−1/2 e−[n¯x

2

2

/(n+1)+s2 ]/2σ2

−n/2

/(n+1)+s2 ]/2σ2

dμdσ 2 σ −n−3 √ 2π

Γ (n/2) σ −n−2 dσ 2

−n/2 n¯ x2 + s2 Γ (n/2) 2 2 −n/2 x /(n + 1) + s2 −1/2 n¯ (n + 1) Γ (n/2) 2 = 2 −n/2 n¯ x + s2 Γ (n/2) 2 n/2 n¯ x2 + s2 , = (n + 1)−1/2 n¯ x2 /(n + 1) + s2 =

taking once again advantage of the normalizing constant of the gamma distribution (see also ¯2 /s2 , √ Exercise 2.8). It therefore increases to inﬁnity with x starting from 1/ n + 1 when x¯ = 0. The value of this Bayes factor for Illingworth’s data is given by > ratio=n*mean(shift)^2/((n-1)*var(shift)) > ((1+ratio)/(1+ratio/(n+1)))^(n/2)/sqrt(n+1) [1] 0.1466004 which conﬁrms the assessment that the model with μ = 0 is to be preferred.

2.4 Monte Carlo Methods While, as seen in Sect. 2.3, the Bayes factor and the posterior probability are the only quantities used in the assessment of models (and hypotheses), the analytical derivation of those objects is not always possible, since they involve integrating the likelihood (θ|Dn ) both on the sets Θ1 and Θ2 , under the respective priors π1 and π2 . Fortunately, there exist special numerical techniques for the computation of Bayes factors, which are, mathematically speaking, simply ratios of integrals. We now detail the techniques used in the approximation of intractable integrals, but refer to Chen et al. (2000) and Robert and Casella (2004, 2009) for book-length presentations.

2.4

Monte Carlo Methods

47

2.4.1 An Approximation Based on Simulations The technique that is most commonly used for integral approximations in statistics is called the Monte Carlo method6 and relies on computer simulations of random variables to produce an approximation technique that converges with the number of simulations. Its justiﬁcation is thus the law of large numbers, that is, if x1 , . . . , xN are independent and distributed from g, then the empirical average ˆN = (h(x1 ) + . . . + h(xN ))/N I converges (almost surely) to the integral I= h(x)g(x) dx . We will not expand on the foundations of the random number generators in this book, except for an introduction to accept–reject methods in Chap. 5 because of their links with Markov chain Monte Carlo techniques (see, instead, Robert and Casella, 2004). The connections of utmost relevance here are (a) that softwares like R can produce pseudo-random series that are indistinguishable from truly random series with a given distribution, as illustrated in Table 1.1 and (b) that those software packages necessarily cover a limited collection of distributions. Therefore, other methods must be found for simulating distributions outside this collection, while relying on the distributions already available, ﬁrst and foremost the uniform U (0, 1) distribution. The implementation of the Monte Carlo method is straightforward, at least on a formal basis, with the following algorithmic representation: Algorithm 2.1 Basic Monte Carlo Method For i = 1, . . . , N , simulate xi ∼ g(x). Take ˆ N = (h(x1 ) + . . . + h(xN ))/N I to approximate I. as long as the (computer-generated) pseudo-random generation from g is feasible and the h(xi ) values are computable. When simulation from g is a problem because g is nonstandard and usual techniques such as accept–reject algorithms (see Chap. 5) are diﬃcult to devise, more advanced techniques such as Markov Chain Monte Carlo (MCMC) are required. We will introduce those 6

This method is named in reference to the central district of Monaco, where the famous Monte-Carlo casino lies.

48

2 Normal Models

in both next chapters. When the diﬃculty is with the intractability of the function h, the solution is often to use an integral representation of h and to expand the random variables xi in (xi , yi ), where yi is an auxiliary variable. The use of such representations will be detailed in Chap. 6. Example 2.2 (Continuation of Example 2.1). As computed in Examπ (Dn ) can be simpliﬁed into ple 2.1, the Bayes factor B21 √

−n −ξ2 /2τ 2 (μ − ξ − x ¯)2 + (μ + ξ − y¯)2 + s2xy e dμ dξ/τ 2π π B21 (Dn ) =

−n dμ (μ − x ¯)2 + (μ − y¯)2 + s2xy √

−n+1/2 −ξ2 /2τ 2 (2ξ + x ¯ − y¯)2 + 2 s2xy e dξ/τ 2π = ,

−n+1/2 (¯ x − y¯)2 + 2 s2xy and we are left with a single integral in the numerator that involves the normal N (0, τ 2 ) density and can thus be represented as an expectation against this distribution. This means that simulating a normal N (0, τ 2 ) sample of ξi ’s π (Dn ) with (i = 1, . . . , N ) and replacing B21

−n+1/2 N 1 ¯ − y¯)2 + 2 s2xy +2 i=1 (2ξi + x N π ˆ B21 (Dn ) =

−n+1/2 (¯ x − y¯)2 + 2 s2xy is an asymptotically valid approximation scheme.

In normaldata, if we compare the ﬁfth and the sixth sessions, both with n = 10 observations, we obtain > illing=as.matrix(normaldata) > xsam=illing[illing[,1]==5,2] > xbar=mean(xsam) [1] -0.041 > ysam=illing[illing[,1]==6,2] > ybar=mean(ysam) [1] -0.025 > Ssquar=9*(var(xsam)+var(ysam))/10 [1] 0.101474 Picking τ = 0.75 as earlier, we get the following approximation to the Bayes factor > Nsim=10^4 > tau=0.75 > xis=rnorm(Nsim,sd=tau) > BaFa=mean(((2*xis+xbar-ybar)^2+2*Ssquar)^(-8.5))/ + ((xbar-ybar)^2+2*Ssquar)^(-8.5) [1] 0.0763622

2.4

Monte Carlo Methods

49

π (D ) implies that ξ = 0, i.e. μ = μ is much more likely This value of B21 n x y for the data at hand than μx = μy . Note that, if we use τ = 0.1 instead, the approximated Bayes factor is 0.4985 which slightly reduces the argument in favor of μx = μy .

Obviously, this Monte Carlo estimate of I is not exact, but generating a suﬃciently large number of random variables can render this approximation error arbitrarily small in a suitable probabilistic sense. It is also possible to assess the size of this error for a given number of simulations. If |h(x)|2 g(x) dx < ∞ , √ ˆN − I] is also normally distributed, the central limit theorem shows that N [I ˆN , estiand this can be used to construct asymptotic conﬁdence regions for I mating the asymptotic variance from the simulation output. π For the approximation of B21 (Dn ) proposed above, its variability is illustrated in Fig. 2.4, based on 500 replications of the simulation of N = 1000 normal variables used in the approximation and obtained as follows

> > + > > + > +

xis=matrix(rnorm(500*10^3,sd=tau),nrow=500) BF=((2*xis+xbar-ybar)^2+2*Ssquar)^(-8.5)/ ((xbar-ybar)^2+2*Ssquar)^(-8.5) estims=apply(BF,1,mean) hist(estims,nclass=84,prob=T,col="wheat2", main="",xlab="Bayes Factor estimates") curve(dnorm(x,mean=mean(estims),sd=sd(estims)), col="steelblue2",add=TRUE)

As can be seen on this ﬁgure, the value of 0.076 reported in the previous Monte Carlo approximation is in the middle of the range of possible values. More in connection with the above point, the shape of the histogram is clearly compatible with the normal approximation, as shown by the ﬁtted normal density.

2.4.2 Importance Sampling An important feature of Example 2.2 is that, for the Monte Carlo approximaπ (Dn ), we exhibited a normal density within the integral and hence tion of B21 derived a representation of this integral as an expectation under this normal distribution. This seems like a very restrictive constraint in the approximation of integrals but this is only an apparent restriction in that we will now show that there is no need to simulate directly from the normal density and furthermore that there is no intrinsic density corresponding to a given integral, but rather an inﬁnity of densities!

2 Normal Models

40 0

20

Density

60

80

50

0.06

0.07

0.08

0.09

0.10

Bayes Factor estimates

Fig. 2.4. Dataset normaldata: Histogram of 500 realizations of the approximation B 21 (Dn ) based on N = 1000 simulations each and normal ﬁt of the sample

Indeed, an arbitrary integral I=

H(x) dx

can be represented in inﬁnitely many ways as an expectation, since, for an arbitrary probability density γ, we always have H(x) I= γ(x) dx , (2.8) γ(x) under the minimal condition that γ(x) > 0 when H(x). Therefore, the generation of a sample from γ can provide a converging approximation to E and the Monte Carlo method applies in a very wide generality. This method is called importance sampling when applied to an expectation under a density g, I = h(x)g(x) dx , H(x) = h(x)g(x) since the values xi simulated from γ are weighted by the importance weights g(xi )/γ(xi ) in the approximation

2.4

Monte Carlo Methods

51

N g(xi ) ˆN = 1 h(xi ) . I N i=1 γ(xi )

While the representation (2.8) holds for any density γ with a support larger than

ˆN can deteriorate the support of H, the performance of the empirical average I considerably when the ratio h(x)g(x)/γ(x) is not bounded as this raises the possibility for inﬁnite variance in the resulting estimator. When using importance ˆN . sampling, one must always take heed of a potentially inﬁnite variance of I

An additional incentive in using importance sampling is that this method does not require the density g (or γ) to be known completely. Those densities can be known only up to a normalizing constant, g(x) ∝ g˜(x) and γ(x) ∝ γ˜ (x), since the ratio n n h(xi )˜ g (xi )/˜ γ (xi ) g˜(xi )/˜ γ (xi ) i=1

i=1

also converges to I when n goes to inﬁnity and when the xi ’s are generated from γ. The equivalent of Algorithm 2.1 for importance sampling is as follows: Algorithm 2.2 Importance Sampling Method For i = 1, . . . , N , simulate xi ∼ γ(x); compute ωi = g˜(xi )/γ(xi ) . Take N N ˆN = I ωi h(xi ) ωi i=1

i=1

to approximate I. This algorithm is straightforward to implement. Since it oﬀers a degree of freedom in the selection of γ, simulation from a manageable distribution can be imposed, keeping in mind the constraint that γ should have ﬂatter tails than g. Unfortunately, as the dimension of x increases, diﬀerences between the target density g and the importance density γ have a larger and larger impact. Example 2.3. Consider almost the same setting as in Exercise 2.11: Dn = (x1 , . . . , xn ) is an iid sample from C (θ, 1) and the prior on θ is a ﬂat prior. We can use a normal importance function from a N (μ, σ 2 ) distribution to produce a sample θ1 , . . . , θN that approximates the Bayes estimator of θ, i.e. its posterior mean, by

52

2 Normal Models

!n N 2 2 −1 t=1 θt exp (θt − μ) /2 i=1 [1 + (xi − θt ) ] π ˆ . δ (Dn ) = N ! n 2 2 −1 t=1 exp {(θt − μ) /2} i=1 [1 + (xi − θt ) ] But this is a very poor estimation (see Exercise 2.17 for an analytic explanation) and it degrades considerably when μ increases. If we run an R simulation experiment producing a sample of estimates when μ increases, as follows, > > > > > > > > + + + + + + + + + > > + > >

Nobs=10 obs=rcauchy(Nobs) Nsim=250 Nmc=500 sampl=matrix(rnorm(Nsim*Nmc),nrow=1000) # normal samples raga=riga=matrix(0,nrow=50,ncol=2) # ranges mu=0 for (j in 1:50){ prod=1/dnorm(sampl-mu) # importance sampling for (i in 1:Nobs) prod=prod*dt(obs[i]-sampl,1) esti=apply(sampl*prod,2,sum)/apply(prod,2,sum) raga[j,]=range(esti) riga[j,]=c(quantile(esti,.025),quantile(esti,.975)) sampl=sampl+0.1 mu=mu+0.1 } mus=seq(0,4.9,by=0.1) plot(mus,0*mus,col="white",xlab=expression(mu), ylab=expression(hat(theta)),ylim=range(raga)) polygon(c(mus,rev(mus)),c(raga[,1],rev(raga[,2])),col="grey50") polygon(c(mus,rev(mus)),c(riga[,1],rev(riga[,2])),col="pink3")

as shown by Fig. 2.5, not only does the range of the approximation increase, but it ends up missing the true value θ = 0 when μ is far enough from 0.

2.4.3 Approximation of Bayes Factors Bayes factors being ratios of integrals, they can be approximated by regular importance sampling tools. However, given their speciﬁcity as ratios of marginal likelihoods, hence of normalizing constants of the posterior distributions, there exist more specialized techniques, including a fairly generic method called bridge sampling, developed by Gelman and Meng (1998). When comparing two models with sampling densities f1 (Dn |θ1 ) (model M1 ) and f2 (Dn |θ2 ) (model M2 ), assume that both models share the same parameter space Θ. This is for instance the case when comparing the ﬁt of two densities with the same number of parameters (modulo a potential reparameterization of one of the models). In this setting, if the corresponding prior

Monte Carlo Methods

53

1

2

2.4

−1

0

θ

((1+ratio)/(1+ratio/(n+1)))^(-n/2)*sqrt(n+1) [1] 6.821262 while the bridge sampling solution is obtained as > > > > > > > > > > > > > > + + + > + + > > > > > + + +

n=64 xbar=mean(shift) sqar=(n-1)*var(shift) Nmc=10^7 # Simulation from model M2: sigma2=1/rgamma(Nmc,shape=n/2,rate=(n*xbar^2/(n+1)+sqar)/2) mu2=rnorm(Nmc,n*xbar/(n+1),sd=sqrt(sigma2/(n+1))) # Simulation from model M1: sigma1=1/rgamma(Nmc,shape=n/2,rate=(n*xbar^2+sqar)/2) muhat=mean(mu2) tauat=sd(mu2) mu1=rnorm(Nmc,mean=muhat,sd=tauat) #tilde functions tildepi1=function(sigma,mu){ exp(-.5*((n*xbar^2+sqar)/sigma+(n+2)*log(sigma))+ dnorm(mu,muhat,tauat,log=T)) } tildepi2=function(sigma,mu){ exp(-.5*((n*(xbar-mu)^2+sqar+mu^2)/sigma+(n+3)*log(sigma)+ log(2*pi)))} #Bayes Factor loop K=diff=1 rationum=tildepi2(sigma1,mu1)/tildepi1(sigma1,mu1) ratioden=tildepi1(sigma2,mu2)/tildepi2(sigma2,mu2) while (diff>0.01*K){ BF=mean(1/(1+K*rationum))/mean(1/(K+ratioden)) diff=abs(K-BF) K=BF}

and returns the value > BF [1] 6.820955 which is deﬁnitely close to the true value! The second possible trick to overcome the dimension diﬃculty while using the bridge sampling strategy is to introduce artiﬁcial posterior distributions in each of the parameters spaces and to process each marginal likelihood as an integral ratio in itself. For instance, if η1 (θ1 ) is an arbitrary normalized density on θ1 , and α is an arbitrary function, we have

56

2 Normal Models

m1 (Dn ) =

π ˜1 (θ1 |Dn ) dθ1 =

π ˜1 (θ1 |Dn )α(θ1 )η1 (θ1 ) dθ1 η1 (θ1 )α(θ1 )π1 (θ1 |Dn ) dθ1

by application of (2.9). Therefore, the optimal choice of α leads to the approximation

N η η η ˜1 (θ1i |Dn ) {m1 (Dn )˜ π 1 (θ1i |Dn ) + η(θ1i )} i=1 π m ˆ 1 (Dn ) = N

π 1 (θ1i |Dn ) + η(θ1i )} i=1 η(θ1i ) {m1 (Dn )˜ η ∼ η(θ1 ). The choice of the density η is obviwhen θ1i ∼ π1 (θ1 |Dn ) and θ1i ously fundamental and it should be close to the true posterior π1 (θ1 |Dn ) to guarantee good convergence approximation. Using a normal approximation to the posterior distribution of θ or a non-parametric approximation based on a sample from π1 (θ1 |Dn ), or yet again an average of MCMC proposals (see Chap. 4) are reasonable choices.

The R implementation of this approach can be done as follows > > > > > > > > + > > > > > >

sigma1=1/rgamma(Nmc,shape=n/2,rate=(n*xbar^2+sqar)/2) sihat=mean(log(sigma1)) tahat=sd(log(sigma1)) sigma1b=exp(rnorm(Nmc,sihat,tahat)) #tilde function tildepi1=function(sigma){ exp(-.5*((n*xbar^2+sqar)/sigma+(n+2)*log(sigma)))} K=diff=1 rnum=dnorm(log(sigma1b),sihat,tahat)/ (sigma1b*tildepi1(sigma1b)) rden=sigma1*tildepi1(sigma1)/dnorm(log(sigma1),sihat,tahat) while (diff>0.01*K){ BF=mean(1/(1+K*rnum))/mean(1/(K+rden)) diff=abs(K-BF) K=BF} m1=BF

when using a normal distribution on log(σ 2 ) as an approximation to π1 (θ1 |Dn ). When considering the unconstrained model, a bivariate normal density can be used, as in > sigma2=1/rgamma(Nmc,shape=n/2,rate=(n*xbar^2/(n+1)+sqar)/2) > mu2=rnorm(Nmc,n*xbar/(n+1),sd=sqrt(sigma2/(n+1))) > temean=c(mean(mu2),mean(log(sigma2)))

2.4

> > > >

Monte Carlo Methods

57

tevar=cov.wt(cbind(mu2,log(sigma2)))$cov te2b=rmnorm(Nmc,mean=temean,tevar) mu2b=te2b[,1] sigma2b=exp(te2b[,2])

leading to > m1/m2 [1] 6.824417 The performances of both extensions are obviously highly dependent on the choice of the completion factors, η1 and η2 on the one hand and π1∗ on the other hand. The performances of the ﬁrst solution, which bridges both models via π1∗ , are bound to deteriorate as the dimension gap between those models increases. The impact of the dimension of the models is less keenly felt for the other solution, as the approximation remains local. As a simple illustration of the performances of both methods, we produce here a comparison between the completions based on a single pseudoconditional and on two local approximations to the posteriors, by running repeated approximations for normaldata and tracing the resulting boxplot as a measure of the variability of those methods. As shown in Fig. 2.6, the variability is quite comparable for both solutions in this speciﬁc case. Note that there exist many other approaches to the approximative computation of marginal likelihoods and of Bayes factors that we cannot cover here. We want however to point out the dangers of the harmonic mean approximation. This approach proceeds from the interesting identity ϕ1 (θ1 ) π1 (θ1 )1 (θ1 |Dn ) ϕ1 (θ1 ) π1 Dn = dθ1 E π1 (θ1 )1 (θ1 |Dn ) π1 (θ1 )1 (θ1 |Dn ) m1 (Dn ) 1 , = m1 (Dn ) which holds, no matter what the density ϕ1 (θ1 ) is—provided ϕ1 (θ1 ) = 0 when π1 (θ1 )1 (θ1 |Dn ) = 0—. The most common implementation in approximations of the marginal likelihood uses ϕ1 (θ1 ) = π1 (θ1 ), leading to the approximation m ˆ 1 (Dn ) = 1

N −1

N j=1

1 . 1 (θ1j |Dn )

While very tempting, since it allows for a direct processing of simulations from the posterior distribution, this approximation is unfortunately most often associated with an inﬁnite variance (Exercise 2.19) and, thus, should not be used. On the opposite, using ϕ1 ’s with supports constrained to the 25 % HPD

58

2 Normal Models

single

double

Fig. 2.6. Dataset normaldata: Boxplot of the variability of the approximations to the Bayes factor assessing whether or not μ = 0, based on a single and on a double completions. Each approximation is based on 105 simulations and the boxplots are based on 250 approximations. The dotted line corresponds to the true π (Dn ) value of B12

regions—approximated by the convex hull of the 10 % or of the 25 % highest simulations—is both completely appropriate and implementable (Marin and Robert, 2010).

2.5 Outlier Detection The above description of inference in normal models is only an introduction both to Bayesian inference and to normal structures. Needless to say, there exists a much wider range of possible applications. For instance, we will meet the normal model again in Chap. 4 as the original case of the (generalized) linear model. Before that, we conclude this chapter with a simple extension of interest, the detection of outliers. Since normal modeling is often an approximation to the “real thing,” there may be doubts about its adequacy. As already mentioned above, we will deal later with the problem of checking that the normal distribution is appropriate for the whole dataset. Here, we consider the somehow simpler problem of separately assessing whether or not each point in the dataset is compatible with

2.5

Outlier Detection

59

normality. There are many diﬀerent ways of dealing with this problem. We choose here to use the predictive distribution: If an observation xi is unlikely under the predictive distribution based on the other observations, then we can argue against its distribution being equal to the distribution of the other observations. If xn+1 is a future observation from the same distribution f (·|θ) as the sample Dn , its predictive distribution given the current sample is deﬁned as f π (xn+1 |Dn ) = f (xn+1 |θ, Dn )π(θ|Dn ) dθ = f (xn+1 |θ)π(θ|Dn ) dθ . This deﬁnition is coherent with the Bayesian approach, which considers xn+1 as an extra unknown and then integrates out θ if xn+1 is the “parameter” of interest. For the normal N (μ, σ 2 ) setup, using a conjugate prior on (μ, σ 2 ) of the form (σ 2 )−λσ −3/2 exp − λμ (μ − ξ)2 + α /2σ 2 , the corresponding posterior distribution on (μ, σ 2 ) given Dn is λμ ξ + nxn σ2 nλμ 2 2 , (x − ξ) /2 , N × I G λσ + n/2, α + s + λμ + n λμ + n λμ + n denoted by N ξ(Dn ), σ 2 /λμ (Dn ) × I G (λσ (Dn )/2, α(Dn )/2) , and the predictive on xn+1 is derived as f π (xn+1 |Dn ) ∝ (σ 2 )−λσ (Dn )/2−1−1 exp −(xn+1 − μ)2 /2σ 2 × exp − λμ (Dn )(μ − ξ(Dn ))2 + α(Dn ) /2σ 2 d(μ, σ 2 ) ∝

(σ 2 )−λσ (Dn )/2−3/2 exp − (λμ (Dn ) + 1)(xn+1 − ξ(Dn ))2

λμ (Dn ) + α(Dn ) /2σ 2 dσ 2

−(λσ (Dn )+1)/2 λμ (Dn ) + 1 ∝ α(Dn ) + . (xn+1 − ξ(Dn ))2 λμ (Dn ) Therefore, the predictive of xn+1 given the sample Dn is a Student t distribution with mean ξ(Dn ) and λσ (Dn ) degrees of freedom. In the special case of the noninformative prior, corresponding to the limiting values λμ = λσ = α = 0, the predictive is −(n+1)/2 n+1 f π (xn+1 |Dn ) ∝ s2 + 1(xn+1 − xn )2 . (2.11) n

60

2 Normal Models

It is therefore a Student’s t distribution with n degrees of freedom, a mean equal to xn and a scale factor equal to (n − 1)s2 /n, which is equivalent to a variance equal to (n − 1)s2 /n2 (to compare with the maximum likelihood estimator σ ˆn2 = s2 /n). In the outlier problem, we process each observation xi ∈ Dn as if it was a “future” observation. Namely, we consider fiπ (x|Dni ) as being the predictive distribution based on Dni = (x1 , . . . , xi−1 , xi+1 , . . . , xn ). Considering fiπ (xi |Dni ) or the corresponding cdf Fiπ (xi |Dni ) (in dimension one) gives an indication of the level of compatibility of the observation xi with the sample. To quantify this level, we can, for instance, approximate the distribution of Fiπ (xi |Dni ) as being uniform over [0, 1] since Fiπ (·|Dni ) does converge to the true cdf of the model. Simultaneously checking all Fiπ (xi |Dni ) over i may signal outliers.

The detection of outliers must pay attention to the Bonferroni fallacy, which is that extreme values do occur in large enough samples. This means that, as n increases, we will see smaller and smaller values of Fiπ (xi |Dni ) occurring, and this even when the whole sample is from the same distribution. The signiﬁcance level must therefore be chosen in accordance with this observation, for instance using a bound a on Fiπ (xi |Dni ) such that 1 − (1 − a)n = 1 − α , where α is the nominal level chosen for outlier detection.

Considering normaldata, we can compute the predictive cdf for each of the 64 observations, considering the 63 remaining ones as data. > n=length(shift) > outl=rep(0,n) > for (i in 1:n){ + lomean=-mean(shift[-i]) + losd=sd(shift[-i])*sqrt((n-2)/n) + outl[i]=pt((shift[i]-lomean)/losd,df=n-1) + } Figure 2.7 provides the qq-plot of the Fiπ (xi |Dni )’s against the uniform quantiles and compares it with a qq-plot based on a dataset truly simulated from the uniform U (0, 1). > + > >

plot(c(0,1),c(0,1),lwd=2,ylab="Predictive",xlab="Uniform", type="l") points((1:n)/(n+1),sort(outl),pch=19,col="steelblue3") points((1:n)/(n+1),sort(runif(n)),pch=19,col="tomato")

There is no clear departure from uniformity when looking at this graph, except of course for the multiple values found in normaldata.

Exercises

61

0.6 0.4 0.0

0.2

Predictive

0.8

1.0

2.6

0.0

0.2

0.4

0.6

0.8

1.0

Uniform

Fig. 2.7. Dataset normaldata: qq-plot of the sample of the Fiπ (xi |Dni ) for a uniform U (0, 1) distribution (blue dots) and comparison with a qq-plot for a uniform U (0, 1) sample (red dots)

2.6 Exercises 2.1 Show that, if μ|σ 2 ∼ N (ξ, σ 2 /λμ ) ,

σ 2 ∼ I G (λσ /2, α/2) ,

then μ ∼ T (λσ , ξ, α/λμ λσ ) a t distribution with λσ degrees of freedom, location parameter ξ and scale parameter α/λμ λσ . 2.2 Show that, if σ 2 ∼ I G (α, β), then E[σ 2 ] = β/(α − 1). Derive from the density of I G (α, β) that the mode is located in β/(α + 1). ˆ 2 |Dn ] proˆ n )) the posterior expectation Eπ [||θ − θ|| 2.3 Show that minimizing (in θ(D ˆ duces the posterior expectation as the solution in θ. 2.4 Show that the Fisher information on θ = (μ, σ 2 ) for the normal N (μ, σ 2 ) distribution is given by

62

2 Normal Models 1/σ 2 1/σ 2 0 2(x − μ)/2σ 4 I F (θ) = Eθ = 4 2 6 4 4 2(x − μ)/2σ (μ − x) /σ − 1/2σ 0 1/2σ

and deduce that Jeﬀreys’ prior is π J (θ) ∝ 1/σ 3 . 2.5 Derive each line of Table 2.1 by an application of Bayes’ formula, π(θ|x) ∝ π(θ)f (x|θ), and the identiﬁcation of the standard distributions. 2.6 A Weibull distribution W (α, β, γ) is deﬁned as the power transform of a gamma G (α, β) distribution: If x ∼ W (α, β, γ), then xγ ∼ G (α, β). Show that, when γ is known, W (α, β, γ) allows for a conjugate family, but that it does not an exponential family when γ is unknown. 2.7 Show that, when the prior on θ = (μ, σ 2 ) is N (ξ, σ 2 /λμ ) × I G (λσ , α), the marginal prior on μ is a Student t distribution T (2λσ , ξ, α/λμ λσ ) (see Example 2.18 for the deﬁnition of a Student t density). Give the corresponding marginal prior on σ 2 . For an iid sample Dn = (x1 , . . . , xn ) from N (μ, σ 2 ), derive the parameters of the posterior distribution of (μ, σ 2 ). 2.8 Show that the normalizing constant for a Student T (ν, μ, σ 2 ) distribution is Γ ((ν + 1)/2)/Γ (ν/2) √ . σ νπ Deduce that the density of the Student t distribution T (ν, θ, σ 2 ) is fν (x) =

Γ ((ν + 1)/2) √ σ νπ Γ (ν/2)

−(ν+1)/2 (x − θ)2 1+ . νσ 2

2.9 Show that, for location and scale models, the speciﬁc noninformative priors are special cases of Jeﬀreys’ generic prior, i.e., that π J (θ) = 1 and π J (θ) = 1/θ, respectively. 2.10 Show that, when π(θ) is a probability density, (2.5) necessarily holds for all datasets Dn . 2.11 Consider a dataset Dn from the Cauchy distribution, C (μ, 1). 1. Show that the likelihood function is

(μ|Dn ) =

n i=1

fμ (xi ) =

πn

n

i=1 (1

1 . + (xi − μ)2 )

2. Examine whether or not there is a conjugate prior for this problem. (The answer is no.) 3. Introducing a normal prior on μ, say N (0, 10), show that the posterior distribution is proportional to exp(−μ2 /20) . π ˜ (μ|Dn ) = n 2 i=1 (1 + (xi − μ) ) 4. Propose a numerical solution for solving π ˜ (μ|Dn ) = k. (Hint: A simple trapezoidal integration can be used: based on a discretization size Δ, computing π ˜ (μ|Dn ) on a regular grid of width Δ and summing up.)

2.6

Exercises

63

2.12 Show that the limit of the posterior probability Pπ (μ < 0|x) of (2.7) when τ goes to ∞ is Φ(−x/σ). Show that, when ξ varies in R, the posterior probability can take any value between 0 and 1. 2.13 Deﬁne a function BaRaJ of the ratio rat when z=mean(shift)/.75 in the function BaFa. Deduce from a plot of the function BaRaJ that the Bayes factor is always less than one when rat varies. (Note: It is possible to establish analytically that the Bayes factor is maximal and equal to 1 for τ = 0.) 2.14 In the application part of Example 2.1 to normaldata, plot the approximated Bayes factor as a function of τ . (Hint: Simulate a single normal N (0, 1) sample and recycle it for all values of τ .) 2.15 In the setup of Example 2.1, show that, when ξ ∼ N (0, σ 2 ), the Bayes factor can be expressed in closed form using the normalizing constant of the t distribution (see Exercise 2.8) 2.16 Discuss what happens to the importance sampling approximation when the support of g is larger than the support of γ. 2.17 Show that, when γ is the normal N (0, ν/(ν − 2)) density and fν is the density of the t distribution with ν degrees of freedom, the ratio 2

ex (ν−2)/2ν fν2 (x) ∝ γ(x) [1 + x2 /ν](ν+1) does not have a ﬁnite integral. What does this imply about the variance of the importance weights? Deduce that the importance weights of Example 2.3 have inﬁnite variance. 2.18 If fν denotes the density of the Student t distribution T (ν, 0, 1) (see Exercise 2.8), consider the integral x I= 1 − x fν (x) dx . 1. Show that I is ﬁnite but that

|x| fν (x) dx = ∞ . |1 − x|

2. Discuss the respective merits of the following importance functions γ – the density of the Student T (ν, 0, 1) distribution, – the density of the Cauchy C (0, 1) distribution, – the density of the normal N (0, ν/(ν − 2)) distribution. In particular, show via an R simulation experiment that these diﬀerent choices all lead to unreliable estimates of I and deduce that the three corresponding estimators have inﬁnite variance. 3. Discuss the alternative choice of a gamma distribution folded at 1, that is, the distribution of x symmetric around 1 and such that |x − 1| ∼ Ga(α, 1) .

64

2 Normal Models Show that

√ f 2 (x) ∝ x fν2 (x) |1 − x|1−α−1 exp |1 − x| γ(x) is integrable around x = 1 when α < 1 but not at inﬁnity. Run a simulation experiment to evaluate the performances of this new proposal. h(x)

2.19 Evaluate the harmonic mean approximation N

m ˆ 1 (Dn ) = 1 N −1 j=1 2

1 .

1 (θ1j |Dn )

when applied to the N (0, σ ) model, normaldata, and an I G (1, 1) prior on σ 2 .

3 Regression and Variable Selection

You see, I always keep my sums. —Ian Rankin, Strip Jack.—

Roadmap Linear regression is one of the most widely used tools in statistics for analyzing the (linear) inﬂuence of some variables or some factors on others and thus to uncover explanatory and predictive patterns. This chapter details the Bayesian analysis of the linear (or regression) model both in terms of prior speciﬁcation (Zellner’s G-prior) and in terms of variable selection, the next chapter appearing as a sequel for nonlinear dependence structures. The reader should be warned that, given that these models are the only conditional models where explicit computation can be conducted, this chapter contains a fair amount of matrix calculus. The photograph at the top of this page is a picture of processionary caterpillars, in connection (for once!) with the benchmark dataset used in this chapter.

J.-M. Marin and C.P. Robert, Bayesian Essentials with R, Springer Texts in Statistics, DOI 10.1007/978-1-4614-8687-9 3, © Springer Science+Business Media New York 2014

65

66

3 Regression and Variable Selection

3.1 Linear Models A large proportion of statistical analyses deal with the representation of dependences among several observed quantities. For instance, which social factors inﬂuence unemployment duration and the probability of ﬁnding a new job? Which economic indicators are best related to recession occurrences? Which physiological levels are most strongly correlated with aneurysm strokes? From a statistical point of view, the ultimate goal of these analyses is thus to ﬁnd a proper representation of the conditional distribution, f (y|θ, x), of an observable variable y given a vector of observables x, based on a sample of x and y. While the overall estimation of the conditional density f is usually beyond our ability, the estimation of θ and possibly of restricted features of f is possible within the Bayesian framework, as shown in this chapter. The variable of primary interest, y, is called the response or the outcome variable; we assume here that this variable is continuous, but we will completely relax this assumption in the next chapter. The variables x = (x1 , . . . , xp ) are called explanatory variables and may be discrete, continuous, or both. One sometimes picks a single variable xj to be of primary interest. We then call it the treatment variable, labeling the other components of x as control variables, meaning that we want to address the (linear) inﬂuence of xj on y once the linear inﬂuence of all the other variables has been taken into account (as in medical studies). The distribution of y given x is typically studied in the context of a set of units or experimental subjects, i = 1, . . . , n, such as patients in a hospital ward, on which both yi and xi1 , . . . , xip are measured. The dataset is then made up of the reunion of the vector of outcomes y = (y1 , . . . , yn ) and the n × p matrix of explanatory variables ⎡ x11 x12 ⎢ x21 x22 ⎢ ⎢ X = [x1 . . . xp ] = ⎢ x31 x32 ⎢ .. .. ⎣ . .

... ... ... .. .

⎤ x1p x2p ⎥ ⎥ x3p ⎥ ⎥. .. ⎥ . ⎦

xn1 xn2 . . . xnp The caterpillar dataset exploited in this chapter was extracted from a 1973 study on pine processionary1caterpillars: it assesses the inﬂuence of some forest settlement characteristics on the development of caterpillar colonies. This dataset was ﬁrst published and studied in Tomassone et al. (1993). The response variable is the logarithmic transform of the average number of nests of caterpillars per tree (as the one in the picture at the beginning of this chapter) in an area of 500 m2 (which corresponds to the last column in caterpillar). There are p = 8 potential explanatory variables deﬁned on n = 33 areas, as follows 1

These caterpillars derive their name from their habit of moving over the ground in incredibly long head-to-tail monk-like processions when leaving their nest to create a new colony.

3.1

x1 x2 x3 x4 x5 x6 x7 x8

is is is is is is is is

the the the the the the the the

Linear Models

67

altitude (in meters), slope (in degrees), number of pine trees in the area, height (in meters) of the tree sampled at the center of the area, orientation of the area (from 1 if southbound to 2 otherwise), height (in meters) of the dominant tree, number of vegetation strata, mix settlement index (from 1 if not mixed to 2 if mixed).

The goal of the regression analysis is to decide which explanatory variables have a strong inﬂuence on the number of nests and how these inﬂuences overlap with one another. As shown by Fig. 3.1, some of these variables clearly have a restricting inﬂuence on the number of nests, as for instance with x5 , x7 and x8 . We use the following R code to produce Fig. 3.1 (the way we created the objects y and X will be described later). > par(mfrow=c(2,4),mar=c(4.2,2,2,1.2)) > for (j in 1:8) plot(X[,j],y,xlab=vnames[j],pch=19, + col="sienna4",xaxt="n",yaxt="n") While many models and thus many dependence structures can be proposed for dependent datasets like caterpillar, in this chapter we only focus on the Gaussian linear regression model, namely the case when E[y|x, θ] is linear in x and the noise is normal. The ordinary normal linear regression model is such that, using a matrix representation, y|α, β, σ 2 ∼ Nn α1n + Xβ, σ 2 In , where Nn denotes the normal distribution in dimension n, and thus the yi ’s are independent normal random variables with E[yi |α, β, σ 2 ] = α + β1 xi1 + . . . + βp xip ,

V[yi |α, β, σ 2 ] = σ 2 .

Given that the models studied in this chapter are all conditional on the regressors, we omit the conditioning on X to simplify the notations. For caterpillar, where n = 33 and p = 8, we thus assume that the expected lognumber yi of caterpillar nests per tree over an area is modeled as a linear combination of an intercept and eight predictor variables (i = 1, . . . , n), E[yi |α, β, σ 2 ] = α +

8 j=1

βj xij ,

68

3 Regression and Variable Selection

x1

x2

x3

x4

x5

x6

x7

x8

Fig. 3.1. Dataset caterpillar: Plot of the pairs (xj , y) (1 ≤ j ≤ 8)

while the variation around this expectation is supposed to be normally distributed. Note that it is also customary to assume that the yi ’s are conditionally independent. The caterpillar dataset is called by the command data(caterpillar) and is made of the following rows: 1200 1342 .... 1229 1310

22 1 4 1.1 5.9 1.4 1.4 2.37 28 8 4.4 1.5 6.4 1.7 1.7 1.47 21 11 5.8 1.8 10 2.3 2 0.21 36 17 5.2 1.9 10.3 2.6 2 0.03

The ﬁrst eight columns correspond to the explanatory variables and the last column is the response variable, i.e. the lognumber of caterpillar nests. The following R code is an example for starting with this caterpillar dataset:

3.2

Classical Least Squares Estimator

69

> y=log(caterpillar$y) > X=as.matrix(caterpillar[,1:8]) There is a diﬀerence between using ﬁnite-valued regressors like x7 in caterpillar and using categorical variables (or factors), which also take a ﬁnite number of values but whose range has no numerical meaning. For instance, if x denotes the socio-professional category of an employee, this variable may range from 1 to 9 for a rough grid of socio-professional activities, or it may range from 1 to 89 on a ﬁner grid, and the numerical values are not comparable. It thus makes little sense to involve x directly in the regression, and the usual approach is to replace the single regressor x (taking values in {1, . . . , m}, say) with m indicator (or dummy) variables x1 = I1 (x), . . ., xm = Im (x). In essence, a diﬀerent constant (or intercept) βj is used in the regression for each class of categorical variable: it is invoked in the linear regression under the form . . . + β1 I1 (x) + . . . + βm Im (x) + . . . . Note that there is an identiﬁability issue related with this model since the sum of the indicators is always equal to one. In a Bayesian perspective, identiﬁability can be achieved via the prior distribution. However, we can also impose an identiﬁability constraint on the parameters, for instance by omitting one class (such as β1 = 0). We pursue this direction further in Sects. 4.5.1 and 6.2.

3.2 Classical Least Squares Estimator Before fully launching into the description of the Bayesian approach to the linear model, we recall the basics of the classical processing of this model (in particular, to relate the Bayesian perspective to the results provided by standard software such as R lm output). For instance, the parameter β can obviously be estimated via maximum likelihood estimation. In order to avoid non-identiﬁability and uniqueness problems, we assume that [1n X] is of full rank, that is, rank [1n X] = p+1. This also means that there is no redundant structure among the explanatory variables.2 We suppose in addition that p + 1 < n in order to obtain well-deﬁned estimates for all parameters. Notice that, since the inferential process is conditioned on the design matrix X, we choose to standardize the data, namely to center and to scale the columns of X so that the estimated values of β are truly comparable. For this purpose, we use the R function scale: > X=scale(X)

2

Hence, the exclusion of one of the classes for categorical variables.

70

3 Regression and Variable Selection

The likelihood (α, β, σ 2 |y) of the standard normal linear model is provided by the following matrix representation: 1 1 T exp − 2 (y − α1n − Xβ) (y − α1n − Xβ) . (3.1) n/2 2σ (2πσ 2 ) The maximum likelihood estimators of α and β are then the solution of the (least squares) minimization problem T

min (y − α1n − Xβ) (y − α1n − Xβ) α,β

= min α,β

n

2

(yi − α − β1 xi1 − . . . − βp xip ) ,

i=1 n

1 yi the empirical mean of the yi ’s and recall that, n i=1

¯ = If we denote by y

T 1T n X = 0n because of the standardization step, we have a Pythagorean decomposition of the above norm as T

(y−α1n −Xβ) (y−α1n −Xβ) y−α)1n )T (y−¯ y1n −Xβ+(¯ y−α)1n ) = (y−¯ y1n −Xβ+(¯ T

= (y−¯ y1n −Xβ) (y−¯ y1n −Xβ) +2(¯ y−α)1T y1n −Xβ) +n(¯ y−α)2 n (y−¯ = (y−¯ y1n −Xβ)T (y−¯ y1n −Xβ) +n(¯ y−α)2 . ¯ 1n − Xβ) = (n¯ y − n¯ y) = 0. Therefore, the likelihood Indeed, 1T n (y − y (α, β, σ 2 |y) is given by ) ( n 1 1 T 2 . exp − (y−¯ y 1 −Xβ) (y−¯ y 1 −Xβ) exp − (¯ y −α) n n n/2 2σ 2 2σ 2 (2πσ 2 ) We get from the above decomposition that ¯, α ˆ=y

ˆ = (XT X)−1 XT (y − y¯) . β

ˆ is the orthogonal projection of y on the linear In geometrical terms, (ˆ α, β) subspace spanned by the columns of [1n X]. It is quite simple to check that ˆ is an unbiased estimator of (α, β). In fact, the Gauss–Markov theorem (ˆ α, β) ˆ is the best linear unbiased (see, e.g., Christensen, 2002) states that (ˆ α, β) estimator of (α, β). This means that, for all a ∈ Rp+1 , and with the abuse of ˆ represents a column vector, notation that, here, (ˆ α, β) ˆ ˜ α, β)|α, β, σ 2 ) ≤ V(aT (˜ α, β)|α, β, σ 2 ) V(aT (ˆ ˜ of (α, β). (Note that the property of for any unbiased linear estimator (˜ α, β) unbiasedness is not particularly appealing when considered on its own.)

3.2

Classical Least Squares Estimator

71

An unbiased estimator of σ 2 is σ ˆ2 =

1 s2 ˆ T (y − α1 ˆ = (y − α ˆ 1n − Xβ) , ˆ n − Xβ) n−p−1 n−p−1

ˆ Note that the MLE and σ ˆ 2 (XT X)−1 approximates the covariance matrix of β. 2 2 2 2 of σ is not σ ˆ but σ ˜ = s /n. The standard t-statistics are deﬁned as (j = 1, . . . , p) βˆj − βj Tj = * ∼ T (n − p − 1, 0, 1) , σ ˆ 2 ωjj where ωjj denotes the (j, j)-th element of the matrix (XT X)−1 . These tstatistic are used in classical tests, for instance for testing H0 : βj = 0 versus H1 : βj = 0, the former being accepted at level γ if √ −1 σ ωjj < Fn−p−1 (1 − γ/2) |βˆj |/ˆ the (1 − γ/2)th quantile of the Student’s t T (n − p − 1, 0, 1) distribution (with location parameter 0 and scale parameter 1). The frequentist argument in using this bound (see Casella and Berger, 2001) is that the so-called p-value is smaller than γ, pj = PH0 (|Tj | > |tj |) < γ. Note that these statistics Tj can also be used when constructing marginal frequentist conﬁdence intervals on the βj ’s like √ √ −1 −1 βj ; βj − βˆj ≤ σ (1 − γ/2) = βj ; |Tj | ≤ σ ˆ ωjj Fn−p−1 (1 − γ/2) . ˆ ωjj Fn−p−1

From

a Bayesian perspective, we far from advocate the use of p-values in Bayesian settings or elsewhere since they suﬀer many defects (exposed for instance in Robert, 2007, Chap. 5), one being that they are often wrongly interpreted as probabilities of the null hypotheses.

For caterpillar, the unbiased estimate of σ 2 is equal to 0.7781 and the maximum likelihood estimates of α and of the components βj produced by the R command > summary(lm(y~X)) are given in Fig. 3.2, along with the least squares estimates of their respective standard deviations and p-values. According to the classical paradigm, the coeﬃcients β1 , β2 and β7 are the only ones considered to be signiﬁcant.

72

3 Regression and Variable Selection

We stress here that conditioning on X is valid only when X is exogenous, that is, only when we can write the joint distribution of (y, X) as f (y, X|α, β, σ 2 , δ) = f (y|α, β, σ 2 , X)f (X|δ) , where (α, β, σ 2 ) and δ are ﬁxed parameters. We can thus ignore f (X|δ) if the parameter δ is only a nuisance parameter since this part is independent3 of (α, β, σ 2 ). The practical advantage of using a regression model as above is that it is much easier to specify a realistic conditional distribution of one variable given p others rather than a joint distribution on all p + 1 variables. Note that if X is not exogenous, for instance when X involves past values of y (see Chap. 7), the joint distribution must be used instead. Residuals: Min 1Q Median -1.4710 -0.4474 -0.1769

3Q 0.6121

Max 1.5602

3Q 0.6121

Max 1.5602

lm(formula = y ˜ X) Residuals: Min 1Q Median -1.4710 -0.4474 -0.1769

Coefficients: Estimate Std. Error t value Pr(>|t|) (Intercept) -0.81328 0.15356 -5.296 1.97e-05 *** Xx1 -0.52722 0.21186 -2.489 0.0202 * Xx2 -0.39286 0.16974 -2.315 0.0295 * Xx3 0.65133 0.38670 1.684 0.1051 Xx4 -0.29048 0.31551 -0.921 0.3664 Xx5 -0.21645 0.16865 -1.283 0.2116 Xx6 0.29361 0.53562 0.548 0.5886 Xx7 -1.09027 0.47020 -2.319 0.0292 * Xx8 -0.02312 0.17225 -0.134 0.8944 --Signif. codes: 0 ?***? 0.001 ?**? 0.01 ?*? 0.05 ?.? 0.1 ? ? 1 Residual standard error: 0.8821 on 24 degrees of freedom Multiple R-squared: 0.6234,Adjusted R-squared: 0.4979

Fig. 3.2. Dataset caterpillar: R output providing the least squares estimates of the regression coeﬃcients along with their standard signiﬁcance analysis

3 From a Bayesian point of view, note that we would also need to impose prior independence between (α, β, σ 2 ) and δ to achieve this separation.

3.3

The Jeﬀreys Prior Analysis

73

3.3 The Jeﬀreys Prior Analysis Considering only the case of a complete lack of prior information on the parameters of the linear model, we ﬁrst describe a noninformative solution based on the Jeﬀreys prior. It is rather easy to show that the Jeﬀreys prior in this case is π J (α, β, σ 2 ) ∝ σ −2 , which is equivalent to a ﬂat prior on (α, β, log σ 2 ). We recall that 1 1 T 2 ¯ 1n − Xβ) (y − y ¯ 1n − Xβ) × exp − 2 (y − y (α, β, σ |y) = n/2 2σ (2πσ 2 ) ) ( n y − α)2 exp − 2 (¯ 2σ T 1 1 ˆ ˆ = exp − 2 y − α ˆ 1n − Xβ y−α ˆ 1n − Xβ × n/2 2σ (2πσ 2 ) n 1 ˆ T XT X(β − β) ˆ exp − 2 (ˆ α − α)2 − 2 (β − β) . 2σ 2σ The corresponding posterior distribution is therefore −n/2 1 ˆ T (y − α1 ˆ × π J (α, β, σ 2 |y) ∝ σ −2 exp − 2 (y − α ˆ 1n − Xβ) ˆ n − Xβ) 2σ n 1 ˆ T XT X(β − β) ˆ α − α)2 − 2 (β − β) σ −2 exp − 2 (ˆ 2σ 2σ −2 −p/2 1 ˆ T XT X(β − β) ˆ × ∝ σ exp − 2 (β − β) 2σ ) ( n −2 −1/2 exp − 2 (ˆ α − α)2 σ 2σ −2 −(n−p−1)/2−1 1 2 σ exp − 2 s . 2σ From this expression, we deduce the following (conditional and marginal) posterior distributions α|σ 2 , y ∼ N α ˆ , σ 2 /n , ˆ σ 2 (XT X)−1 , β|σ 2 , y ∼ Np β, σ 2 |y ∼ I G ((n − p − 1)/2, s2 /2) .

As in

every analysis involving an improper prior, one needs to check that the corresponding posterior distribution is proper. In this case, π(α, β, σ 2 |y) is proper when both n > p + 1 and rank [1n X] = p + 1. The former constraint requires that there be at least as many data points as there are parameters in the model, and, as already explained above, the latter is obviously necessary for identiﬁability reasons.

74

3 Regression and Variable Selection

The corresponding Bayesian estimates of α, β and σ 2 are thus given by Eπ [α|y] = α ˆ,

ˆ Eπ [β|y] = β

and Eπ [σ 2 |y] =

s2 , n−p−3

respectively. Unsurprisingly, the Jeﬀreys prior estimate of α is the empirical mean. Further, the posterior expectation of β is the maximum likelihood estimate. Note also that the Jeﬀreys prior estimate of σ 2 is larger (and thus more pessimistic) than both the maximum likelihood estimate s2 /n and the classical unbiased estimate s2 /(n − p − 1). The marginal posterior distribution of βj associated with the above joint distribution is T (n − p − 1, βˆj , ωjj s2 /(n − p − 1)) , (recall that ωjj = (XT X)−1 (j,j) ). Hence, the similarity with a frequentist analysis of this model is very strong since the classical (1 − γ) conﬁdence interval and the Bayesian HPD interval on βj coincide, even though they have diﬀerent interpretations. They are both equal to + −1 (1 − γ/2) ωjj s2 /(n − p − 1) . βj ; |βj − βˆj | ≤ Fn−p−1 For caterpillar, the Bayes estimate of σ 2 is equal to 0.8489. Figure 3.3 provides the corresponding (marginal) 95 % HPD intervals for each component of β. (It is obtained by the plotCI function, part of the gplots package.) Note that while some of these credible intervals include the value βj = 0 (represented by the dashed line), they do not necessarily validate acceptance of the null hypothesis H0 : βj = 0, which must be tested through a Bayes factor, as described below. This distinction is a major diﬀerence from the classical approach, where conﬁdence intervals are dual sets of acceptance regions.

3.4 Zellner’s G-Prior Analysis From this section onwards,4 we concentrate on a diﬀerent noninformative approach which was proposed by Arnold Zellner5 to handle linear regression from a Bayesian perspective. This approach is a middle-ground perspective where some prior information may be available on β and it is called Zellner’s G-prior, the “G” being the symbol used by Zellner in the prior variance. 4 In order to keep this coverage of G-priors simple and self-contained, we made several choices in the presentation that the most mature readers will possibly ﬁnd arbitrary, but this cannot be avoided if we want to keep the chapter at a reasonable length. 5 Arnold Zellner was a famous Bayesian econometrician, who wrote two reference books on Bayesian econometrics (Zellner, 1971, 1984)

Zellner’s G-Prior Analysis

3.4

75

l l

2

l l

4

l l

6

l

l

8

l

−2.0

−1.5

−1.0

−0.5

0.0

0.5

1.0

1.5

Fig. 3.3. Dataset caterpillar: Range of the credible 95 % HPD intervals for α (top row ) and each component of β when using the Jeﬀreys prior

3.4.1 A Semi-noninformative Solution When considering the likelihood (3.1) its shape is both Gaussian and Inverse Gamma, indeed, β given σ 2 appears in a Gaussian-like expression, while σ 2 involves an Inverse Gamma expression. This structure leads to a natural conjugate prior family, of the form ˜ σ 2 M −1 ) , (α, β)|σ 2 ∼ Np+1 ((˜ α, β), conditional on σ 2 , where M is a (p + 1, p + 1) positive deﬁnite symmetric matrix, and for σ 2 , a, b > 0 . σ 2 ∼ I G (a, b), (The conjugacy can be easily checked by the reader.) Even in the presence of genuine information on the parameters, the hyperparameters M , a and b are very diﬃcult to specify. Moreover, the posterior distributions, notably the posterior variances are sensitive to the speciﬁcation of these hyper-parameters.

76

3 Regression and Variable Selection

Therefore, given that a natural conjugate prior for the linear regression model has severe limitations, a more elaborate strategy is called for. The idea at the core of Zellner’s G-prior modeling is to allow the experimenter to introduce (possibly weak) information about the location parameter of the regression but to bypass the most diﬃcult aspects of the prior speciﬁcation, namely the derivation of the prior correlation structure. This structure is ﬁxed in Zellner’s proposal since the prior corresponds to ˜ gσ 2 (XT X)−1 , (3.2) β|α, σ 2 ∼ Np β, and a noninformative prior distribution is imposed on the pair (α, σ 2 ), π α, σ 2 ∝ σ −2 . (3.3) Zellner’s G-prior is thus decomposed as a (conditional) Gaussian prior for β and an improper (Jeﬀreys) prior for (α, σ 2 ). This modelling somehow appears as a data-dependent prior through its dependence on X, but this is not a genuine issue6 since the whole model is conditional on X. The experimenter ˜ and of the constant g. thus restricts prior determination to the choices of β As we will see once the posterior distribution is constructed, the factor g can be interpreted as being inversely proportional to the amount of information available in the prior relative to the sample. For instance, setting g = n gives the prior the same weight as one observation of the sample. We will use this as our default value.

Genuine data-dependent priors are not acceptable in a Bayesian analysis because they use the data twice and fail to enjoy the basic convergence properties of the Bayes estimators. (See Carlin and Louis, 1996, for a comparative study of the corresponding so-called empirical Bayes estimators.)

Note that, in the initial proposition of Zellner (1984), the parameter α is not modelled by a ﬂat prior distribution. It was instead considered to be a component of the vector β. (This was also the approach adopted in Marin and Robert 2007.) However, endowing α with a ﬂat prior ensures the location-scale invariance of the analysis, which means that changes in location or scale on y (like a switch from Celsius to Fahrenheit degrees for temperatures) do not impact on the resulting inference. We are now engaging into some algebra that will expose the properties of the G-posterior. First, we assume p > 0, meaning that there is at least one ex−1 T X . planatory variable inthe model. We deﬁne the matrix P = X XT X The prior π α, β, σ 2 can then be decomposed as 6 This choice is more problematic when conditioning on X is no longer possible, as for instance when X contains lagged dependent variables (Chap. 7) or endogenous variables.

3.4

Zellner’s G-Prior Analysis

1 ( T T ˜ β X Xβ − 2βT XT PXβ π α, β, σ 2 ∝ (σ 2 )−p/2 exp − 2gσ 2 1 ˜T T ˜ , X PX β β σ −2 exp − 2gσ 2

)

77

×

since XT PX = XT X. Therefore, π α, β, σ 2 |y ∝ (σ 2 )−n/2−p/2−1 1 T ¯ 1n − Xβ) (y − y ¯ 1n − Xβ) × exp − 2 (y − y 2σ ( n ) 1 ˜T T 2 ˜ exp − 2 (¯ y − α) × exp − β X PXβ × 2σ 2gσ 2 1 T T T ˜ exp − βTX Xβ − 2β X PXβ . 2gσ 2 Since 1T n X = 0p , we deduce that

1 T T 2 2 −n/2−p/2−1 T π α, β, σ |y ∝ (σ ) exp − 2 β X Xβ − 2y Xβ × 2σ 1 T ¯ 1n ) (y − y ¯ 1n ) × exp − 2 (y − y 2σ ) ( n 1 ˜T T 2 ˜ y − α) × exp − exp − 2 (¯ β X PXβ × 2σ 2gσ 2 1 T T T T ˜ exp − β X Xβ − 2β X PXβ . 2gσ 2

Since PX = X, we deduce that, conditionally on y, X and σ 2 , the parameters α and β are independent and such that ¯ , σ 2 /n , α|σ 2 , y ∼ N1 y σ2 g −1 g ˆ 2 T ˜ X X β|y, σ ∼ Np , β + Xβ/g , g+1 g+1 ˆ = XT X −1 XT y is the maximum likelihood (and least squares) where β estimator of β. The posterior independence between α and β is due to the fact that X is centered and that α and β are a priori independent. Moreover, the posterior distribution of σ 2 is given by

˜ − β) ˆ T XT X(β ˜ − β) ˆ (g + 1) σ 2 |y ∼ IG (n − 1)/2, s2 + (β where IG (a, b) is an inverse Gamma distribution with mean b/(a − 1) and ˆ T (y − y ˆ corresponds to the (classical) ¯ 1n − Xβ) ¯ 1n − Xβ) where s2 = (y − y residual sum of squares.

78

3 Regression and Variable Selection

The previous derivation assumes that p > 0. In the special case p = 0, which will later be used as a null model in hypothesis testing, similar arguments lead to ¯ , σ 2 /n , α|y, σ 2 ∼ N y ¯ 1n )T (y − y ¯ 1n ) 2 . σ 2 |y ∼ IG (n − 1)/2, (y − y (There is no β when p = 0, as this corresponds to the constant mean model.)

Recalling the double expectation formulas E [E [X|Y ]] = E [X]

and V(X) = V[E(X|Y )] + E[V(X|Y )] for V(X|Y ) = E[(X − E(X|Y ))2 Y ], we can derive from the previous derivations that ¯ Eπ [α|y] = Eπ Eπ α|σ 2 , y |y = Eπ [¯ y|y] = y and that Vπ (α|y) = V(¯ y|y) + E

σ 2 y = κ n(n − 3) , n

where ) 1 ( ˜ T XT PXβ ˜ − 2yT PXβ ˜ −gyT Py + β g+1

2 T T ˜ ˆ ˜ ˆ (g + 1) . = s + (β − β) X X(β − β)

¯ 1n ) + ¯ 1n )T (y − y κ = (y − y

With a bit of extra algebra, we can recover the whole distribution of α from ) ( κ ) ( n ¯ )2 exp − 2 κ , π(α, σ 2 |y) ∝ (σ −2 )(n−1)/2+1+1/2 exp − 2 (α − y 2σ 2σ namely

−n/2 ¯ )2 n(α − y . π(α|y) ∝ 1 + κ

This means that the marginal posterior distribution of α—the distribution of α given only y and X—is a Student’s t distribution with n − 1 degrees ¯ and a scale parameter equal to of freedom, a location parameter equal to y κ n(n − 1). If we now turn to the parameter β, by the same double expectation formula, we derive that Eπ [β|y] = Eπ Eπ β|σ 2 , y y g ˆ ˜ = Eπ (β + β/g)y g+1 g ˆ ˜ = (β + β/g) . g+1

3.4

Zellner’s G-Prior Analysis

79

This result gives its meaning to the above point relating g with the amount of information contained in the dataset. For instance, when g = 1, the prior information has the same weight as this amount. In this case, the Bayesian estimate of β is the average between the least square estimator and the prior expectation. The larger g is, the weaker the prior information and the closer the Bayesian estimator is to the least squares estimator. For instance, when ˆ g goes to inﬁnity, the posterior mean converges to β. Based on similar derivations, we can compute the posterior variance of β. Indeed, gσ 2 g ˆ ˜ π T −1 (β + β/g)|y + E (X X) V (β|y) = V g+1 g+1 κg = (XT X)−1 . (g + 1)(n − 3) Once more, it is possible to integrate out σ 2 in T T g + 1 π π β − E π(β, σ 2 |y) ∝ (σ 2 )−p/2 exp − [β|y] X X β − E [β|y] 2gσ 2 1 ×(σ 2 )−(n−1)/2−1 exp − 2 κ , 2σ leading to

T T g + 1 π π β − E [β|y] X X β − E [β|y] π(β|y) ∝ 1 + . gκ Therefore, the marginal posterior distribution of β is also a multivariate Student’s t distribution with n − 1 degrees of freedom, location parameter equal g ˆ ˜ gκ to (β + β/g) and scale parameter equal to (XT X)−1 . g+1 (g + 1)(n − 1) The standard Bayes estimator of σ 2 for this model is the posterior expectation

˜ − β) ˆ T XT X(β ˜ − β) ˆ (g + 1) 2 s2 + (β κ π = . E σ |y = n−3 n−3

In the special case p = 0, by using similar arguments, we get (y − y ¯ 1n ) ¯ 1n )T (y − y s2 Eπ σ 2 y = = , n−3 n−3 which is the same expectation as with the Jeﬀreys prior.

HPD regions on subvectors of the parameter β can be derived in a straightforward manner from this marginal posterior distribution of β. For a single parameter, we have for instance

80

3 Regression and Variable Selection

βj |y ∼ T

g n − 1, g+1

β˜j + βˆj g

gκ ωjj , (n − 1)(g + 1)

,

where ωjj is the (j, j)-th element of the matrix (XT X)−1 . If we set

˜ + g β) ˆ (g + 1) ζ = (β the transform βj − ζj

,

gκ ωjj (n − 1)(g + 1)

is (marginally) distributed as a standard t distribution with n − 1 degrees of freedom. A (1 − γ) HPD interval on βj has therefore gκ −1 ωjj Fn−1 (1 − γ/2) ζj ± (n − 1)(g + 1) −1 as bounds, where Fn−1 denotes the quantile function of the T (n − 1, 0, 1) distribution.

3.4.2 The BayesReg R Function We have created in bayess an R function called BayesReg to implement Zellner’s G-prior analysis within R. The purpose is dual: ﬁrst, this R function shows how easily automated this approach can be. Second, it also illustrates how it is possible to get exactly the same type of output as the standard R function summary(lm(y~X)). The following R code is extracted from this function BayesReg and used to calculate the Bayes estimates. As an aside, notice that we use the function stop in order to end the calculations if the matrix XT X is not invertible. if (det(t(X)%*%X) res1=BayesReg(y,X)

Intercept x1 x2 x3 x4 x5 x6 x7 x8

PostMean PostStError Log10bf EvidAgaH0 -0.8133 0.1407 -0.5039 0.1883 0.7224 (**) -0.3755 0.1508 0.5392 (**) 0.6225 0.3436 -0.0443 -0.2776 0.2804 -0.5422 -0.2069 0.1499 -0.3378 0.2806 0.4760 -0.6857 -1.0420 0.4178 0.5435 (**) -0.0221 0.1531 -0.7609

Posterior Mean of Sigma2: 0.6528 Posterior StError of Sigma2: 0.939

3.4.4 Prediction The prediction of m ≥ 1 future observations from units for which the explana˜ ˜ —have been observed or set tory variables X—but not the outcome variable y is also based on the posterior distribution. Logically enough, were α, β and ˜ would then have a Gaussian distribution σ 2 known quantities, the m-vector y ˜ and variance σ 2 Im . The predictive distribution on y ˜ is with mean α1m + Xβ deﬁned as the marginal in y of the joint posterior distribution on (˜ y, α, β, σ 2 ). ˜ of future observations has a Gaussian disConditional on σ 2 , the vector y tribution and we can derive its expectation—used as our Bayesian estimator— by averaging over α and β, Eπ [˜ y|σ 2 , y] = Eπ [Eπ (˜ y|α, β, σ 2 , y)|σ 2 , y] 2 ˜ = Eπ [α1m + Xβ|σ , y] ˜ ˆ ˜ β + gβ , =α ˆ 1m + X g+1 which is independent from σ 2 . This representation is quite intuitive, being the ˜ by the Bayesian estimator product of the matrix of explanatory variables X of β. Similarly, we can compute Vπ (˜ y|σ 2 , y) = Eπ [Vπ (˜ y|α, β, σ 2 , y)|σ 2 , y] +Vπ (Eπ (˜ y|α, β, σ 2 )|σ 2 , y) 2 ˜ , y) = Eπ [σ 2 Im |σ 2 , y] + Vπ (α1m + Xβ|σ g ˜ T X)−1 X ˜T . X(X = σ 2 Im + g+1

3.5

Markov Chain Monte Carlo Methods

85

Due to this factorization, and the fact that the conditional expectation does not depend on σ 2 , we thus obtain g ˜ T −1 ˜ T π 2 X(X X) X V (˜ y|y) = σ ˆ Im + . g+1 This decomposition of the variance makes perfect sense: Conditionally on σ 2 , the posterior predictive variance has two terms, the ﬁrst term being σ 2 Im , which corresponds to the sampling variation, and the second one g ˜ ˜ T , which corresponds to the uncertainty about β. X(XT X)−1 X being σ 2 g+1 HPD credible regions and tests can then be conducted based on this conditional predictive distribution ˜ |σ 2 , y, σ 2 ∼ N Eπ [˜ y y], Vπ (˜ y|y, σ 2 ) . ˜ leads to a mulIntegrating σ 2 out to produce the marginal distribution of y tivariate Student’s t distribution ˜ ˜ |y ∼ Tm n, α y + 1), ˆ 1m + g β/(g ) ˆ T XT Xβ ˆ ( s2 + β ˜T . ˜ T X)−1 X Im + X(X n (following a straightforward but lengthy derivation that is very similar to the one we conducted at the end of Chap. 2, see (2.11)).

3.5 Markov Chain Monte Carlo Methods Given the complexity of most models encountered in Bayesian modeling, standard simulation methods are not a suﬃciently versatile solution. We now present the rudiments of a technique that emerged in the late 1980s as the core of Bayesian computing and that has since then revolutionized the ﬁeld. This technique is based on Markov chains, but we will not make many incursions into the theory of Markov chains (see Meyn and Tweedie, 1993), focusing rather on the practical implementation of these algorithms and trusting that the underlying theory is sound enough to validate them (Robert and Casella, 2004). At this point, it is suﬃcient to recall that a Markov chain (xt )t∈N is a sequence of dependent random vectors whose dependence on the past values x0 , . . . , xt−1 stops at the value immediately before, xt−1 , and that is entirely deﬁned by its kernel—that is, the conditional distribution of xt given xt−1 . The central idea behind these new methods, called Markov chain Monte Carlo (MCMC) algorithms, is that, to simulate from a distribution π (for instance, the posterior distribution), it is actually suﬃcient to produce a Markov chain (xt )t∈N whose stationary distribution is π: when xt is marginally distributed according to π, then xt+1 is also marginally distributed according to

86

3 Regression and Variable Selection

π. If an algorithm that generates such a chain can be constructed, the ergodic theorem guarantees that, in almost all settings, the average T 1 g(xt ) T t=1

converges to E[g(x)], no matter what the starting value.8 More informally, this property means that, for large enough t, xt is approximately distributed from π and can thus be used like the output from a more standard simulation algorithm (even though one must take care of the correlation between the xt ’s created by the Markovian structure). For integral approximation purposes, the diﬀerence from regular Monte Carlo approximations is that the variance structure of the estimator is more complex because of the Markovian dependence. These methods being central to the cases studied from this stage onward, we hope that the reader will become suﬃciently proﬁcient with them by the end of the book! In this chapter, we detail a particular type of MCMC algorithm, the Gibbs sampler, that is currently suﬃcient for our needs. The next chapter will introduce a more universal type of algorithm. 3.5.1 Conditionals A ﬁrst remark that motivates the use of the Gibbs sampler9 is that, within structures such as π(x1 ) = π1 (x1 |x2 )˜ π (x2 ) dx2 , (3.5) to simulate from the joint distribution π(x1 , x1 ) = π1 (x1 |x2 )˜ π (x2 )

(3.6)

automatically produces (marginal) simulation from π(x1 ). Therefore, in settings where (3.5) holds, it is not necessary to simulate from π(x1 ) when one can jointly simulate (x1 , x2 ) from (3.6). For example, consider (x1 , x2 ) ∈ N×[0, 1] distributed from the joint density n π(x1 , x2 ) ∝ xx1 +α−1 (1 − x2 )n−x1 +β−1 . x1 2 This is a joint distribution where x1 |x2 ∼ B(n, x2 ) and x2 |α, β ∼ Be(α, β) . 8 In probabilistic terms, if the Markov chains produced by these algorithms are irreducible, then these chains are both positive recurrent with stationary distribution π and ergodic, that is, asymptotically independent of the starting value x0 . 9 In the literature, both the denominations Gibbs sampler and Gibbs sampling can be found. In this book, we will use Gibbs sampling for the simulation technique and Gibbs sampler for the simulation algorithm.

3.5

Markov Chain Monte Carlo Methods

87

Therefore, although π(x1 ) =

n B(α + x1 , β + n − x1 ) x1 B(α, β)

is available in closed form as the beta-binomial distribution, it is unnecessary (i) (i) to work with this marginal when one can simulate an iid sample (x1 , x2 ) (t = 1, . . . , N ) as (t)

(t)

(t)

x2 ∼ Be(α, β) and x1 ∼ B(n, x2 ) . Integrals such as E[x1 /(x1 + 1)] can then be approximated by N (t) 1 x1 , N i=1 x(t) + 1 1

using a regular Monte Carlo approach. Unfortunately, even when one works with a representation such as (3.6) that is naturally associated with the original model, it is often the case that the mixing density π ˜ (x2 ) itself is neither available in closed form nor amenable to simulation. However, both conditional posterior distributions, π1 (x1 |x2 )

and

π2 (x2 |x1 ) ,

can often be simulated, and the following method takes full advantage of this feature. 3.5.2 Two-Stage Gibbs Sampler The availability of both conditionals of (3.6) in terms of simulation can be exploited to build a transition kernel and a corresponding Markov chain, somewhat analogous to the derivation of the maximum of a multivariate function via an iterative device that successively maximizes the function in each of its arguments until a ﬁxed point is reached. The corresponding Markov kernel is built by simulating successively from each conditional distribution, with the conditioning variable being updated on the run. It is called the two-stage Gibbs sampler or sometimes the data augmentation algorithm, although both terms are rather misleading.10

10

Gibbs sampling got its name from Gibbs ﬁelds, used in image analysis, when Geman and Geman (1984) proposed an early version of this algorithm, while data augmentation refers to Tanner’s (1996) special use of this algorithm in missing-data settings, as seen in Chap. 6.

88

3 Regression and Variable Selection

Algorithm 3.3 Two-Stage Gibbs Sampler (0)

Initialization: Start with an arbitrary value x2 . (t−1) , generate Iteration t: Given x2 (t) (t−1) 1. x1 according to π1 (x1 |x2 ), (t) (t) 2. x2 according to π2 (x2 |x1 ) . (t)

Note that, in the second step of the algorithm, x2 is generated conditional (t) (t−1) . The validation of this algorithm is that, for both on x1 = x1 , not x1 generations, π is a stationary distribution. Therefore, the limiting distribution (t) (t) of the chain (x1 , x2 )t is π if the chain is irreducible; that is, if it can reach any region in the support of π in a ﬁnite number of steps. (Note that there is a diﬀerence between the stationary distribution and the limiting distribution only in cases when the chain is not ergodic, as shown in Exercise 3.9.) The practical implementation of Gibbs sampling involves solving two types of diﬃculties: the ﬁrst type corresponds to deriving an eﬃcient decomposition of the joint distribution in easily-simulated conditionals and the second one to deciding when to stop the algorithm. Evaluating the eﬃciency of the decomposition includes assessing the ease of simulating from both conditionals and the level of correlation between the x(t) ’s, as well as the mixing behavior of the chain, that is, its ability to explore the support of π suﬃciently fast. While deciding whether or not a given conditional can be simulated is easy enough, it is not always possible to ﬁnd a manageable conditional, and more robust alternatives such as the Metropolis–Hastings algorithm will be described in the following chapters (see Sect. 4.2). Choosing a stopping rule also relates to the mixing performances of the algorithm, as well as to its ability to approximate posterior expectations under π. Many indicators have been proposed in the literature (see Robert and Casella, 2004, Chap. 12) to signify convergence, or lack thereof, although none of these is foolproof. In the easiest cases, the lack of convergence is blatant and can be spotted on the raw plot of the sequence of the x(t) ’s, while, in other cases, the Gibbs sampler explores very satisfactorily one mode of the posterior distribution but fails altogether to visit the other modes of the posterior: we will encounter such cases in Chap. 6 with mixtures of distributions. Throughout this chapter and the following ones, we give hints on how to implement these recommendations in practice. Consider the posterior distribution derived in Exercise 2.11, for n = 2 observations, 2

e−μ /20 . π(μ|D2 ) ∝ {1 + (x1 − μ)2 )(1 + (x2 − μ)2 } Even though this is a univariate distribution, it can still be processed by a Gibbs sampler through a data augmentation step, thus illustrating the idea behind (3.5). In fact, since (j = 1, 2)

3.5

1 = 1 + (xj − μ)2

Markov Chain Monte Carlo Methods

∞

89

2

e−ωj [1+(xj −μ) ] dωj ,

0

we can deﬁne ω = (ω1 , ω2 ) and envision π(μ|D2 ) as the marginal distribution of 2 2 2 π(μ, ω|D2 ) ∝ e−μ /20 × e−ωj [1+(xj −μ) ] . j=1

For this multivariate distribution, a corresponding Gibbs sampler is associated with the following two steps: 1. Generate μ(t) ∼ π(μ|ω (t−1) , D2 ) . 2. Generate ω(t) ∼ π(ω|μ(t) , D2 ) . The second step is straightforward: the ωi ’s are conditionally independent and distributed as E xp(1 + (xi − μ(t) )2 ). The ﬁrst step is also well-deﬁned since π(μ|ω, D2 ) isa normal distribution with mean i ωi xi /( i ωi + 1/20) and variance 1/(2 i ωi +1/10). The corresponding R program then simpliﬁes into two lines

Fig. 3.4. (Top) Last 100 iterations of the chain (μ(t) ); (bottom) histogram of the chain (μ(t) ) and comparison with the target density for 10,000 iterations

90

3 Regression and Variable Selection

> mu = rnorm(1,sum(x*omega)/sum(omega+.05), + sqrt(1/(.1+2*sum(omega)))) > omega = rexp(2,1+(x-mu)^2) and the output of the simulation is represented in Fig. 3.4, with a very satisfying ﬁt between the histogram of the simulated values and the target. A detailed zoom on the last 100 iterations shows how the chain (μ(t) ) moves around, alternatively visiting each mode of the target.

When running a Gibbs sampler, the number of iterations should never be ﬁxed in advance: it is usually impossible to predict the performance of a given sampler before producing a corresponding chain. Deciding on the length of an MCMC run is therefore a sequential process where output behaviors are examined after pilot runs and new simulations (or new samplers) are chosen on the basis of these pilot runs.

3.5.3 The General Gibbs Sampler For a joint distribution π(x1 , . . . , xp ) with full conditionals π1 , . . . , πp where πj is the distribution of xj conditional on (x1 , . . . , xj−1 , xj+1 , . . . , xp ), the Gibbs sampler simulates successively from all conditionals, modifying one component of x at a time. The corresponding algorithmic representation is given in Algorithm 3.4. Algorithm 3.4 Gibbs Sampler (0)

(0)

Initialization: Start with an arbitrary value x(0) = (x1 , . . . , xp ) . (t−1) (t−1) Iteration t: Given (x1 , . . . , xp ), generate (t) (t−1) (t−1) , . . . , xp ), 1. x1 according to π1 (x1 |x2 (t) (t) (t−1) (t−1) , . . . , xp ), 2. x2 according to π2 (x2 |x1 , x3 .. . (t)

(t)

(t)

p. xp according to πp (xp |x1 , . . . , xp−1 ) .

Quite logically, the validation of this generalization of Algorithm 3.3 is identical: for each of the p steps of the t-th iteration, the joint distribution π(x) is stationary. Under the same restriction on the irreducibility of the chain, it also converges to π for every possible starting value. Note that the order in which the components of x are simulated can be modiﬁed at each iteration, either deterministically or randomly, without putting the validity of the algorithm in jeopardy.

3.6

Variable Selection

91

The two-stage Gibbs sampler naturally appears as a special case of Algorithm 3.4 for p = 2. It is, however, endowed with higher theoretical properties, as detailed in Robert and Casella (2004, Chap. 9) and Robert and Casella (2009, Chap. 7). To conclude this section, let us stress that the impact of MCMC on Bayesian statistics has been considerable. Since the 1990s, which saw the emergence of MCMC methods in the statistical community, the occurrence of Bayesian methods in applied statistics has greatly increased, and the frontier between Bayesian and “classical” statistics is now so fuzzy that in some ﬁelds, it has completely disappeared. From a Bayesian point of view, the access to far more advanced computational means has induced a radical modiﬁcation of the way people work with models and prior assumptions. In particular, it has opened the way to process much more complex structures, such as graphical models and latent variable models (see Chap. 6). It has also freed inference by opening for good the possibility of Bayesian model choice (see, e.g., Robert, 2007, Chap. 7). This expansion is much more visible among academics than among applied statisticians, though, given that the use of the MCMC technology requires some “hard” thinking to process every new problem. The availability of speciﬁc software such as BUGS has nonetheless given access to MCMC techniques to a wider community, starting with the medical ﬁeld. New modules in R and other languages like Python are also helping to bridge the gap.

3.6 Variable Selection 3.6.1 Deciding on Explanatory Variables In an ideal world, when building a regression model, we should include all relevant pieces of information, which in the regression context means including all predictor variables that might possibly help in explaining y. However, there are obvious drawbacks to the advice of increasing the number of explanatory variables. For one thing, in noninformative settings, this eventually clashes with the constraint p < n. For another, using a huge number of explanatory variables leaves little information available to obtain precise estimators. In other words, when we increase the explanatory scope of the regression model, we do not necessarily increase its explanatory power because it gets harder and harder to estimate the coeﬃcients.11 It is thus important to be 11 This phenomenon is related to the principle of parsimony, also called Occam’s razor, which states that, among two models with similar explanatory powers, the simplest one should always be preferred. It is also connected with the learning curve eﬀect found in information theory and neural networks, where the performance of a model increases on the learning dataset but decreases on a testing dataset as its complexity increases.

92

3 Regression and Variable Selection

able to decide which variables—within a large pool of potential explanatory variables—should be kept in a model that balances good explanatory power with good estimation performance. This is truly a decision problem in that all potential models have to be considered in parallel against a criterion that ranks them. This variable-selection problem can be formalized as follows. We consider a dependent random variable y and a set of p potential explanatory variables. At this stage, we assume that every subset of q explanatory variables could make a proper set of explanatory variables for the regression of y. The only restriction we impose is that the intercept (that is, the constant variable) is included in every model. There are thus 2p models in competition and we are looking for a procedure that selects the “best” model, that is, the “most relevant” explanatory variables. Note that this variable-selection procedure can alternatively be seen as a two-stage estimation setting where we ﬁrst estimate the indicator of the model (within the collection of models), which also amounts to estimating variable indicators, as detailed below, and we then estimate the parameters corresponding to this very model. Each of the 2p models under comparison is in fact associated with a binary indicator vector γ ∈ Γ = {0, 1}p , where γj = 1 means that the variable xj is included in the model, denoted by Mγ . This notation is quite handy since γ=(1,0,1,0,0,. . . ,1,0) clearly indicates which explanatory variables are in and which are not. We also use the notation qγ = 1T pγ for computing the number of variables included in the model Mγ . We deﬁne βγ as a sub-vector of β containing only the components such that xj is included in the model Mγ and Xγ as the sub-matrix of X where only the columns such that xj is included in the model Mγ have been left. The model Mγ is thus deﬁned as y|α, β γ , σ 2 , γ ∼ Nn α1n + βγ Xγ βγ , σ 2 In .

Once

again, and apparently in contradiction to our basic tenet that diﬀerent models should enjoy completely diﬀerent parameters, we are compelled to denote by σ 2 and α the variance and intercept terms common to all models, respectively. Although this is more of a mathematical trick than a true modeling reason, the prior independence of (α, σ 2 ) and γ allows for the simultaneous use of Bayes factors and an improper prior. Despite the possibly confusing notation, β γ and β are completely unrelated in that they are parameters of diﬀerent models.

3.6

Variable Selection

93

3.6.2 G-Prior Distributions for Model Choice Because so many models are in competition and thus considered in the global model all at once, we cannot expect a practitioner to specify one’s own prior on every model Mγ in a completely subjective and autonomous manner. We thus now proceed to derive all priors from a single global prior associated with the so-called full model that corresponds to γ = (1, . . . , 1). The argument goes as follows: (1) For the full model, we use Zellner’s G-prior as deﬁned in Sect. 3.4, ˜ gσ 2 (XT X)−1 ) and π(α, σ 2 ) ∝ σ −2 . β|σ 2 ∼ Np (β, (2) For each (sub-)model Mγ , the prior distribution of βγ conditional on σ 2 is ﬁxed as −1 γ γ 2 2 γT γ ˜ β |σ , γ ∼ Nqγ β , gσ X X , −1 ˜ and we use the same prior on (α, σ 2 ). ˜ γ = Xγ T Xγ ˜β Xγ T X where β

This distribution is conditional on γ; in particular, this implies that, while the variance notation σ 2 is common to all models, its distribution varies with γ.

Although there are many possible ways of deﬁning the prior on the model index12 γ, we opt for the uniform prior π(γ) = 2−p . The posterior distribution of γ (that is, the distribution of γ given y) is central to the variable-selection methodology since it is proportional to the marginal density of y in Mγ . In addition, for prediction purposes, the prediction distribution can be obtained by averaging over all models, the weights being the model probabilities (this is called model averaging). The posterior distribution of γ is π(γ|y) ∝ f (y|γ)π(γ) ∝ f (y|γ) ∝ (g + 1)−(qγ +1)/2 yT y −

−1 g yT Xγ Xγ T Xγ Xγ T y g+1 −(n−1)/2 1 ˜γ T γ T γ ˜ γ − . (3.7) β X X β g+1

When the number of explanatory variables is less than 15, say, the exact derivation of the posterior probabilities for all submodels can be undertaken. 12

For instance, one could instead use a uniform prior on the number qγ of explanatory variables or a more parsimonious prior such as π(γ) = 1/qγ .

94

3 Regression and Variable Selection

Indeed, 215 = 32768 means that the problem remains tractable. The following R code (part of the function ModChoBayesReg) is used to calculate those posterior probabilities and returns the top most probable models. The integrated likelihood for the null model is computed as intlike0. intlike=rep(intlike0,2^p) for (j in 2:2^p){ gam=as.integer(intToBits(i-1)[1:p]==1) pgam=sum(gam) Xgam=X[,which(gam==1)] Ugam=solve(t(Xgam)%*%Xgam)%*%t(Xgam) betatildegam=b1=Ugam%*%X%*%betatilde betamlgam=b2=Ugam%*%y s2gam=t(y-alphaml-Xgam%*%b2)%*%(y-alphaml-Xgam%*%b2) kappagam=as.numeric(s2gam+t(b1-b2)%*%t(Xgam)%*% Xgam%*%(b1-b2)/(g+1)) intlike[j]=(g+1)^(-pgam/2)*kappagam^(-(n-1)/2) } intlike=intlike/sum(intlike) modcho=order(intlike)[2^p:(2^p-9)] probtop10=intlike[modcho] The above R code uses the generic function intToBits to turn an integer i into the indicator vector gam. The remainder of the code is quite similar to the model choice code when computing the Bayes factors. For the caterpillar data, we set β˜ = 08 and g = 1. The models corresponding to the top 10 posterior probabilities are then given by > ModChoBayesReg(y,X,g=1) Number of variables less than 15 Model posterior probabilities are calculated exactly Top10Models 1 1 2 3 7 2 1 2 3 5 7 3 1 2 7 4 1 2 3 4 7 5 1 2 3 4 5 7 6 1 2 5 7 7 1 2 3 7 8 8 1 2 3 6 7 9 1 2 3 5 6 7 10 1 2 3 5 7 8

PostProb 0.0142 0.0138 0.0117 0.0112 0.0110 0.0108 0.0104 0.0102 0.0100 0.0098

3.6

Variable Selection

95

In a basic 0 − 1 decision setup, we would choose the model Mγ with the highest posterior probability—that is, the model with explanatory variables x1 , x2 , x3 and x7 —which corresponds to the variables – – – –

altitude, slope, the number of pine trees in the area, and the number of vegetation strata.

The model selected by the procedure thus fails to correspond to the three variables identiﬁed in the R output at the end of Sect. 3.4. But interestingly, even under this strong shrinkage prior g = 1 (where the prior has the same weight as the data), all top ten models contain the explanatory variables x1 , x2 and x7 , which have the most stars in this R analysis. Now, the default or noninformative calibration of the G-prior corresponds to the choice β˜ = 0p and g = n, which reduces the prior input to the equivalent of a single observation. Pushing g to a smaller value results in a paradoxical behaviour of the procedure which then usually picks the simpler model: this is another illustration of the Jeﬀreys-Lindley paradox, mentioned in Chap. 2. ˜ = 0p and g = n, the ten most likely models and their posterior For β probabilities are: > ModChoBayesReg(y,X) Number of variables less than 15 Models’s posterior probabilities are calculated exactly

1 2 3 4 5 6 7 8 9 10

Top10Models PostProb 1 2 7 0.0767 1 7 0.0689 1 2 3 7 0.0686 1 3 7 0.0376 1 2 6 0.0369 1 2 3 5 7 0.0326 1 2 5 7 0.0294 1 6 0.0205 1 2 4 7 0.0201 7 0.0198

For this diﬀerent prior modelling, we chose the same model as the lm classical procedure, rather than when g = 1; however, the posterior probabilities of the most likely models are much lower for g = 1, which is logical given that the current prior is less informative. Therefore, the top model is not as strongly supported as in the informative case. Once again, we stress that the choice g = 1 is rather arbitrary and that it is used here merely for illustrative purposes. The default value we recommend is g = n.

96

3 Regression and Variable Selection

3.6.3 A Stochastic Search for the Most Likely Model When the number p of variables is large, it becomes impossible to compute the posterior probabilities for the whole series of 2p models. We then need a tailored algorithm that samples from π(γ|y) and thus selects the most likely models, without computing ﬁrst all the values of π(γ|y). This can be done rather naturally by Gibbs sampling, given the availability of the full conditional posterior probabilities of the γj ’s. Indeed, if γ −j (1 ≤ j ≤ p) is the vector (γ1 , . . . , γj−1 , γj+1 , . . . , γp ), the full conditional distribution π(γj |y, γ −j ) of γj is proportional to π(γ|y) and can be computed in both γj = 0 and γj = 1 at no cost (since these are the only possible values of γj ). Algorithm 3.5 Gibbs Sampler for Variable Selection Initialization: Draw γ 0 from the uniform distribution on Γ . (t−1) (t−1) Iteration t: Given (γ1 , . . . , γp ), generate (t) (t−1) (t−1) , . . . , γp ), 1. γ1 according to π(γ1 |y, γ2 (t) (t) (t−1) (t−1) , . . . , γp ), 2. γ2 according to π(γ2 |y, γ1 , γ3 .. . (t)

(t)

(t)

p. γp according to π(γp |y, γ1 , . . . , γp−1 ) .

After a large number of iterations of this algorithm (that is, when the sampler is supposed to have converged or, more accurately, when the sampler has suﬃciently explored the support of the target distribution), its output can be used to approximate the posterior probabilities π(γ|y, X) by empirical averages based on the Gibbs output, ˆ π (γ = γ ∗ |y) = P

1 T − T0 + 1

T

Iγ (t) =γ ∗ ,

t=T0

where the T0 ﬁrst values are eliminated as burn-in. (The number T0 is therefore the number of iterations roughly needed to “reach” convergence.) The Gibbs output can also be used to approximate the inclusion of a given variable, P π (γj = 1|y, X), as ˆ π (γj = 1|y) = P

1 T − T0 + 1

T t=T0

Iγ (t) =1 , j

with the same asymptotic validation. The following R code (again part of the function ModChoBayesReg) describes our implementation of the above variable-selection Gibbs sampler.

3.6

Variable Selection

97

The code uses the null model with only the intercept α as a reference, based on the integrated likelihood intlike0 as above. It then starts at random in the collection of models: gamma=rep(0,niter) mcur=sample(c(0,1),p,replace=TRUE) gamma[1]=sum(2^(0:(p-1))*mcur)+1 pcur=sum(mcur) and computes the corresponding integrated likelihood intlikecur if (pcur==0) intlikecur=intlike0 else{ #integrated likelihood Xcur=X[,which(mcur==1)] Ucur=solve(t(Xcur)%*%Xcur)%*%t(Xcur) betatildecur=b1=Ucur%*%X%*%betatilde betamlcur=b2=Ucur%*%y s2cur=t(y-alphaml-Xcur%*%b2)%*%(y-alphaml-Xcur%*%b2) kappacur=as.numeric(s2cur+t(b1-b2)%*%t(Xcur)%*% Xcur%*%(b1-b2)/(g+1)) intlikecur=(g+1)^(-pcur/2)*kappacur^(-(n-1)/2) } It then proceeds according to Algorithm 3.5, proposing to change one variable indicator γj and accepting this move with a Metropolis–Hastings (deﬁned and justiﬁed in Chap. 4) probability: if (runif(1) 0 .

Taking advantage of the matrix identities −1 −1 −1 M + XT X = M−1 − M−1 M−1 + (XT X)−1 M −1 T −1 = (XT X)−1 − (XT X)−1 M−1 + (XT X)−1 (X X) and −1 XT X(M + XT X)−1 M = M−1 (M + XT X)(XT X)−1 −1 = M−1 + (XT X)−1 , establish that

ˆ + Mβ}, ˜ σ 2 (M + XT X)−1 β|y, σ 2 ∼ Np (M + XT X)−1 {(XT X)β

(3.8)

ˆ = (XT X)−1 XT y and where β 2 ˜ − β) ˆ ˜ − β) ˆ T M−1 + (XT X)−1 −1 (β ( β n s σ 2 |y ∼ I G + a, b + + 2 2 2 (3.9) ˆ T (y − βX) ˆ where s2 = (y − βX) are the correct posterior distributions. Give a (1 − α) HPD region on β. 3.5 The regression model of Exercise 3.4 can also be used in a predictive sense: ˜ i.e., when predicting m unobserved for a given (m, p + 1) explanatory matrix X, ˜ can be inferred through the predictive variates y˜i , the corresponding outcome y distribution π(˜ y |σ 2 , y). Show that π(˜ y|σ 2 , y) is a Gaussian density with mean ˜ ˆ + Mβ) ˜ Eπ [˜ y|σ 2 , y] = X(M + XT X)−1 (XT Xβ

100

3 Regression and Variable Selection

and covariance matrix ˜ T) . ˜ y|σ 2 , y) = σ 2 (Im + X(M + XT X)−1 X Vπ (˜ Deduce that

˜ ˆ + Mβ), ˜ ˜ |y ∼ Tm n + 2a, X(M + XT X)−1 (XT Xβ y ˜ − β) ˆ T M−1 + (XT X)−1 −1 (β ˜ − β) ˆ 2b + s2 + (β n + 2a ) ( T ˜T . ˜ + X X)−1 X × Im + X(M

3.6 Show that the marginal distribution of y associated with (3.8) and (3.9) is given by b −1 T ˜ y ∼ Tn 2a, Xβ, (In + XM X ) . a 3.7 Show that the matrix (In + gX(XT X)−1 XT ) has 1 and g + 1 as only eigenvalues. (Hint: Show that the eigenvectors associated with g + 1 are of the form Xβ and that the eigenvectors associated with 1 are those orthogonal to X.) Deduce that the determinant of the matrix (In + gX(XT X)−1 XT ) is indeed (g + 1)p+1 . ˜ , m dimensional 3.8 Under the Jeﬀreys prior, give the predictive distribution of y ˜ vector corresponding to the (m, p) matrix of explanatory variables X. 3.9 If (x1 , x2 ) is distributed from the uniform distribution on (x1 , x2 ); (x1 − 1)2 + (x2 − 1)2 ≤ 1 ∪ (x1 , x2 ); (x1 + 1)2 + (x2 + 1)2 ≤ 1 , show that the Gibbs sampler does not produce an irreducible chain. For this distribution, ﬁnd an alternative Gibbs sampler that works. (Hint: Consider a rotation of the coordinate axes.) 3.10 If a joint density g(y1 , y2 ) corresponds to the conditional distributions g1 (y1 |y2 ) and g2 (y2 |y1 ), show that it is given by g(y1 , y2 ) =

g2 (y2 |y1 ) . g2 (v|y1 )/g1 (y1 |v) dv

3.11 Considering the model η|θ ∼ Bin(n, θ) ,

θ ∼ Be(a, b),

derive the joint distribution of (η, θ) and the corresponding full conditional distributions. Implement a Gibbs sampler associated with those full conditionals and compare the outcome of the Gibbs sampler on θ with the true marginal distribution of θ.

3.7

Exercises

101

3.12 Take the posterior distribution on (θ, σ 2 ) associated with the joint model xi |θ, σ 2 ∼ N (θ, σ 2 ), θ ∼ N (θ0 , τ 2 ) ,

i = 1, . . . , n, σ 2 ∼ IG (a, b) .

Show that the full conditional distributions are given by σ2 σ2 τ 2 nτ 2 2 ¯, 2 x θ|x, σ ∼ N θ0 + 2 σ 2 + nτ 2 σ + nτ 2 σ + nτ 2

and σ |x, θ ∼ IG 2

1 n 2 + a, (xi − θ) + b , 2 2 i

where x¯ is the empirical average of the observations. Implement the Gibbs sampler associated with these conditionals.

4 Generalized Linear Models

This was the sort of thing that impressed Rebus: not nature, but ingenuity. —Ian Rankin, A Question of Blood.—

Roadmap Generalized linear models are extensions of the linear regression model described in the previous chapter. In particular, they avoid the selection of a single transformation of the data that must achieve the possibly conﬂicting goals of normality and linearity imposed by the linear regression model, which is for instance impossible for binary or count responses. The trick that allows both a feasible processing and an extension of linear regression is ﬁrst to turn the covariates into a real number by a linear projection and then to transform this value so that it ﬁts the support of the response. We focus here on the Bayesian analysis of probit and logit models for binary data and of log-linear models for contingency tables. On the methodological side, we present a general MCMC method, the Metropolis–Hastings algorithm, which is used for the simulation of complex distributions where both regular and Gibbs sampling fail. This includes in particular the random walk Metropolis–Hastings algorithm, which acts like a plain vanilla MCMC algorithm.

J.-M. Marin and C.P. Robert, Bayesian Essentials with R, Springer Texts in Statistics, DOI 10.1007/978-1-4614-8687-9 4, © Springer Science+Business Media New York 2014

103

104

4 Generalized Linear Models

4.1 A Generalization of the Linear Model 4.1.1 Motivation

Status

Bottom margin width (mm)

In the previous chapter, we modeled the connection between a response variable y and a vector x of explanatory variables by a linear dependence relation with normal perturbations. There are many instances where both the linearity and the normality assumptions are not appropriate, especially when the support of y is restricted to R+ or N. For instance, in dichotomous models, y takes its values in {0, 1} as it represents the indicator of occurrence of a particular event (death in a medical study, unemployment in a socioeconomic study, migration in a capture–recapture study, etc.); in this case, a linear conditional expectation E[y|x, β] = xT β would be fairly cumbersome to handle, both in terms of the constraints on β and the corresponding distribution of the error ε = y − E[y|x, β].

Bottom margin width (mm)

Status

Fig. 4.1. Dataset bank: (left) Plot of the status indicator versus the bottom margin width; (right) boxplots of the bottom margin width for both counterfeit statuses

The bank dataset we analyze in the ﬁrst part of this chapter comes from Flury and Riedwyl (1988) and is made of four measurements on 100 genuine Swiss banknotes and 100 counterfeit ones. The response variable y is thus the status of the banknote, where 0 stands for genuine and 1 stands for counterfeit, while the explanatory factors are the length of the bill x1 , the width of the left edge x2 , the width of the right edge x3 , and the bottom margin width x4 ,

4.1

A Generalization of the Linear Model

105

all expressed in millimeters. We want a probabilistic model that predicts the type of banknote (i.e., that detects counterfeit banknotes) based on the four measurements above. To motivate the introduction of the generalized linear models, we only consider here the dependence of y on the fourth measure, x4 , which again is the bottom margin width of the banknote. To start, the yi ’s being binary, the conditional distribution of y given x4 cannot be normal. Nonetheless, as shown by Fig. 4.1, the variable x4 clearly has a strong inﬂuence on whether the banknote is or is not counterfeit. To model this dependence in a proper manner, we must devise a realistic (if not real!) connection between y and x4 . The fact that y is binary implies a speciﬁc form of dependence: Indeed, both its marginal and conditional distributions necessarily are Bernoulli distributions. This means that, for instance, the conditional distribution of y given x4 is a Bernoulli B(p(x4 )) distribution; that is, for x4 = x4i , there exists 0 ≤ pi = p(x4i ) ≤ 1 such that P(yi = 1|x4 = x4i ) = pi , which turns out to be also the conditional expectation of yi , E[yi |x4i ]. If we do impose a linear dependence on the pi ’s, namely, p(x4i ) = β0 + β1 x4i , the maximum likelihood estimates of β0 and β1 are then equal to −2.02 and 0.268, leading to the estimated prediction equation pˆi = −2.02 + 0.268xi4 .

(4.1)

This implies that a banknote with bottom margin width equal to 8 is counterfeit with probability −2.02 + 0.268 × 8 = 0.120 . Thus, this banknote has a relatively small probability of having been counterfeited, which coincides with the intuition drawn from Fig. 4.1. However, if we now consider a banknote with bottom margin width equal to 12, (4.1) implies that this banknote is counterfeited with probability −2.02 + 0.268 × 12 = 1.192 , which is certainly embarrassing for a probability estimate! We could try to modify the result by truncating the probability to (0, 1) and by deciding that this value of x4 almost certainly indicates a counterfeit, but still there is a fundamental diﬃculty with this model. The fact that an ordinary linear

106

4 Generalized Linear Models

dependence can predict values outside (0, 1) suggests that the connection between this explanatory variable and the probability of a counterfeit cannot be modeled through a linear function but rather can be achieved using functions of x4i that take their values within the interval (0, 1). 4.1.2 Link Functions As shown by the previous analysis, while linear models are nice to work with, they also have strong limitations. Therefore, we need a broader class of models to cover various dependence structures. The class selected for this chapter is called the family of generalized linear models (GLM), which has been formalized in McCullagh and Nelder (1989). This nomenclature stems from the fact that the dependence of y on x is partly linear in the sense that the conditional distribution of y given x is deﬁned in terms of a linear combination xT β of the components of x, y|β ∼ f (y|xT β) . As in the previous chapter, we use the notation y = (y1 , . . . , yn ) for a sample of n responses and ⎤ ⎡ x11 x12 . . . x1k ⎢ x21 x22 . . . x2k ⎥ ⎥ ⎢ ⎥ ⎢ X = [x1 . . . xk ] = ⎢ x31 x32 . . . x3k ⎥ ⎢ .. .. .. .. ⎥ ⎣ . . . . ⎦ xn1 xn2 . . . xnk for the n × k matrix of corresponding explanatory variables, possibly with x11 = . . . = xn1 = 1. We use y and x as generic notations for single-response and covariate vectors, respectively. Once again, we will omit the dependence on x or X to simplify notations. A generalized linear model is speciﬁed by two functions: 1. a conditional density f of y given x that belongs to an exponential family (Sect. 2.2.3) and that is parameterized by an expectation parameter μ = μ(x) = E[y|x] and possibly a dispersion parameter ϕ > 0 that does not depend on x; and 2. a link function g that relates the mean μ = μ(x) of f and the covariate vector, x, as g(μ) = (xT β), β ∈ Rk . For identiﬁability reasons, the link function g is a one-to-one function and we have E[y|β, ϕ] = g −1 xT β . We can thus write the (conditional) likelihood as (β, ϕ|y) =

n f yi |xiT β, ϕ i=1

4.1

A Generalization of the Linear Model

107

if we choose to reparameterize f with the transform g(μi ) of its mean and if we denote by xi the covariate vector for the ith observation.1 The ordinary linear regression is obviously a special case of GLM where g(x) = x, ϕ = σ 2 and y|β, σ 2 ∼ N xT β, σ 2 . However, outside the linear model, the interpretation of the coeﬃcients βi is much more delicate because these coeﬃcients do not relate directly to the observables, due to the presence of a link function that cannot be the identity. For instance, in the logistic regression model (deﬁned in the following paragraph), the linear dependence is deﬁned in terms of the log-odds ratio log{p1 /(1 − pi )}. The most widely used GLMs are presumably those that analyze binary data, as in bank, that is, when yi ∼ B(1, pi ) (with μi = pi = p(xiT β)). The mean function p thus transforms a real value into a value between 0 and 1, and a possible choice of link function is the logit transform, g(p) = log{p/(1 − p)} , associated with the logistic regression model. Because of the limited support of the responses yi , there is no dispersion parameter in this model and the corresponding likelihood function is yi 1−yi n exp(xiT β) 1 1 + exp(xiT β) 1 + exp(xiT β) i=1 n n

iT = exp 1 + exp(xiT β) . yi x β

(β|y) =

i=1

(4.2)

i=1

It thus fails to factorize conveniently because of the denominator: there is no manageable conjugate prior for this model, called the logit model. There exists a speciﬁc form of link function for each exponential family which is called the canonical link. This canonical function is chosen as the function g of the expectation parameter that appears in the exponent of the natural exponential family representation of the probability density, namely g (μ) = θ

if

f (y|μ, ϕ) = h(y) exp ϕ{T (y) · θ − Ψ (θ)} .

Since the logistic regression model can be written as pi f (yi |pi ) = exp yi log + log(1 − pi ) , 1 − pi the logit link function is the canonical version for the Bernoulli model. Note that, while it is customary to use the canonical link, there is no compelling reason to do so, besides following custom! 1

This upper indexing allows for the distinction between xi , the ith component of the covariate vector, and xi , the ith vector of covariates in the sample.

108

4 Generalized Linear Models

For binary response variables, many link functions can be substituted for the logit link function. For instance, the probit link function, g(μi ) = Φ−1 (μi ), where Φ is the standard normal cdf, is often used in econometrics. The corresponding likelihood is (β|y) ∝

n

1−yi Φ(xiT β)yi 1 − Φ(xiT β) .

(4.3)

i=1

Although this alternative is also quite arbitrary and any other cdf could be used as a link function (such as the logistic cdf associated with (4.2)), the probit link function enjoys a missing-data (Chap. 6) interpretation that clearly boosted its popularity: This model can indeed be interpreted as a degraded linear regression model in the sense that observing yi = 1 corresponds to the case zi ≥ 0, where zi is a latent (that is, unobserved) variable such that zi ∼ N xiT β, 1 . In other words, y = I(zi ≥ 0) appears as a dichotomized linear regression response. Of course, this perspective is only an interpretation of the probit model in the sense that there may be no hidden zi ’s at all in the real world! In addition, the probit and logistic regression models have quite similar behaviors, diﬀering mostly in the tails. Another type of GLM deals with unbounded integer-valued variables. The Poisson regression model starts from the assumption that the yi ’s are Poisson P(μi ) and it selects a link function connecting R+ bijectively with R, such as, for instance, the logarithmic function, g(μi ) = log(μi ). This model is thus a count model in the sense that the responses are integers, for instance the number of deaths due to lung cancer in a county or the number of speeding tickets issued on a particular stretch of highway, and it is quite common in epidemiology. The corresponding likelihood is n 1 (β|y) = exp yi xiT β − exp(xiT β) , y ! i i=1 where the factorial terms (1/yi !) are irrelevant for both likelihood and posterior computations. Note that it does not factorize conveniently because of the exponential terms within the exponential. The three examples above are simply illustrations of the versatility of generalized linear modeling. In this chapter, we discuss only two types of data for which generalized linear modeling is appropriate. We refer the reader to McCullagh and Nelder (1989) and Gelman et al. (2013) for a much more detailed coverage.

4.2 Metropolis–Hastings Algorithms As partly hinted by the previous examples, posterior inference in GLMs is much harder than in linear models because of less manageable (and nonfactorizing) likelihoods, which explains the longevity and versatility of linear

4.2

Metropolis–Hastings Algorithms

109

model studies over the past centuries! Working with a GLM typically requires speciﬁc numerical or simulation tools. We take the opportunity of this requirement to introduce a universal MCMC method called the Metropolis–Hastings algorithm. Its range of applicability is incredibly broad (meaning that it is by no means restricted to GLM applications) and its inclusion in the Bayesian toolbox in the early 1990s has led to considerable extensions of the Bayesian ﬁeld.2 4.2.1 Definition When compared with the Gibbs sampler, Metropolis–Hastings algorithms are generic (or oﬀ-the-shelf) MCMC algorithms in the sense that they can be tuned toward a much wider range of possibilities. Those algorithms are also a natural extension of standard simulation algorithms such as accept–reject (see Chap. 5) or sampling importance resampling methods since they are all based on a proposal distribution. However, a major diﬀerence is that, for the Metropolis–Hastings algorithms, the proposal distribution is Markov, with kernel density q(x, y). If the target distribution has density π, the Metropolis– Hastings algorithm is as follows: Algorithm 4.6 Generic Metropolis–Hastings Sampler Initialization: Choose an arbitrary starting value x(0) . Iteration t (t ≥ 1): ˜ ∼ q(x(t−1) , x). 1. Given x(t−1) , generate x 2. Compute π(˜ x)/q(x(t−1) , x ˜) (t−1) ,1 . ,x ˜) = min ρ(x π(x(t−1) )/q(˜ x, x(t−1) ) 3. With probability ρ(x(t−1) , x ˜), accept x˜ and set x(t) = x˜; otherwise reject x˜ and set x(t) = x(t−1) .

The distribution q is also called the instrumental distribution. As in the accept–reject method (Sect. 5.4), we only need to know either π or q up to a proportionality constant since both constants cancel in the calculation of ρ. Note also the advantage of this approach compared with the Gibbs sampler: it is not necessary to use the conditional distributions of π. The strong appeal of this algorithm is that it is rather universal in its formulation as well as in its use. Indeed, we only need to simulate from a 2 This algorithm had been used by particle physicists, including Metropolis, since the late 1940s, but, as is often the case, the connection with statistics was not made until much later!

110

4 Generalized Linear Models

proposal q that can be chosen quite freely. There is, however, a theoretical constraint, namely that the chain produced by this algorithm must be able to explore the support of π(y) in a ﬁnite number of steps. As discussed below, there also are many practical diﬃculties that are such that the algorithm may lose its universal feature and that it may require some speciﬁc tuning for each new application. The theoretical validation of this algorithm is the same as with other MCMC algorithms: The target distribution π is the limiting distribution of the Markov chain produced by Algorithm 4.6. This is due to the choice of the acceptance probability ρ(x, y) since the so-called detailed balance equation π(x)q(x, y)ρ(x, y) = π(y)q(y, x)ρ(y, x) holds and thus implies that π is stationary by integrating out x. While theoretical guarantees that the algorithm converges are very high, the choice of q remains essential in practice. Poor choices of q may indeed result either in a very high rejection rate, meaning that the Markov chain (x(t) )t hardly moves, or in a myopic exploration of the support of π, that is, in a dependence on the starting value x(0) such that the chain is stuck in a neighborhood region of x(0) . A particular choice of proposal q may thus work well for one target density but be extremely poor for another one. While the algorithm is indeed universal, it is impossible to prescribe applicationindependent strategies for choosing q. We thus consider below two speciﬁc cases of proposals and brieﬂy discuss their pros and cons (see Robert and Casella, 2004, Chap. 7, for a detailed discussion). 4.2.2 The Independence Sampler The choice of q closest to the accept–reject method (see Algorithm 5.9) is to pick a constant q that is independent of its ﬁrst argument, q(x, y) = q(y) . In that case, ρ simpliﬁes into π(y)/q(y) . ρ(x, y) = min 1, π(x)/q(x) In the special case in which q is proportional to π, we obtain ρ(x, y) = 1 and the algorithm reduces, as expected, to iid sampling from π. The analogy with the accept–reject algorithm is that the maximum of the ratio π/q is replaced with the current value π(x(t−1) )/q(x(t−1) ) but the sequence of accepted x(t) ’s is not iid because of the acceptance step. The convergence properties of the algorithm depend on the density q. First, q needs to be positive everywhere on the support of π. Second, for good

4.2

Metropolis–Hastings Algorithms

111

exploration of this support, it appears that the ratio π/q needs to be bounded (see Robert and Casella, 2004, Theorem 7.8). Otherwise, the chain may take too long to reach some regions with low q/π values. This constraint obviously reduces the appeal of using an independence sampler, even though the fact that it does not require an explicit upper bound on π/q may sometimes be a plus. This type of MH sampler is thus very model-dependent, and it suﬀers from the same drawbacks as the importance sampling methodology, namely that tuning the “right” proposal becomes much harder as the dimension increases. 4.2.3 The Random Walk Sampler Since the independence sampler requires too much global information about the target distribution that is diﬃcult to come by in complex or highdimensional problems, an alternative is to opt for a local gathering of information, clutching to the hope that the accumulated information will provide, in the end, the global picture. Practically, this means exploring the neighborhood of the current value x(t) in search of other points of interest. The simplest exploration device is based on random walk dynamics. A random walk proposal is based on a symmetric transition kernel q(x, y) = qRW (y − x) with qRW (x) = qRW (−x). Symmetry implies that the acceptance probability ρ(x, y) reduces to the simpler form

ρ(x, y) = min 1, π(y) π(x) . The appeal of this scheme is obvious when looking at the acceptance probability, since it only depends on the target π and since this version accepts all proposed moves that increase the value of π. There is considerable ﬂexibility in the choice of the distribution qRW , at least in terms of scale (i.e., the size of the neighborhood of the current value) and tails. Note that while from a probabilistic point of view random walks usually have no stationary distribution, the algorithm biases the random walk by moving toward modes of π more often than moving away from them. The ambivalence of MCMC methods like the Metropolis–Hastings algorithm is that they can be applied to virtually any target. This is a terriﬁc plus in that they can tackle new models, but there is also a genuine danger that they simultaneously fail to converge and fail to signal that they have failed to converge! Indeed, these algorithms can produce seemingly reasonable results, with all outer aspects of stability, while they are missing major modes of the target distribution. For instance, particular attention must be paid to models where the number of parameters exceeds by far the size of the dataset. 4.2.4 Output Analysis and Proposal Design An important problem with the implementation of an MCMC algorithm is to gauge when convergence has been achieved; that is, to assess at what point the distribution of the chain is suﬃciently close to its asymptotic distribution

112

4 Generalized Linear Models

for all practical purposes or, more practically, when it has covered the whole support of the target distribution with suﬃcient regularity. The number of iterations T0 that is required to achieve this goal is called the burn-in period. It is usually sensible to discard simulated values within this burn-in period in the Monte Carlo estimation so that the bias caused by the starting value is reduced. However, and this is particularly true in high dimensions, the empirical assessment of MCMC convergence is extremely delicate, to the point that it is rarely possible to be certain that an algorithm has converged.3 Nevertheless, some partial convergence diagnostic procedures can be found in the literature (see Robert and Casella, 2004, Chap. 12, and Robert and Casella, 2009, Chap. 8). In particular, the latter describes the R package coda in Sect. 8.2.4. A ﬁrst way to assess whether or not a chain is in its stationary regime is to visually compare trace plots of sequences started at diﬀerent values, as it may expose diﬃculties related, for instance, to multimodality. In practice, when chains of length T from two starting values have visited substantially diﬀerent parts of the state space, the burn-in period for at least one of the chains should be greater than T . Note, however, that the problem of obtaining overdispersed starting values can be diﬃcult when little is known about the target density, especially in large dimensions. Autocorrelation plots of particular components provide in addition good indications of the chain’s mixing behavior. If ρk (k ∈ N∗ ) denotes the kthorder autocorrelation, ρk = cov x(t) , x(t+k) , these quantities can be estimated from the observed chain itself,4 at least for small values of k, and an eﬀective sample size factor can be deduced from these estimates, −1/2 T0 T ess = T 1 + 2 ρˆk , k=1

where ρˆk is the empirical autocorrelation function. This quantity represents the sample size of an equivalent iid sample when running T iterations. Conversely, the ratio T /T ess indicates the multiplying factor on the minimum number of iid iterations required to run a simulation. Note, however, that this is only a partial indicator: Chains that remain stuck in one of the modes of the target distribution may well have a high eﬀective ratio. While we cannot discuss at length the selection of the proposal distribution (see Robert and Casella, 2004, Chap. 7, and Robert and Casella, 2009, 3 Guaranteed convergence as in accept–reject algorithms is sometimes achievable with MCMC methods using techniques such as perfect sampling or renewal. But such techniques require a much more advanced study of the target distribution and the transition kernel of the algorithm. These conditions are not met very often in practice (see Robert and Casella 2004, Chap. 13). 4 In R, this estimation can be conducted using the acf function.

4.2

Metropolis–Hastings Algorithms

113

Chap. 6), we stress that this is an important choice that has deep consequences for the convergence properties of the simulated Markov chain and thus for the exploration of the target distribution. As for prior distributions, we advise the simultaneous use of diﬀerent kernels to assess their performances on the run. When considering a random walk proposal, for instance, a quantity that needs to be calibrated against the target distribution is the scale of this random walk. Indeed, if the variance of the proposal is too small with respect to the target distribution, the exploration of the target support will be small and may fail in more severe cases. Similarly, if the variance is too large, this means that the proposal will most often generate values that are outside the support of the target and that the algorithm will reject a large portion of attempted transitions.

It seems reasonable to tune the proposal distribution in terms of its past performances, for instance by increasing the variance if the acceptance rate is high or decreasing it otherwise (or moving the location parameter toward the mean estimated over the past iterations). This must not be implemented outside a burn-in step, though, because a permanent modiﬁcation of the proposal distribution amounts to taking into account the whole past of the sequence and thus it cancels both its Markovian nature and its convergence guarantees.

Consider, solely for illustration purposes, the standard normal distribution N (0, 1) as a target. If we use Algorithm 4.6 with a normal random walk, i.e., x˜|x(t−1) ∼ N x(t−1) , σ 2 , the performance of the sampler depends on the value σ. An R function that implements the associated Hastings–Metropolis sampler is coded as hm=function(n,x0,sigma2){ x=rep(x0,n) for (i in 2:n){ y=rnorm(1,x[i-1],sqrt(sigma2)) if (runif(1) mod=summary(glm(y~X,family=binomial(link="probit"))) ˆ and mod$cov.unscaled to Σ. ˆ The with mod$coeff[,1] corresponding to β following code is then reproducing the above algorithm in R:: hmflatprobit=function(niter,y,X,scale){ p=dim(X)[2] mod=summary(glm(y~-1+X,family=binomial(link="probit"))) beta=matrix(0,niter,p) beta[1,]=as.vector(mod$coeff[,1]) Sigma2=as.matrix(mod$cov.unscaled) 7

A choice of parameters that depend on the data for the Metropolis–Hastings proposal is completely valid, both from an MCMC point of view (meaning that this is not a self-tuning algorithm) and from a Bayesian point of view (since the parameters of the proposal are not those of the prior).

4.3

The Probit Model

117

for (i in 2:niter){ tildebeta=rmnorm(1,beta[i-1,],scale*Sigma2) llr=probitll(tildebeta,y,X)-probitll(beta[i-1,],y,X) if (runif(1) mod=summary(glm(y~-1+X,family=poisson())) ˆ as mod$coeﬀ[,1] and Σ ˆ as mod$cov.unscaled). which provides β For airquality, we ﬁrst consider the most general nonsaturated model, as described in Sect. 4.5.1. Taking into account the identiﬁability constraints, there are therefore 1+(2−1)+(2−1)+(5−1)+(2−1)×(2−1)+(2−1)×(5−1)+(2−1)×(5−1) , i.e., 16, free parameters in the model (to be compared with the 20 counts in the contingency table). Given the dimension of the simulated parameter, it is impossible to provide a complete picture of the convergence properties of the algorithm, and we represented in Fig. 4.8 the traces and histograms for the marginal posterior distributions of the parameters βi based on 10,000 iterations using a scale factor equal to τ 2 = 0.5. (This value was obtained by trial and error, producing a smooth trace for all parameters. Larger values of τ required a larger number of iterations since the acceptance rate was lower, as the reader can check using the BCoRe package.) Note that some of the traces represented in Fig. 4.8 show periodic patterns that indicate that more iterations could be necessary. However, the corresponding histograms remain

132

4 Generalized Linear Models

Fig. 4.8. Dataset airquality: Traces (top) and histograms (bottom) of the simulations from the posterior distributions of the components of β using a ﬂat prior and a random walk Metropolis–Hastings algorithm with scale factor τ 2 = 0.5 (same order row-wise as in Table 4.1)

4.5

Log-Linear Models

133

Table 4.1. Dataset airquality: Bayes estimates of the parameter β using a random walk MH algorithm with scale factor τ 2 = 0.5 Eﬀect λ λu2 λv2 λw 2 λw 3 λw 4 λw 5 λuv 22 λuw 22 λuw 23 λuw 24 λuw 25 λvw 22 λvw 23 λvw 24 λvw 25

Post. mean 2.8041 –1.0684 –5.8652 –1.4401 –2.7178 –1.1031 –0.0036 3.3559 -1.6242 –0.3456 –0.2473 –1.3335 4.5493 6.8479 4.6557 3.9558

Post. var. 0.0612 0.2176 1.7141 0.2735 0.7915 0.2295 0.1127 0.4490 1.2869 0.8432 0.6658 0.7115 2.1997 2.5881 1.7201 1.7128

quite stable over iterations. Both the approximated posterior means and the posterior variances for the 16 parameters as deduced from the MCMC run are given in Table 4.1. A few histograms in Fig. 4.8 are centered at 0, signaling a potential lack of signiﬁcance for the corresponding βi ’s. 4.5.3 Model Choice and Significance of the Parameters If we try to compare diﬀerent levels of association (or interaction), or if we simply want to test the signiﬁcance of some parameters βi , the ﬂat prior is once again inappropriate. The G-prior alternative proposed for the probit and logit models is still available, though, and we can thus replace the posterior distribution of the previous section with −k/2 π −k/2 π(β|y) ∝ |XT X|1/2 Γ (k/2) βT (XT X)β ⎧ ⎫ T n n ⎨ ⎬ exp yi xi β− exp(xiT β) ⎩ ⎭ i=1

as an alternative posterior.

i=1

(4.9)

134

4 Generalized Linear Models

Table 4.2. Dataset airquality: Metropolis–Hastings approximations of the posterior means under the G-prior Eﬀect λ λu2 λv2 λw 2 λw 3 λw 4 λw 5 λuv 22 λuw 22 λuw 23 λuw 24 λuw 25 λvw 22 λvw 23 λvw 24 λvw 25

Post. mean 2.7202 –1.1237 –4.5393 –1.4245 –2.5970 –1.1373 0.0359 2.8902 –0.9385 0.1942 0.0589 –1.0534 3.2351 5.3978 3.5831 2.8051

Post. var. 0.0603 0.1981 0.9336 0.3164 0.5596 0.2301 0.1166 0.3221 0.8804 0.6055 0.5345 0.5220 1.3664 1.3506 1.0452 1.0061

For airquality and the same model as in the previous analysis, namely the maximum nonsaturated model with 16 parameters, Algorithm 4.7 can be used with (4.9) as target and τ 2 = 0.5 as the scale in the random walk. The result of this simulation over 10,000 iterations is presented in Fig. 4.9. The traces of the components of β show the same slow mixing as in Fig. 4.8, with similar occurrences of large deviances from the mean value that may indicate the weak identiﬁability of some of these parameters. Note also that the histograms of the posterior marginal distributions are rather close to those associated with the ﬂat prior, as shown in Fig. 4.8. The MCMC approximations to the posterior means and the posterior variances are given in Table 4.2 for all 16 parameters, based on the last 9,000 iterations. While the ﬁrst parameters are quite close to those provided by Table 4.1, the estimates of the interaction coeﬃcients vary much more and are associated with much larger variances. This indicates that much less information is available within the contingency table about interactions, as can be expected. If we now consider the very reason why this alternative to the ﬂat prior was introduced, we are facing the same diﬃculty as in the probit case for the computation of the marginal density of y. And, once again, the same solution applies: using an importance sampling experiment to approximate the integral works when the importance function is a multivariate normal (or t) distribution with mean (approximately) E[β|y] and covariance matrix (approximately) 2 × V(β|y) using the Metropolis–Hastings approximations reported in Table 4.2. We can therefore approximate Bayes factors for testing all possible structures of the log-linear model.

4.5

Log-Linear Models

135

Fig. 4.9. Dataset airquality: Same legend as Fig. 4.8 for the posterior distribution (4.9) as target

136

4 Generalized Linear Models

For airquality, we illustrate this ability by testing the presence of twoby-two interactions between the three variables. We thus compare the largest non-saturated model with each submodel where one interaction is removed. An ANOVA-like output is Effect log10(BF) u:v u:w v:w

6.0983 (****) -0.5732 6.0802 (****)

evidence against H0: (****) decisive, (***) strong, (**) substantial, (*) poor which means that the interaction between u and w (that is, ozone and month) is too small to be signiﬁcant given all the other eﬀects. (Note that it would be excessive to derive from this lack of signiﬁcance a conclusion of independence between u and w because this interaction is conditional on all other interactions in the complete nonsaturated model.) The above was obtained by the following R code: ﬁrst we simulated an importance sample towards approximating the full model integrated likelihood mklog=apply(noinfloglin,2,mean) vklog=var(noinfloglin) simk=rmnorm(100000,mklog,2*vklog) usk=loglinnoinflpost(simk,counts,X)dmnorm(simk,mklog,2*vklog,log=T) then reproduced this computation for the three corresponding submodels, namely noinfloglin1=hmnoinfloglin(10^4,counts,X[,-8],0.5) mk1=apply(noinfloglin1,2,mean) vk1=var(noinfloglin1) simk1=rmnorm(100000,mk1,2*vk1) usk1=loglinnoinflpost(simk1,counts,X[,-8])dmnorm(simk1,mk1,2*vk1,log=T) bf1loglin=mean(exp(usk))/mean(exp(usk1)) and the same pattern with noinfloglin2=hmnoinfloglin(10^4,counts,cbind(X[,-(9:12)],0.5) and noinfloglin3=hmnoinfloglin(10^4,counts,X[,1:12],0.5)

4.6

Exercises

137

4.6 Exercises n 4.1 Show that, for the logistic regression model, the statistic i=1 yi xi is sufﬁcient when conditioning on the xi ’s (1 ≤ i ≤ n), and give the corresponding family of conjugate priors. 4.2 Show that the logarithmic link is the canonical link function in the case of the Poisson regression model. 4.3 Suppose y1 , . . . , yk are independent Poisson P(μi ) random variables. Show k that, conditional on n = i=1 yi , y = (y1 , . . . , yk ) ∼ Mk (n; α1 , . . . , αk ) , and determine the αi ’s. 4.4 For π the density of an inverse normal distribution with parameters θ1 = 3/2 and θ2 = 2, π(x) ∝ x−3/2 exp(−3/2x − 2/x)Ix>0 , write down and implement an independence MH sampler * with a Gamma proposal with parameters (α, β) = (4/3, 1) and (α, β) = (0.5 4/3, 0.5). 4.5 Consider x1 , x2 , and x3 iid C (θ, 1), and π(θ) ∝ exp(−θ2 /100). Show that the posterior distribution of θ, π(θ|x1 , x2 , x3 ), is proportional to exp(−θ2 /100)[(1 + (θ − x1 )2 )(1 + (θ − x2 )2 )(1 + (θ − x3 )2 )]−1 and that it is trimodal when x1 = 0, x2 = 5, and x3 = 9. Using a random walk based on the Cauchy distribution C (0, σ 2 ), estimate the posterior mean of θ using diﬀerent values of σ 2 . In each case, monitor the convergence. 4.6 Estimate the mean of a G a(4.3, 6.2) random variable using 1. direct sampling from the distribution via the R command > x=rgamma(n,4.3,scale=6.2) 2. Metropolis–Hastings with a G a(4, 7) proposal distribution; 3. Metropolis–Hastings with a G a(5, 6) proposal distribution. In each case, monitor the convergence of the cumulated average. 4.7 For a standard normal distribution as target, implement a Hastings– Metropolis algorithm with a mixture of ﬁve random walks with variances σ = 0.01, 0.1, 1, 10, 100 and equal weights. Compare its output with the output of Fig. 4.3. 4.8 For the probit model under ﬂat prior, ﬁnd conditions on the observed pairs (xi , yi ) for the posterior distribution above to be proper. 4.9 For the probit model under non-informative prior, ﬁnd conditions on i yi and i (1 − yi ) for the posterior distribution deﬁned by (4.4) to be proper.

138

4 Generalized Linear Models

4.10 Include an intercept in the probit analysis of bank and run the corresponding version of Algorithm 4.7 to discuss whether or not the posterior variance of the intercept is high. 4.11 Using variable representation of the probit model, introduce the latent zi |β ∼ N xiT β, 1 (1 ≤ i ≤ n) such that yi = Izi ≤0 . Deduce that N+ xiT β, 1, 0 if yi = 1 , zi |yi , β ∼ N− xiT β, 1, 0 if yi = 0 , where N+ (μ, 1, 0) and N− (μ, 1, 0) are the normal distributions with mean μ and variance 1 that are left-truncated and right-truncated at 0, respectively. Check that those distributions can be simulated using the R commands > xp=qnorm(runif(1)*pnorm(mu)+pnorm(-mu))+mu > xm=qnorm(runif(1)*pnorm(-mu))+mu Under the ﬂat prior π(β) ∝ 1, show that β|y, z ∼ Nk (XT X)−1 XT z, (XT X)−1 , where z = (z1 , . . . , zn ), and derive the corresponding Gibbs sampler, sometimes called the Albert–Chib sampler. (Hint: A good starting point is the maximum likelihood estimate of β.) Compare the application to bank with the output in Fig. 4.4. (Note: Account for diﬀerences in computing time.) 4.12 For the bank dataset and the probit model, compute the Bayes factor associated with the null hypothesis H0 : β2 = β3 = 0.

˜ iT β {1 + exp x ˜ iT β} 4.13 In the case of the logit model—i.e., when pi = exp x (1 ≤ i ≤ k)—derive the prior distribution on β associated with the prior (4.6) on (p1 , . . . , pk ). 4.14 Examine whether or not the suﬃcient conditions for propriety of the posterior distribution found in Exercise 4.9 for the probit model are the same for the logit model. 4.15 For the bank dataset and the logit model, compute the Bayes factor associated with the null hypothesis H0 : β2 = β3 = 0 and compare its value with the value obtained for the probit model in Exercise 4.12. 4.16 Given a contingency table with four categorical variables, determine the number of submodels to consider. 4.17 In the case of a 2 × 2 contingency table with ﬁxed total count n = n11 + n12 + n21 + n22 , we denote by θ11 , θ12 , θ21 , θ22 the corresponding probabilities. If the prior on those probabilities is a Dirichlet D4 (1/2, . . . , 1/2), give the corresponding marginal distributions of α = θ11 + θ12 and β = θ11 + θ21 . Deduce the associated Bayes factor if H0 is the hypothesis of independence between the factors and if the priors on the margin probabilities α and β are those derived above.

5 Capture–Recapture Experiments

He still couldn’t be sure that he hadn’t landed in a trap. —Ian Rankin, Resurrection Men.—

Roadmap This chapter deals with a very special case of survey models. Surveys are used in many settings to evaluate some features of a given population, including its main characteristic, the size of the population. In the case of capture–recapture surveys, individuals are observed and identiﬁed either once or several times and the repeated observations can be used to draw inference on the population size and its dynamic characteristics. Along with the original model, we will also introduce extensions that can be seen as a ﬁrst entry into hidden Markov chain models, detailed further in Chap. 6. In particular, we cover the generic Arnason–Schwarz model that is customarily used by biologists for open populations. On the methodological side, we provide an introduction to the accept–reject method, which is the central simulation technique behind most standard random generators and relates to the Metropolis–Hastings methodology in many ways.

J.-M. Marin and C.P. Robert, Bayesian Essentials with R, Springer Texts in Statistics, DOI 10.1007/978-1-4614-8687-9 5, © Springer Science+Business Media New York 2014

139

140

5 Capture–Recapture Experiments

5.1 Inference in a Finite Population In this chapter, we consider the problem of estimating the unknown size, N , of a population, based on a survey; that is, on a partial observation of this population. To be able to evaluate a population size without going through the enumeration of all its members is obviously very appealing, both timewise and moneywise, especially when sampling those members has a perturbing eﬀect on them.1 A primary type of survey (which we do not study in this chapter) is based on knowledge of the structure of the population. For instance, in a political survey about voting intentions, we build a sample of 1,000 individuals, say, such that the main sociological groups (farmers, civil servants, senior citizens, etc.) are represented in proportion in the sample. In that situation, there is no statistical inference, so to speak, except about the variability of the responses, which are in the simplest cases binomial variables. Obviously, such surveys require primary knowledge of the population, which can be obtained either by a (costly) census, like those that states run every 5 or 10 years, or by a preliminary exploratory survey that aims at uncovering these hidden structures. This secondary type of survey is the purpose of this chapter, under the name of capture–recapture (or capture–mark–recapture) experiments, where a few individuals sampled at random from the population of interest bring some information about the characteristics of this population and in particular about its size. The capture–recapture models were ﬁrst used in biology and ecology to estimate the size of animal populations, such as herds of caribous (e.g., for culling) or of whales (e.g., for the International Whaling Commission to determine ﬁshing quotas), cod populations, and the number of diﬀerent species in a particular area. While our illustrative dataset will be related to a biological problem, we stress that these capture–recapture models apply in a much wider range of domains, such as, for instance, – sociology and demography, where the estimation of the size of populations at risk is always delicate (e.g., homeless people, prostitutes, illegal migrants, drug addicts, etc.); – oﬃcial statistics for reducing the cost of a census2 or improving its eﬃciency on delicate or rare subcategories (as in the U.S. census undercount procedure and the new French census); – ﬁnance (e.g., in credit scoring, defaulting companies, etc.) and marketing (consumer habits, telemarketing, etc.); 1 In the most extreme cases, sampling an individual may lead to its destruction, as for instance in forestry when estimating the volume of trees or in meat production when estimating the content of fat in meat. 2 Even though a census is formally a deterministic process since it aims at the complete enumeration of a given population, it inevitably involves many random components at the selection, collection, and processing levels (S¨ arndal et al., 2003).

5.1

Inference in a Finite Population

141

– fraud detection (e.g., phone, credit card, etc.) and document authentication (historical documents, forgery, etc.); and – software debugging, to determine an evaluation of the number of bugs in a computer program. In these diﬀerent examples, the size N of the whole population is unknown but samples (with ﬁxed or random sizes) can easily be extracted from the population. For instance, in a computer program, the total number N of bugs is unknown but one can record the number n1 of bugs detected in a given perusal. Similarly, the total number N of homeless people in a city like Philadelphia at a given time is not known but it is possible to count the number n1 of homeless persons in a given shelter on a precise night, to record their ID, and to cross this sample with a sample of n2 persons collected the night after in order to detect how many persons n12 were present in the shelter on both nights. The dataset we consider throughout this chapter is called eurodip and is related to a population of birds called European dippers (Cinclus cinclus). These birds are closely dependent on streams, feeding on underwater invertebrates, and their nests are always close to water. The capture–recapture data on the European dipper contained in eurodip covers 7 years (1981–1987 inclusive) of observations in a zone of 200 km2 in eastern France. The data consist of markings and recaptures of breeding adults each year during the breeding period from early March to early June. Birds were at least 1 year old when initially banded. In eurodip, each row of seven digits corresponds to a capture–recapture story for a given dipper, 0 indicating an absence of capture that year and, in the case of a capture, 1, 2, or 3 representing the zone where the dipper is captured. For instance, the three lines from eurodip 1 0 0 0 0 0 0 1 3 0 0 0 0 0 0 2 2 2 1 2 2 indicate that the ﬁrst dipper was only captured the ﬁrst year in zone 1 and that the second dipper was captured in years 1981 and 1982 and moved from zone 1 to zone 3 between those years. The third dipper was captured every year but 1981 and moved between zones 1 and 2 during the remaining year. In conclusion, we hope that the introduction above was motivating enough to convince the reader that population sampling models are deeply relevant in statistical practice. Besides, these models also provide an interesting application of Bayesian modeling and in particular they allow for the inclusion of often available prior information.

142

5 Capture–Recapture Experiments

5.2 Sampling Models 5.2.1 The Binomial Capture Model We start with the simplest model of all, namely the independent observation or capture3 of n+ individuals from a population of size N . For instance, a trap is positioned on a rabbit track for ﬁve hours and n+ rabbits are found in the trap. While the population size N ∈ N is the parameter of interest, there exists a nuisance parameter, namely the probability p ∈ [0, 1] with which each individual is captured. (This model assumes that catching the ith individual is independent of catching the jth individual.) For this model, n+ ∼ B(N, p) and the corresponding likelihood is + + N pn (1 − p)N −n IN ≥n+ . (N, p|n+ ) = n+ Obviously, with a single observation n+ , we cannot say much on (N, p), but the posterior distribution is still well-deﬁned. For instance, if we use the vague prior π(N, p) ∝ N −1 IN (N )I[0,1] (p) , the posterior distribution of N is N! N −1 IN ≥n+ IN∗ (N ) (N − n+ )!n+ ! (N − 1)! (N − n+ )! ∝ IN ≥n+ ∨1 (N − n+ )! (N + 1)! 1 + I = , N (N + 1) N ≥n ∨1

1

π(N |n+ ) ∝

+

+

pn (1 − p)N −n dp

0

(5.1)

where n+ ∨ 1 = max(n+ , 1). Note that this posterior distribution is deﬁned even when n+ = 0. If we use the (more informative) uniform prior π(N, p) ∝ I{1,...,S} (N )I[0,1] (p) , the posterior distribution of N is π(N |n+ ) ∝

1 I + (N ) . N + 1 {n ∨1,...,S}

3 We use the original terminology of capture and individuals, even though the sampling mechanism may be far from genuine capture, as in whale sightseeing or software bug detection.

5.2

Sampling Models

143

For illustrative purposes, consider the case of year 1981 in eurodip (which is the ﬁrst column in the ﬁle): > data(eurodip) > year81=eurodip[,1] > nplus=sum(year81>0) [1] 22 where n+ = 22 dippers were thus captured. By using the binomial capture model and the vague prior π(N, p) ∝ N −1 , the number of dippers N can be estimated by the posterior median. (Note that the mean of (5.1) does not exist, no matter what n+ is.) > N=max(nplus,1) > rangd=N:(10^4*N) > post=1/(rangd*(rangd+1)) > 1/sum(post) [1] 22.0022 > post=post/sum(post) > min(rangd[cumsum(post)>.5]) [1] 43 For this year 1981, the estimate of N is therefore 43 dippers. (See Exercise 5.1 for theoretical justiﬁcations as to why the sum of the probabilities is equal to n+ and why the median is exactly 2n+ −1.) If we use the ecological information that there cannot be more than 400 dippers in this region, we can take the prior π(N, p) ∝ I{1,...,400} (N )I[0,1] (p) and estimate the number of dippers N by its posterior expectation: > pbino=function(nplus){ + prob=c(rep(0,max(nplus,1)-1),1/(max(nplus,1):400+1)) + prob/sum(prob) + } > sum((1:400)*pbino(nplus)) [1] 130.5237

5.2.2 The Two-Stage Capture–Recapture Model A logical extension to the capture model above is the capture–mark–recapture model, which considers two capture periods plus a marking stage, as follows: 1. n1 individuals from a population of size N are “captured”, that is, sampled without replacement. 2. Those individuals are “marked”, that is, identiﬁed by a numbered tag (for birds and ﬁshes), a collar (for mammals), or another device (like the Social Security number for homeless people or a picture for whales), and they are then released into the population.

144

5 Capture–Recapture Experiments

3. A second and similar sampling (once again without replacement) is conducted, with n2 individuals captured. 4. m2 individuals out of the n2 ’s bear the identiﬁcation mark and are thus characterized as having been captured in both experiments. If we assume a closed population (that is, a ﬁxed population size N throughout the capture experiment), a constant capture probability p for all individuals, and complete independence between individuals and between captures, we end up with a product of binomial models, n1 ∼ B(N, p) ,

m2 |n1 ∼ B(n1 , p) ,

and n2 − m2 |n1 , m2 ∼ B(N − n1 , p) . If nc = n1 + n2

and n+ = n1 + (n2 − m2 )

denote the total number of captures over both periods and the total number of captured individuals, respectively, the corresponding likelihood (N, p|n1 , n2 , m2 ) is N − n1 n2 −m2 (1 − p)N −n1 −n2 +m2 I{0,...,N −n1 } (n2 − m2 ) p n 2 − m2 n 1 m2 n1 −m2 N p (1 − p) pn1 (1 − p)N −n1 I{0,...,N } (n1 ) × n1 m2 N! ∝ pn1 +n2 (1 − p)2N −n1 −n2 IN ≥n+ (N − n1 − n2 + m2 )! c c N ∝ pn (1 − p)2N −n IN ≥n+ , n+ which shows that (nc , n+ ) is a suﬃcient statistic. If we choose the prior π(N, p) = π(N )π(p) such that π(p) is a U ([0, 1]) density, the conditional posterior distribution on p is such that c

c

π(p|N, n1 , n2 , m2 ) = π(p|N, nc ) ∝ pn (1 − p)2N −n ; that is, p|N, nc ∼ Be(nc + 1, 2N − nc + 1). Unfortunately, the marginal posterior distribution of N is more complicated. For instance, if π(N ) = IN∗ (N ), it satisﬁes N c + π(N |n1 , n2 , m2 ) = π(N |n , n ) ∝ B(nc + 1, 2N − nc + 1)IN ≥n+ ∨1 , n+

5.2

Sampling Models

145

where B(a, b) denotes the beta function This distribution is called a beta-Pascal distribution, but it is not very tractable. The same diﬃculty occurs if π(N ) = N −1 IN∗ (N ). The intractability in the posterior distribution π(N |n1 , n2 , m2 ) is due to the inﬁnite summation resulting from the unbounded support of N . A feasible approximation is to replace the missing normalizing factor by a ﬁnite sum with a large enough bound on N , the bound being determined by a lack of perceivable impact on the sum. But the approximation errors due to the computations of terms such as nN+ or B(nc + 1, 2N − nc + 1) can become a serious problem when n+ is large. However, > prob=lchoose((471570:10^7),471570)+lgamma(2*(471570:10^7)+ 582681+1)-lgamma(2*(471570:10^7)+2) > range(prob) [1] -7886469 -7659979 shows that relatively large populations are manageable. If we have information about an upper bound S on N and use the corresponding uniform prior, π(N ) ∝ I{1,...,S} (N ) , the posterior distribution of N is thus proportional to N Γ (2N − nc + 1) I{n+ ∨1,...,S} (N ) , π(N |n+ ) ∝ n+ Γ (2N + 2) and, in this case, it is possible to calculate the posterior expectation of N with no approximation error. For the ﬁrst 2 years of the eurodip experiment, which correspond to the ﬁrst two columns and the ﬁrst 70 rows of the dataset, n1 = 22, n2 = 60, and m2 = 11. Hence, nc = 82 and n+ = 71. Therefore, within the frame of the two-stage capture–recapture model4 and the uniform prior U ({1, . . . , 400}) × U ([0, 1]) on (N, p), the posterior expectation of N is derived as follows: > > > > > > + +

n1=sum(eurodip[,1]>0) n2=sum(eurodip[,2]>0) m2=sum((eurodip[,1]>0) & (eurodip[,2]>0)) nc=n1+n2 nplus=nc-m2 pcapture=function(T,nplus,nc){ #T is the number of capture episodes lprob=lchoose(max(nplus,1):400,nplus)+ lgamma(T*max(nplus,1):400-nc+1)+ lgamma(T*max(nplus,1):400+2) + prob=c(rep(0,max(nplus,1)-1),exp(lprob-max(lprob))) 4

This analysis is based on the assumption that all birds captured in the second year were already present in the population during the ﬁrst year.

146

5 Capture–Recapture Experiments

+ prob/sum(prob) + } > sum((1:400)*pcapture(2,nplus,nc)) [1] 165.2637 A simpler model used in capture–recapture settings is the hypergeometric model, also called the Darroch model. This model can be seen as a conditional version of the two-stage model when conditioning on both sample sizes n1 and n2 since (see Exercise 5.3) m2 |n1 , n2 ∼ H (N, n2 , n1 /N ) ,

(5.2)

the hypergeometric distribution If we choose the uniform prior U ({1, . . . , 400}) on N , the posterior distribution of N is thus N − n1 N π(N |m2 ) ∝ I{n+ ∨1,...,400} (N ) , n2 n 2 − m2 and posterior expectations can be computed numerically by simple summations. For the ﬁrst 2 years of the eurodip dataset and S = 400, the posterior distribution of N for the Darroch model is given by

π(N |m2 ) ∝ (n − n1 )!(N − n2 )! {(n − n1 − n2 + m2 )!N !} I{71,...,400} (N ) , the normalization factor being the inverse of 400

(k − n1 )!(k − n2 )! {(k − n1 − n2 + m2 )!k!} .

k=71

We thus have a closed-form posterior distribution and the posterior expectation of N is given by pdarroch=function(n1,n2,m2){ prob=c(rep(0,max(n1+n2-m2,1)-1), choose(n1,m2)*choose(max((n1+n2-m2),1):400-n1,n2-m2)/ choose(max((n1+n2-m2),1):400,n2)) prob/sum(prob) } > sum((1:400)*pdarroch(n1,n2,m2)) [1] 137.5962

5.2

Sampling Models

147

Table~5.1 shows the evolution of this posterior expectation for different values of m_2, obtained by > for (i in 6:16) print(round(sum(pdarroch(n1,n2,i)*1:400))) [1] 277 [1] 252 [1] 224 [1] 197 [1] 172 [1] 152 [1] 135 [1] 122 [1] 111 [1] 101 [1] 94 The number of recaptures is thus highly inﬂuential on the estimate of N . In parallel, Table 5.2 shows the evolution of the posterior expectation for diﬀerent values of S (taken equal to 400 in the above). When S is large enough, say larger than S = 250, the estimate of N is quite stable, as expected.

Table 5.1. Dataset eurodip: Rounded posterior expectation of the dipper population size, N , under a uniform prior U ({1, . . . , 400}) 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 m2 Eπ [N |m2 ] 355 349 340 329 316 299 277 252 224 197 172 152 135 122 110 101

Table 5.2. Dataset eurodip: Rounded posterior expectation of the dipper population size, N , under a uniform prior U ({1, . . . , S}), for m2 = 11 S 100 150 200 250 300 350 400 450 500 Eπ [N |m2 ] 95 125 141 148 151 151 152 152 152

Leaving the Darroch model and getting back to the two-stage capture model with probability p of capture, the posterior distribution of (N, p) associated with the noninformative prior π(N, p) = 1/N is proportional to c (N − 1)! nc p (1 − p)2N −n . + (N − n )!

Thus, if n+ > 0, both conditional posterior distributions are standard distributions since

148

5 Capture–Recapture Experiments

p|nc , N ∼ Be(nc + 1, 2N − nc + 1) N − n+ |n+ , p ∼ N eg(n+ , 1 − (1 − p)2 ) , the latter being a negative binomial distribution. Indeed, as a function of N , + + N −1 (N − 1)! 2N −nc 2 N −n 2 n (1 − p) (1 − p) 1 − (1 − p) ∝ . N − n+ (N − n+ )! Therefore, while the marginal posterior in N is diﬃcult to manage, the joint distribution of (N, p) can be approximated by a Gibbs sampler, as follows: Algorithm 5.8 Two-stage Capture–Recapture Gibbs Sampler Initialization: Generate p(0) ∼ U ([0, 1]). Iteration i (i ≥ 1): 1. Generate N (i) − n+ ∼ N eg(n+ , 1 − (1 − p(i−1) )2 ). 2. Generate p(i) ∼ Be(nc + 1, 2N (i) − nc + 1).

5.2.3 The T -Stage Capture–Recapture Model A further extension to the two-stage capture–recapture model is to consider instead a series of T consecutive captures. In that case, if we denote by nt the number of individuals captured at period t (1 ≤ t ≤ T ) and by mt the number of recaptured individuals (with the convention that m1 = 0), under the same assumptions as in the two-stage model, then n1 ∼ B(N, p) and, conditionally on the j − 1 previous captures and recaptures (2 ≤ j ≤ T ), j−1 j−1 mj ∼ B (nt − mt ), p and nj − mj ∼ B N − (nt − mt ), p . t=1

t=1

The likelihood (N, p|n1 , n2 , m2 . . . , nT , mT ) is thus 9 j−1 T N − t=1 (nt − mt ) N n1 N −n1 p (1 − p) pnj −mj +mj n1 − m n j j j=2 : j−1 j−1 (n − m ) t t N − j−1 (n −m ) (n −m )−m t=1 t t t t j t=1 × (1 − p) (1 − p) t=1 mj ∝

c c N! pn (1 − p)T N −n IN ≥n+ (N − n+ )!

if we denote the suﬃcient statistics as n+ =

T t=1

(nt − mt ) and nc =

T t=1

nt ,

5.2

Sampling Models

149

the total numbers of captured individuals and captures over the T periods, respectively. For a noninformative prior such as π(N, p) = 1/N , the joint posterior satisﬁes π(N, p|n+ , nc ) ∝

c (N − 1)! nc p (1 − p)T N −n IN ≥n+ ∨1 . (N − n+ )!

Therefore, the conditional posterior distribution of p is p|N, n+ , nc ∼ Be(nc + 1, T N − nc + 1) and the marginal posterior distribution of N π(N |n+ , nc ) ∝

(N − 1)! (T N − nc )! + I , (N − n+ )! (T N + 1)! N ≥n ∨1

is computable. Note that the normalization coeﬃcient can also be approximated by summation with an arbitrary precision unless N and n+ are very large. For the uniform prior U ({1, . . . , S}) on N and U ([0, 1]) on p, the posterior distribution of N is then proportional to N (T N − nc )! + I{n+ ∨1,...,S} (N ). π(N |n ) ∝ n+ (T N + 1)! For the whole set of observations in eurodip, we have T = 7, n+ = 294, and nc = 519. Under the uniform prior with S = 400, the posterior expectation of N is given by > sum((1:400)*pcapture(7,294,519)) [1] 372.7384 While this value seems dangerously close to the upper bound of 400 on N and thus leads us to suspect a strong inﬂuence of the upper bound S, the computation of the posterior expectation for S = 2500 > S=2500;T=7;nplus=294;nc=519 > lprob=lchoose(max(nplus,1):S,nplus)+ + lgamma(T*max(nplus,1):S-nc+1)-lgamma(T*max(nplus,1):S+2) > prob=c(rep(0,max(nplus,1)-1),exp(lprob-max(lprob))) > sum((1:S)*prob)/sum(prob) [1] 373.9939 leads to 373.99, which shows the limited impact of this hyperparameter S. Using even a slightly more advanced sampling model may lead to genuine computational diﬃculties. For instance, consider a heterogeneous capture–recapture model where the individuals are captured at time 1 ≤ t ≤ T with

150

5 Capture–Recapture Experiments

probability pt and where both the size N of the population and the probabilities pt are unknown. The corresponding likelihood is T N! (N, p1 , . . . , pT |n1 , n2 , m2 . . . , nT , mT ) ∝ pnt (1 − pt )N −nt . (N − n+ )! t=1 t

If the associated prior on (N, p1 , . . . , pT ) is such that N ∼ P(λ) and (1 ≤ t ≤ T ),

αt = log

pt 1 − pt

∼ N (μt , σ 2 ),

where both σ 2 and the μt ’s are known,5 the posterior distribution satisﬁes T

π(α1 , . . . , αT , N |, n1 , . . . , nT ) ∝

λN N! (1 + eαt )−N (5.3) (N − n+ )! N ! t=1 T 1 × exp αt nt − 2 (αt − μt )2 . 2σ t=1

It is thus much less manageable from a computational point of view, especially when there are many capture episodes. A corresponding Gibbs sampler could simulate easily from the conditional posterior distribution on N since T + + αt N − n |α, n ∼ P λ (1 + e ) , t=1

but the conditionals on the αt ’s (1 ≤ t ≤ T ) are less conventional, αt |N, n ∼ πt (αt |N, n) ∝ (1 + eαt )−N eαt nt −(αt −μt )

2

/2σ2

,

and they require either an accept–reject algorithm (Sect. 5.4) or a Metropolis–Hastings algorithm in order to be simulated. For the prior λN IN (N )I[0,1] (p) , π(N, p) ∝ N! the conditional posteriors are then p|N, nc ∼ Be(nc + 1, T N − nc + 1) and N − n+ |p, n+ ∼ P(λ(1 − p)T ) and a Gibbs sampler similar to the one developed in Algorithm 5.8 can easily be implemented, for instance via the code 5 This assumption can be justiﬁed on the basis that each capture probability is only observed once on the tth round (and so cannot reasonably be associated with a noninformative prior).

5.3

> > > > > > + + +

Open Populations

151

lambda=200 nsimu=10^4 p=rep(1,nsimu); N=p N[1]=2*nplus p[1]=rbeta(1,nc+1,T*lambda-nc+1) for (i in 2:nsimu){ N[i]=nplus+rpois(1,lambda*(1-p[i-1])^T) p[i]=rbeta(1,nc+1,T*N[i]-nc+1) }

For eurodip, we used this Gibbs sampler and obtained the results illustrated by Fig. 5.1. When the chain is initialized at the (unlikely) value N (0) = λ = 200 (which is the prior expectation of N ), the stabilization of the chain is quite clear: It only takes a few iterations to converge toward the proper region that supports the posterior distribution. We can thus visually conﬁrm the convergence of the algorithm and approximate the Bayes estimators of N and p by the Monte Carlo averages > mean(N) [1] 326.9831 > mean(p) [1] 0.2271828 The precision of these estimates can be assessed as in a regular Monte Carlo experiment, but the variance estimate is biased because of the correlation between the simulations. A simple way to assess this eﬀect is to call R function acf() for each component θi of the parameter, as ν =1+2

∞

(1)

(t+1)

cor(θi , θi

)

t=1

evaluates the loss of eﬃciency due to the correlation. The corresponding effective sample size, given by Tess = T /ν, provides the equivalent size of an iid sample. For instance, > 1/(1+2*sum(acf(N)$acf[-1])) [1] 0.599199 > 1/(1+2*sum(acf(p)$acf[-1])) [1] 0.6063236 shows that the current Gibbs sampler oﬀers an eﬃciency of 60% compared with an iid sample from the posterior distribution.

152

5 Capture–Recapture Experiments 360 0.26 340 p

N

0.24 0.22

320 0.20 0.18

300 0

2000 4000 6000 8000

0

2000 4000 6000 8000

40

0.06

Density

Density

30 0.04

0.02

20 10 0

0.00 300 310 320 330 340 350 360 N

0.18

0.20

0.22

0.24

0.26

p

Fig. 5.1. Dataset eurodip: Representation of the Gibbs sampling output for the parameters p (ﬁrst column) and N (second column)

5.3 Open Populations Moving towards more realistic settings, we now consider the case of an open population model, where the population size does not remain ﬁxed over the experiment but, on the contrary, there is a probability q for each individual to leave the population at each time (or, more accurately, between any two capture episodes). Given that the associated likelihood involves unobserved indicators (namely, indicators of survival; see Exercise 5.14), we study here a simpler model where only the individuals captured during the ﬁrst capture experiment are marked and subsequent recaptures are registered. For three successive capture experiments, we thus have n1 ∼ B(N, p) ,

r1 |n1 ∼ B(n1 , q) ,

r2 |n1 , r1 ∼ B(n1 − r1 , q) ,

for the distributions of the ﬁrst capture population size and of the numbers of individuals who vanished between the ﬁrst and second, and the second and third experiments, respectively, and c2 |n1 , r1 ∼ B(n1 − r1 , p),

c3 |n1 , r1 , r2 ∼ B(n1 − r1 − r2 , p) ,

5.3

Open Populations

153

for the number of recaptured individuals during the second and the third experiments, respectively. Here, only n1 , c2 , and c3 are observed. The numbers of individuals removed at stages 1 and 2, r1 and r2 , are not available and must therefore be simulated, as well as the parameters N , p, and q.6 The likelihood (N, p, q, r1 , r2 |n1 , c2 , c3 ) is given by

N n1 N −n1 n1 r1 n1 −r1 n1 − r1 p (1 − p) q (1 − q) pc2 (1 − p)n1 −r1 −c2 n1 r1 c2 n1 − r1 r2 n1 −r1 −r2 n1 − r1 − r2 × q (1 − q) pc3 (1 − p)n1 −r1 −r2 −c3 r2 c3

and, if we use the prior π(N, p, q) ∝ N −1 I[0,1] (p)I[0,1] (q), the associated conditionals are +

π(p|N, q, D ∗ ) ∝ pn (1 − p)u+ , π(q|N, p, D ∗ ) ∝ q r1 +r2 (1 − q)2n1 −2r1 −r2 , (N − 1)! (1 − p)N IN ≥n1 , (N − n1 )! (n1 − r1 )! q r1 (1 − q)−2r1 (1 − p)−2r1 , π(r1 |p, q, n1 , c2 , c3 , r2 ) ∝ r1 !(n1 − r1 − r2 − c3 )!(n1 − c2 − r1 )! q r2 [(1 − p)(1 − q)]−r2 , π(r2 |p, q, n1 , c2 , c3 , r1 ) ∝ r2 !(n1 − r1 − r2 − c3 )! π(N |p, q, D ∗ ) ∝

where D ∗ = (n1 , c2 , c3 , r1 , r2 ) and u1 = N − n1 , u2 = n1 − r1 − c2 , u3 = n1 − r1 − r2 − c3 , n+ = n 1 + c2 + c3 , u + = u 1 + u 2 + u 3 (u stands for unobserved, even though these variables can be computed conditional on the remaining unknowns). Therefore, the full conditionals are p|N, q, D ∗ ∼ Be(n+ + 1, u+ + 1) , q|N, p, D ∗ ∼ Be(r1 + r2 + 1, 2n1 − 2r1 − r2 + 1) , N − n1 |p, q, D ∗ ∼ N eg(n1 , p) , q r2 |p, q, n1 , c2 , c3 , r1 ∼ B n1 − r1 − c3 , , q + (1 − q)(1 − p) which are very easily simulated, while r1 has a less conventional distribution. However, this diﬃculty is minor since, in our case, n1 is not extremely 6 From a theoretical point of view, r1 and r2 are missing variables rather than true parameters. This obviously does not change anything either for simulation purposes or for Bayesian inference.

154

5 Capture–Recapture Experiments

large. It is thus possible to compute the probability that r1 is equal to each of the values in {0, 1, . . . , min(n1 − r2 − c3 , n1 − c2 )}. This means that the corresponding Gibbs sampler can be implemented as well. gibbscap1=function(nsimu,n1,c2,c3){ N=p=q=r1=r2=rep(0,nsimu) N[1]=round(n1/runif(1)) r1[1]=max(c2,c3)+round((n1-c2)*runif(1)) r2[1]=round((n1-r1[1]-c3)*runif(1)) nplus=n1+c2+c3 for (i in 2:nsimu){ uplus=N[i-1]-r1[i-1]-c2+n1-r1[i-1]-r2[i-1]-c3 p[i]=rbeta(1,nplus+1,uplus+1) q[i]=rbeta(1,r1[i-1]+r2[i-1]+1,2*n1-2*r1[i-1]-r2[i-1]+1) N[i]=n1+rnbinom(1,n1,p[i]) rbar=min(n1-r2[i-1]-c3,n1-c2) pq=q[i]/((1-q[i])*(1-p[i]))^2 pr=lchoose(n1-c2,0:rbar)+(0:rbar)*log(pq)+ lchoose(n1-(0:rbar),r2[i-1]+c3) r1[i]=sample(0:rbar,1,prob=exp(pr-max(pr))) r2[i]=rbinom(1,n1-r1[i]-c3,q[i]/(q[i]+(1-q[i])*(1-p[i]))) } list(N=N,p=p,q=q,r1=r1,r2=r2) } We stress that R is quite helpful in simulating from unusual distributions and in particular from those with ﬁnite support. For instance, the conditional distribution of r1 above can be simulated using the following representation of P(r1 = k|p, q, n1 , c2 , c3 , r2 ) (0 ≤ k ≤ r = min(n1 − r2 − c3 , n1 − c2 )), k n1 − k n 1 − c2 q , k (1 − q)2 (1 − p)2 r2 + c3

(5.4)

up to a normalization constant, since the binomial coeﬃcients and the power in k can be computed for all values of k at once, thanks to the matrix capabilities of R, through the command lchoose. The above quantity corresponding to pr=lchoose(n=n1 − c2 ,k=0:r) + (0:r)*log(q1) + lchoose(n=n1-(0:r),k=r2 + c3 ) is the whole vector of the log-probabilities, with q1 = q/(1 − q)2 (1 − p)2 .

In

most computations, it is safer to use logarithmic transforms to reduce the risk of running into overﬂow or underﬂow error messages. For instance, in the example above, the probability vector can be recovered by pr=exp(pr-max(pr))/sum(exp(pr-max(pr)))

5.3

Open Populations

155

while a direct computation of exp(pr) may well produce an Inf value that invalidates the remaining computations.7

Once the probabilities are transformed as in the previous R code, a call to the R command > sample(0:mm,n,prob=exp(pr-max(pr))) is suﬃcient to provide n simulations of r1 . The production of a large Gibbs sample is immediate: > system.time(gibbscap1(10^5,22,11,6)) user system elapsed 12.816 0.000 12.830 Even a large value such as n1 = 1612 used below does not lead to computing diﬃculties since we can run 10,000 iterations of the corresponding Gibbs sampler in a few seconds on a laptop: > system.time(gibbscap1(10^4,1612,811,236)) user system elapsed 10.245 0.028 10.294 For eurodip, we have n1 = 22, c2 = 11, and c3 = 6. We obtain the Gibbs output > gg=gibbscap1(10^5,22,11,6) summarized in Fig. 5.2. The sequences for all components are rather stable and their mixing behavior (i.e., the speed of exploration of the support of the target) is satisfactory, even though we can still detect a trend in the ﬁrst three rows. Since r1 and r2 are integers with only a few possible values, the last two rows show apparently higher jumps than the three other parameters. The MCMC approximations to the posterior expectations of N and p are equal > mean(gg$N) [1] 57.52955 > mean(gg$p) [1] 0.3962891 respectively. Given the large diﬀerence between n1 and c2 and the proximity between c2 and c3 , high values of q are rejected, and the diﬀerence can be attributed 7

This recommendation also applies to the computation of likelihoods that tend to take absolute values that exceed the range of the computer representation of real numbers, while only the relative values are relevant for Bayesian computations. Using a transform such as exp(loglike-max(loglike)) thus helps in reducing the risk of overﬂows.

156

5 Capture–Recapture Experiments

with high likelihood to a poor capture rate. One should take into account the fact that there are only three observations for a model that involves three true parameters plus two missing variables. Figure 5.3 gives another insight into the posterior distribution by representing the joint distribution of the sample of (r1 , r2 )’s > plot(jitter(gg$r1,factor=1),jitter(g2$r2,factor=1),cex=0.5, + xlab=expression(r[1]),ylab=expression(r[2])) using for representation purposes the R function jitter(), which moves each point by a tiny random amount. There is a clear positive correlation between r1 and r2 , despite the fact that r2 is simulated on an (n1 − c3 − r1 ) scale. The mode of the posterior is (r1 , r2 ) = (0, 0), which means that it is likely that no dipper died or left the observation area over the 3-year period.

5.4 Accept–Reject Algorithms In Chap. 2, we mentioned standard random number generators used for the most common distributions and presented importance sampling (Algorithm 2.2) as a possible alternative when such generators are not available. While MCMC algorithms always oﬀer a solution when facing nonstandard distributions, there often exists a possibility that is in fact used in most of the standard random generators and which we now present. It also relates to the independent Metropolis–Hastings algorithm of Sect. 4.2.2. Given a density g that is deﬁned on an arbitrary space (of any dimension), a fundamental identity is that simulating X distributed from g(x) is completely equivalent to simulating (X, U ) uniformly distributed on the set S = {(x, u) : 0 < u < g(x)} (this is called the Fundamental Theorem of Simulation in Robert and Casella, 2004, Chap. 3). The reason for this equivalence is simply that ∞ I0thresh(y,n1,c2,c3,r2,barr)) } echan[i]=y } echan } the average of the acceptance ratios g(k)/M f (k) is equal to 0.12. This is a relatively small value since it corresponds to a rejection rate of about 9/10. The simulation process could thus be a little slow, although > system.time(ardipper(10^5,n1=22,c2=11,c3=6,r2=1,q1=.1)) user system elapsed 8.148 0.024 8.1959 shows this is not the case. (Note that the code ardipper provided here does not produce the rejection rate. It has to be modiﬁed for this purpose.) An histogram of accepted values is shown in Fig. 5.4. Obviously, this method is not hassle-free. For complex densities g, it may prove impossible to ﬁnd a density f such that g(x) ≤ M f (x) and M is small enough. However, there exists a large class of univariate distributions for which a generic choice of f is possible (see Robert and Casella, 2004, Chap. 2).

160

5 Capture–Recapture Experiments

0.4

0.3

0.2

0.1

0.0

0

1

2

3

4

5

r1

Fig. 5.4. Dataset eurodip: Sample from the distribution (5.4) obtained by accept– reject and based on the simulation of 10,000 values from a B(n1 , q1 ) distribution for n1 = 22, c2 = 11, c3 = 6, r2 = 1, and q1 = 0.1

5.5 The Arnason–Schwarz Capture–Recapture Model We consider in this ﬁnal section a more advanced capture–recapture model based on the realistic assumption that, in most capture–recapture experiments, we can tag individuals one by one; that is, we can distinguish each individual at the time of its ﬁrst capture and thus follow its capture history. For instance, when tagging mammals and birds, diﬀerentiated tags can be used, so that there is only one individual with tag, say, 23131932.8 The Arnason–Schwarz model thus considers a capture–recapture experiment as a collection of individual histories. For each individual that has been 8 In a capture–recapture experiment used in Dupuis (1995), a population of lizards was observed in the south of France (Loz`ere). When it was found that plastic tags caused necrosis on those lizards, the biologists in charge of the experiment decided to cut a phalange of one of the ﬁngers of the captured lizards to identify them later. While the number of possibilities, 220 , is limited, it is still much larger than the number of captured lizards in this study. Whether or not the lizards appreciated this ability to classify them is not known.

5.5

The Arnason–Schwarz Capture–Recapture Model

161

captured at least once during the experiment, individual characteristics of interest are registered at each capture. For instance, this may include location, weight, sexual status, pregnancy occurrence, social status, and so on. The probabilistic modeling includes this categorical decomposition by adding what we will call movement probabilities to the survival probabilities already used in the Darroch open population model of Sect. 5.2.2. From a theoretical point of view, this is a ﬁrst example of a (partially) hidden Markov model, a structure studied in detail in Chap. 7. In addition, the model includes the possibility that individuals vanish from the population between two capture experiments. (This is thus another example of an open population model.) As in eurodip, the interest that drives the capture–recapture experiment may be to study the movements of individuals within a zone K divided into k = 3 strata denoted by 1, 2, 3. (This structure is generic: Zones are not necessarily geographic and can correspond to anything from social status, to HIV stage, to university degree.) For instance, four consecutive rows of possible eurodip (individual) capture–recapture histories look as follows: 45 0 3 0 0 46 0 2 2 2 47 0 2 0 0 48 2 1 2 1

000 211 000 000

where 0 denotes a failure to capture. This means that, for dipper number 46, the ﬁrst location was not observed but this dipper was captured for all the other experiments. For dippers number 45 and 47, there was no capture after the second time and thus one or both of them could be dead (or outside the range of the capture area) at the time of the last capture experiment. We also stress that the Arnason–Schwarz model often assumes that individuals that were not part of the population on the ﬁrst capture experiments can be identiﬁed as such.9 We thus have cohorts of individuals that entered the study in the ﬁrst year, the second year, and so on. 5.5.1 Modeling A description of the basic Arnason–Schwarz model involves two types of variables for each individual i (i = 1, . . . , n) in the population: ﬁrst, a variable that describes the location of this individual, zi = (z(i,t) , t = 1, .., τ ) , where τ is the number of capture periods; and, second, a binary variable that describes the capture history of this individual, xi = (x(i,t) , t = 1, .., τ ) . 9

This is the case, for instance, with newborns or new mothers in animal capture experiments.

162

5 Capture–Recapture Experiments

We10 assume that z(i,t) = r means the animal i is alive in stratum r at time t and that z(i,t) = † denotes the case when the animal i is dead at time t. The variable zi is sometimes called the migration process of individual i by analogy with the special case where one is considering animals moving between geographical zones, like some northern birds in spring and fall. Note that xi is entirely observed, while zi is not. For instance, we may have xi = 1 1 0 1 1 1 0 0 0 and zi = 1 2 · 3 1 1 · · · , for which a possible completed zi is zi = 1 2 1 3 1 1 2 † † , meaning that the animal died between the seventh and the eighth capture events. In particular, the Arnason–Schwarz model assumes that dead animals are never observed (although this type of assumption can easily be modiﬁed when processing the model, in what are called tag-recovery experiments). Therefore z(i,t) = † always corresponds to x(i,t) = 0. Moreover, we assume that the (xi , zi )’s (i = 1, . . . , n) are independent and that each random vector zi is a Markov chain taking values in K ∪ {†} with uniform initial probability on K (unless there is prior information to the contrary). The parameters of the Arnason–Schwarz model are thus of two kinds: the capture probabilities pt (r) = P x(i,t) = 1|z(i,t) = r on the one hand and the transition probabilities qt (r, s) = P z(i,t+1) = s|z(i,t) = r r ∈ K, s ∈ K ∪ {†},

qt (†, †) = 1

on the other hand. We derive two further sets of parameters, ϕt (r) = 1−qt (r, †) the survival probabilities and ψt (r, s) the interstrata movement probabilities, deﬁned as qt (r, s) = ϕt (r) × ψt (r, s) r ∈ K, s ∈ K . The likelihood corresponding to the complete observation of the (xi , zi )’s, (p1 , . . . , pτ , q1 , . . . , qτ |(x1 , z1 ), . . . , (xn , zn )), is then given by : 9 τ n τ −1 x(i,t) 1−x(i,t) pt (z(i,t) ) {1 − pt (z(i,t) )} × qt (z(i,t) , z(i,t+1) ) , (5.6) i=1

t=1

t=1

10 Covariates registered once or at each time will not be used here, although they could be introduced via a generalized linear model as in Chap. 4, so we abstain from adding further notations in an already dense section.

5.5

The Arnason–Schwarz Capture–Recapture Model

163

up to a constant. The complexity of the likelihood corresponding to the data actually observed is due to the fact that the zi ’s are not fully observed, hence that (5.6) would have to be summed over all possible values of the missing components of the zi ’s. This complexity can be bypassed by a simulation alternative described below in Sect. 5.5.2. The prior modeling corresponding to these parameters will depend on the information that is available about the population covered by the capture– recapture experiment. For illustration’s sake, consider the use of conjugate priors pt (r) ∼ Be(at (r), bt (r)) ,

ϕt (r) ∼ Be(αt (r), βt (r)) ,

where the hyperparameters, at (r), bt (r) and so on, depend on both time t and location r, and ψt (r) ∼ Dir(γt (r)) , a Dirichlet distribution, where ψt (r) = (ψt (r, s); s ∈ K) with ψt (r, s) = 1 , s∈K

and γt (r) = (γt (r, s); s ∈ K). The determination of these (numerous) hyperparameters is also case-dependent and varies from a noninformative modeling, where all hyperparameters are taken to be equal to 1 or 1/2, to a very informative setting where exact values of these hyperparameters can be chosen from the prior information. The following example is an illustration of the latter. Table 5.3. Prior information about the capture and survival parameters of the Arnason–Schwarz model, represented by prior expectation and prior conﬁdence interval, for a capture–recapture experiment on the migrations of lizards (source: Dupuis, 1995) Episode 2 3 4 5 6 pt Mean 0.3 0.4 0.5 0.2 0.2 95% cred. int. [0.1, 0.5] [0.2, 0.6] [0.3, 0.7] [0.05, 0.4] [0.05, 0.4] Site A B,C t = 1, 3, 5 t = 2, 4 t = 1, 3, 5 t = 2, 4 Episode ϕt (r) Mean 0.7 0.65 0.7 0.7 [0.4, 0.95] 95% cred. int. [0.4, 0.95] [0.35, 0.9] [0.4, 0.95]

Example 5.1. For the capture–recapture experiment described in Footnote 8 on the migrations of lizards between three adjacent zones, there are six capture episodes. The prior information provided by the biologists on the capture and survival probabilities, pt (which are assumed to be zone independent) and ϕt (r), is given by Table 5.3. While this may seem very artiﬁcial, this

164

5 Capture–Recapture Experiments

construction of the prior distribution actually happened that way because the biologists in charge were able to quantify their beliefs and intuitions in terms of prior expectation and prior conﬁdence interval. (The diﬀerences in the prior values on pt are due to diﬀerences in capture eﬀorts, while the diﬀerences between the group of episodes 1, 3 and 5, and the group of episodes 2 and 4 are due to the fact that the odd indices correspond to spring and the even indices to fall and mortality is higher over the winter.) Moreover, this prior information can be perfectly translated in a collection of beta priors by the R divide-and-conquer function probet=function(a,b,c,alpha){ coc=(1-c)/c pbeta(b,alpha,alpha*coc)-pbeta(a,alpha,alpha*coc) } solbeta=function(a,b,c,prec=10^(-3)){ coc=(1-c)/c detail=alpha=1 while (probet(a,b,c,alpha)prec){ alpha=max(alpha-detail,detail/10) detail=detail/10 while (probet(a,b,c,alpha) solbeta(.1,.5,.3,10^(-4)) [1] 5.45300 12.72367 then leads to the hyperparameters given in Table 5.4.

Table 5.4. Hyperparameters of the beta priors corresponding to the information contained in Table 5.3 (source: Dupuis, 1995) Episode 2 3 4 5 6 Dist. Be(5, 13) Be(8, 12) Be(12, 12) Be(3.5, 14) Be(3.5, 14) Site A B t = 2, 4 t = 1, 3, 5 t = 2, 4 Episode t = 1, 3, 5 Dist. Be(6.0, 2.5) Be(6.5, 3.5) Be(6.0, 2.5) Be(6.0, 2.5)

5.5

The Arnason–Schwarz Capture–Recapture Model

165

5.5.2 Gibbs Sampler Given the presence of missing data in the Arnason–Schwarz model, a Gibbs sampler is a natural solution to handle the complexity of the likelihood. It needs to include simulation of the missing components in the vectors zi in order to simulate the parameters from the full conditional distribution π(θ|x, z) ∝ (θ|x, z) × π(θ) , Algorithm 5.10 Arnason–Schwarz Gibbs Sampler Iteration l (l ≥ 1): 1. Parameter simulation Simulate θ(l) ∼ π(θ|z(l−1) , x) as (t = 1, . . . , τ ), (l) (l) pt (r)|x, z(l−1) ∼ Be at (r) + ut (r), bt (r) + vt (r) , ⎛ ⎞ (l) (l) (l) ϕt (r)|x, z(l−1) ∼ Be ⎝αt (r) + wt (r, j), βt (r) + wt (r, †)⎠ , j∈K

(l) ψt (r)|x, z(l−1)

(l) ∼ Dir γt (r, s) + wt (r, s); s ∈ K ,

where (l)

wt (r, s) =

n

(l)

i=1 n

(l)

i=1 n

ut (r) = vt (r) =

i=1

I(z(l−1) =r,z(l−1) (i,t)

(i,t+1)

=s)

I(x

(l−1) (i,t) =1,z(i,t) =r)

I(x

(l−1) (i,t) =0,z(i,t) =r)

,

, .

2. Missing location simulation (l) Generate the unobserved z(i,t) ’s from the full conditional distributions (l)

(l−1)

(l)

(l−1)

(l)

P(z(i,1) = s|x(i,1) , z(i,2) , θ(l) ) ∝ q1 (s, z(i,2) )(1 − p1 (s)) , (l)

(l)

(l−1)

(l)

(l)

P(z(i,t) = s|x(i,t) , z(i,t−1) , z(i,t+1) , θ(l) ) ∝ qt−1 (z(i,t−1) , s) (l−1)

(l)

×qt (s, z(i,t+1) )(1 − pt (s)) , (l)

(l)

(l)

(l)

P(z(i,τ ) = s|x(i,τ ) , z(i,τ −1) , θ(l) ) ∝ qτ −1 (z(i,τ −1) , s)(1 − pτ (s)(l) ) .

166

5 Capture–Recapture Experiments

where x and z denote the collections of the vectors of capture indicators and locations, respectively. This is thus a particular case of data augmentation, where the missing data z are simulated at each step t in order to reconstitute a complete sample (x, z(t) ) for which conjugacy applies. In the setting of the Arnason–Schwarz model, we can simulate the full conditional distributions both of the parameters and of the missing components. The Gibbs sampler is as follows: Note that simulating the missing locations in the zi ’s conditionally on the other locations and on the parameters is not a very complex task because of the good conditioning properties of these vectors (which stem from their Markovian nature). As shown in Step 2 of Algorithm 5.10, the full conditional distribution of z(i,t) only depends on the previous and next locations z(i,t−1) and z(i,t+1) (and obviously on the fact that it is not observed; that is, that x(i,t) = 0). The corresponding part of the R code is based on a latent matrix containing the current values of both the observed and missing locations: for (i in 1:n){ if (z[i,1]==0) latent[i,1]=sample(1:(m+1),1, prob=q[,latent[i,2]]*(1-c(p[s,],0))) for (t in ((2:(T-1))[z[i,-c(1.T)]==0])) latent[i,t]=sample(1:(m+1),1, prob=q[latent[i,t-1],]*q[,latent[i,t+1]]*(1-c(p[s,],0))) if (z[i,T]==0) latent[i,T]=sample(1:(m+1),1, prob=q[latent[i,T-1],]*(1-c(p[s,],0))) }

(The convoluted range for the inner loop replaces an if (z[i,t]==0).) When the number of states s ∈ K is moderate, it is straightforward to simulate from such a distribution. Take K = {1, 2}, n = 4, m = 8 and assume that, for x, we have the following histories: 1 1 · · 1 · · · 1 · 1 · 1 · 2 1 2 1 · 1 2 · · 1 1 · · 1 2 1 1 2

1 2 3 4

Assume also that all (prior) hyperparameters are taken equal to 1. Then one possible instance of a simulated z is 1 1 2 1

1 1 1 2

1 1 2 1

2 2 1 1

1 1 2 2

1 1 1 1

2 1 1 1

† 2 1 2

5.5

The Arnason–Schwarz Capture–Recapture Model

167

and it leads to the following simulation of the parameters: (l)

p4 (1)|x, z(l−1) ∼ Be(1 + 2, 1 + 0) , (l)

ϕ7 (2)|x, z(l−1) ∼ Be(1 + 0, 1 + 1) , (l)

ψ2 (1, 2)|x, z(l−1) ∼ Be(1 + 1, 1 + 2) , in the Gibbs sampler, where the hyperparameters are therefore derived from the (partly) simulated history above. Note that because there are only two possible states, the Dirichlet distribution simpliﬁes into a beta distribution. 15

0.35

10

p(1)

0.30

0.25

5

0.20 0 0

2000

4000

6000

8000

10000

0.20

0.25

0.30

0.35

0.40

80

0.99 φ(2)

0.15 100

1.00

60 0.98 40 0.97 20 0.96 0 0

2000

4000

6000

8000

10000

0.75

0.95

0.96

0.97

0.98

0.99

1.00

6

0.65 ψ(3,3)

0.94 8

4 0.55 2 0.45 0 0

2000

4000

6000

8000

10000

0.4

0.5

0.6

0.7

Fig. 5.5. Dataset eurodip: Representation of the Gibbs sampling output for some parameters of the Arnason–Schwarz model, based on 10,000 iterations, with raw plots (ﬁrst column) and histograms (second column)

168

5 Capture–Recapture Experiments

For eurodip, Lebreton et al. (1992) argue that the capture and survival rates should be constant over time. If we assume that the movement probabilities are also time independent, we are left with 3 + 3 + 3 × 2 = 12 parameters. Figure 5.5 gives the Gibbs output for the parameters p(1), ϕ(2), and ψ(3, 3) using noninformative priors with a(r) = b(r) = α(r) = β(r) = γ(r, s) = 1. The simulation of the parameters is obtained by the following piece of R code, where s is the current index of the Gibbs iteration in the R code below: for (r1 in 1:m){ for (r2 in 1:(m+1)) omega[r2]=sum(latent[,1:(T-1)]==r1 & latent[,2:T]==r2) u=sum(z!=0 & latent==r1) v=sum(z==0 & latent==r1) p[s,r1]=rbeta(1,1+u,1+v) phi[s,r1]=rbeta(1,1+sum(omega[1:m]),1+omega[m+1]) psi[r1,,s]=rdirichlet(1,rep(1,m)+omega[1:m]) } The transition probabilities qt (r, s) are then reconstructed from the survival and movement probabilities, with the special case of the m+1 column corresponding to the absorbing † state: tt=matrix(rep(phi[s,],m),m,byrow=T) q=rbind(tt*psi[,,s],rep(0,m)) q=cbind(q,1-apply(q,1,sum)) The convergence of the Gibbs sampler to the region of interest occurs very quickly, even though we can spot an approximate periodicity in the raw plots on the left-hand side. The MCMC approximations of the estimates of p(1), ϕ(2), and ψ(3, 3), the empirical mean over the last 8,000 simulations, are equal to 0.25, 0.99, and 0.61, respectively.

5.6 Exercises 5.1 Show that the posterior distribution π(N |n+ ) given by (5.1), while associated with an improper prior, is deﬁned for all values of n+ . Show that the normalization factor of (5.1) is n+ ∨ 1, and deduce that the posterior median is equal to 2(n+ ∨ 1) − 1. Discuss the relevance of this estimator and show that it corresponds to a Bayes estimate of p equal to 1/2.

5.6

Exercises

169

5.2 Under the same prior as in Sect. 5.2.1, derive the marginal posterior density of N in the case where n+ 1 ∼ B(N, p) and + + n+ 2 , . . . , nk ∼ B(n1 , p) iid

are observed (the later are in fact recaptures). Apply to the sample + + (n+ 1 , n2 , . . . , n11 ) = (32, 20, 8, 5, 1, 2, 0, 2, 1, 1, 0) ,

which describes a series of tag recoveries over 11 years. 5.3 Show that the conditional distribution of m2 conditional on both sample sizes n1 and n2 is given by (5.2) and does not depend on p. Deduce the expectation Eπ [m2 |n1 , n2 , N ]. 5.4 In order to determine the number N of buses in a town, a capture–recapture strategy goes as follows. We observe n1 = 20 buses during the ﬁrst day and keep track of their identifying numbers. Then we repeat the experiment the following day by recording the number of buses that have already been spotted on the previous day, say m2 = 5, out of the n2 = 30 buses observed the second day. For the Darroch model, give the posterior expectation of N under the prior π(N ) = 1/N . ˆ = 5.5 Show that the maximum likelihood estimator of N for the Darroch model is N n1 / (m2 /n2 ), and deduce that it is not deﬁned when m2 = 0. 5.6 Give the likelihood of the extension of Darroch’s model when the capture–recapture experiments are repeated K times with capture sizes and recapture observations nk (1 ≤ k ≤ K) and mk (2 ≤ k ≤ K), respectively. (Hint: Exhibit ﬁrst the two-dimensional suﬃcient statistic associated with this model.) 5.7 Give both conditional posterior distributions involved in Algorithm 5.8 in the case n+ = 0. 5.8 Show that, for the two-stage capture model with probability p of capture, when the prior on N is a P(λ) distribution, the conditional posterior on N − n+ is P(λ(1 − p)2). 5.9 Reproduce the analysis of eurodip summarized by Fig. 5.1 when switching the prior from π(N, p) ∝ λN /N ! to π(N, p) ∝ N −1 . 5.10 An extension of the T -stage capture–recapture model of Sect. 5.2.3 is to consider that the capture of an individual modiﬁes its probability of being captured from p to q for future recaptures. Give the likelihood (N, p, q|n1 , n2 , m2 . . . , nT , mT ). 5.11 Another extension of the two-stage capture–recapture model is to allow for mark loss.11 If we introduce q as the probability of losing the mark, r as the probability of recovering a lost mark and k as the number of recovered lost marks, give the associated likelihood (N, p, q, r|n1 , n2 , m2 , k). 11

Tags can be lost by marked animals, but the animals themselves could also be lost to recapture either by changing habitat or dying. Our current model assumes that the population is closed; that is, that there is no immigration, emigration, birth, or death within the population during the length of the study. These other kinds of extension are dealt with in Sects. 5.3 and 5.5.

170

5 Capture–Recapture Experiments

5.12 Show that the conditional distribution of r1 in the open population model of Sect. 5.3 is proportional to the product (5.4). 5.13 Show that the distribution of r2 in the open population model of Sect. 5.3 can be integrated out from the joint distribution and that this leads to the following distribution on r1 : π(r1 |p, q, n1 , c2 , c3 ) ∝

(n1 − r1 )!(n1 − r1 − c3 )! r1 !(n1 − r1 − c2 )! r1 q . × (1 − p)(1 − q)[q + (1 − p)(1 − q)]

Compare the computational cost of a Gibbs sampler based on this approach with a Gibbs sampler using the full conditionals. 5.14 Show that the likelihood associated with an open population as in Sect. 5.3 can be written as

(N, p|D ∗ ) =

N T

( it ,δit )it t=1 i=1

it q i(t−1) (1 − q i(t−1) )1− it

× p(1− it )δit (1 − p)(1− it )(1−δit ) , where q0 = q, q1 = 1, and δit and it are the capture and exit indicators, respectively. Derive the order of complexity of this likelihood; that is, the number of elementary operations necessary to compute it.12 5.15 In connection with the presentation of the accept–reject algorithm in Sect. 5.4, show that, for M > 0, if g is replaced with M g in S and if (X, U ) is uniformly distributed on S , the marginal distribution of X is still g. Deduce that the density g only needs to be known up to a normalizing constant. 5.16 For the function g(x) = (1 + sin2 (x))(2 + cos4 (4x)) exp[−x4 {1 + sin6 (x)}] on [0, 2π], examine the feasibility of running a uniform sampler on the set S associated with the accept–reject algorithm in Sect. 5.4. 5.17 Show that the probability of acceptance in Step 2 of Algorithm 5.9 is 1/M and that the number of trials until a variable is accepted has a geometric distribution with parameter 1/M . Conclude that the expected number of trials per simulation is M . 5.18 For the conditional distribution of αt derived from (5.3), construct an accept– reject algorithm based on a normal bounding density f and study its performances for N = 532, nt = 118, μt = −0.5, and σ 2 = 3. 5.19 When uniform simulation on the accept–reject set S of Sect. 5.4 is impossible, construct a Gibbs sampler based on the conditional distributions of u and x. (Hint: Show that both conditionals are uniform distributions.) This special case of the Gibbs sampler is called the slice sampler (see Robert and Casella, 2004, Chap. 8). Apply to the distribution of Exercise 5.16. 12

We will see in Chap. 7 a derivation of this likelihood that enjoys an O(T ) complexity.

5.6

Exercises

171

5.20 Show that the normalizing constant M of a target density f can be deduced from the acceptance rate in the accept–reject algorithm (Algorithm 5.9) under the assumption that g is properly normalized. 5.21 Reproduce the analysis of Exercise 5.20 for the marginal distribution of r1 computed in Exercise 5.13. 5.22 Modify the function ardipper used in Sect. 5.4 to return the acceptance rate as well as a sample from the target distribution. 5.23 Show that, given a mean and a 95% conﬁdence interval in [0, 1], there exists at most one beta distribution Be(a, b) with such a mean and conﬁdence interval. 5.24 Show that, for the Arnason–Schwarz model, groups of consecutive unknown locations are independent of one another, conditional on the observations. Devise a way to simulate these groups by blocks rather than one at a time; that is, using the joint posterior distributions of the groups rather than the full conditional distributions of the states.

6 Mixture Models

I must have missed something. —Ian Rankin, The Hanging Garden.—

Roadmap This chapter covers a class of models where a rather simple distribution is made more complex and less informative by a mechanism that mixes together several known or unknown distributions. This representation is naturally called a mixture of distributions, as illustrated above. Inference about the parameters of the elements of the mixtures and the weights is called mixture estimation, while recovery of the original distribution of each observation is called classiﬁcation (or, more exactly, unsupervised classiﬁcation to distinguish it from the supervised classiﬁcation to be discussed in Chap. 8). Both aspects almost always require advanced computational tools since even the representation of the posterior distribution may be complicated. Typically, Bayesian inference for these models was not correctly treated until the introduction of MCMC algorithms in the early 1990s. This chapter also covers the case of a mixture with an unknown number of components, for which a speciﬁc approximation of the Bayes factor was designed by Chib (1995).

J.-M. Marin and C.P. Robert, Bayesian Essentials with R, Springer Texts in Statistics, DOI 10.1007/978-1-4614-8687-9 6, © Springer Science+Business Media New York 2014

173

174

6 Mixture Models

6.1 Missing Variable Models In some cases, the complexity of a model originates from the fact that some piece of information about an original and more standard (simpler) model is missing. For instance, we have encountered a missing variable model in Chap. 5 with the Arnason–Schwarz model (Sect. 5.5), where the fact of ignoring the characteristics of the individuals outside their capture periods makes inference much harder. Similarly, we have seen in Chap. 4 that the probit model can be reinterpreted as a missing-variable model in that we only observe the sign of a normal variable. Formally, all models that are deﬁned via a marginalization mechanism, that is, such that the density of the observables x, f (x|θ), is given by an integral g(x, z|θ) dz , (6.1) f (x|θ) = Z

can be considered as belonging to a missing variable (or missing data) model.1 This chapter focus on the case of the mixture model, which is the archetypical missing-variable model in that its simple representation (and interpretation) is mirrored by a need for complex processing. Later, in Chap. 7, we will also discuss hidden Markov models that add to the missing structure a temporal dependence dimension. Although image analysis is the topic of Chap. 8, the dataset used in this chapter is derived from an image of a license plate, called License and not available in bayess, as > image(license,col=grey(0:255/255),axes=FALSE,xlab="", ylab="") represented in Fig. 6.1 (top). The actual histogram of the grey levels is concentrated on 256 values because of the poor resolution of the image, but we transformed the original data as > > > +

license=scan("license.txt") license=jitter(license,10) datha=log((license-min(license)+.01)/ (max(license)+.01-license))

where jitter is used to randomize the dataset and avoid repetitions (as already described on page 156). The second line of code is a logit transform.

1

This is not a deﬁnition in the mathematical sense since all densities can formally be represented that way. We thus stress that the model itself must be introduced that way. This point is not to be mistaken for a requirement that the variable z be meaningful for the data at hand. In many cases, for instance the probit model, the missing variable representation remains formal.

6.1

Missing Variable Models

175

The transformed data used in this chapter has been stored in the ﬁle datha.txt. > data(datha) > datha=as.matrix(datha) > hist(datha,nclas=200,xlab="",xlim=c(min(datha),max(datha)), ylab="",prob=TRUE,main="") As seen from Fig. 6.1 (bottom), the resulting structure of the data is compatible with a sample from a mixture of several normal distributions (with at least two components). We point out at this early stage that mixture modeling is often used in image smoothing. Unsurprisingly, the current plate image would instead require feature recognition, for which this modeling does not help, because it requires spatial coherence and thus more complicated models that will be presented in Chap. 8.

0.30

0.20

0.10

0.00 −10

−5

0

5

10

Fig. 6.1. Dataset License: (top) Image of a car license plate and (bottom) histogram of the transformed grey levels of the dataset

176

6 Mixture Models

6.2 Finite Mixture Models We now introduce the speciﬁc case of mixtures as it exempliﬁes the complexity of missing-variable models, both by its nature (in the sense that it is inherently linked with a missing variable) and by its processing, which also requires the incorporation of the missing structure.2 A mixture distribution is a convex combination k

pj fj (x) ,

pj ≥ 0 ,

j=1

k

pj = 1 ,

j=1

of k distributions fj (k > 1). In the simplest situations, the fj ’s are known and inference focuses either on the unknown proportions pj or on the allocations of the points of the sample (x1 , . . . , xn ) to the components fj , i.e. on the probability that xi is generated from fj by opposition to being generated from f , say. In most cases, however, the fj ’s are from a parametric family like the normal or Beta distributions, with unknown parameters θj , leading to the mixture model k pj f (x|θj ) , (6.2) j=1

with parameters including both the weights pj and the component parameters θj (j = 1, . . . , k). It is actually relevant to distinguish the weights pj from the other parameters in that they are solely associated with the missingdata structure of the model, while the others are related to the observations. This distinction is obviously irrelevant in the computation of the likelihood function or in the construction of the prior distribution, but it matters in the interpretation of the posterior output, for instance. There are several motivations for considering mixtures of distributions as a useful extension to “standard” distributions. The most natural approach is to envisage a dataset as made of several latent (that is, missing, unobserved) strata or subpopulations. For instance, one of the earliest occurrences of mixture modeling can be found in Bertillon (1887),3 where the bimodal structure of the heights of (military) conscripts in central France (Doubs) can be explained a posteriori by the aggregation of two populations of young men, one from the plains and one from the mountains. The mixture structure appears because the origin of each observation (that is, the allocation to a speciﬁc subpopulation or stratum) is lost. In the example of the military conscripts, this means that the geographical origin of each young man was not recorded. 2

We will see later that the missing structure of a mixture actually need not be simulated but, for more complex missing-variable structures like hidden Markov models (introduced in Chap. 7), this completion cannot be avoided. 3 The Frenchman Alphonse Bertillon is also the father of scientiﬁc police investigation. For instance, he originated the use of ﬁngerprints in criminal investigations.

6.3

Mixture Likelihoods and Posteriors

177

Depending on the setting, the inferential goal associated with a sample from a mixture of distributions may be either to reconstitute the groups by estimating the missing component z, an operation usually called classiﬁcation (or clustering), to provide estimators for the parameters of the diﬀerent groups, or even to estimate the number k of groups. A completely diﬀerent (if more involved) approach to the interpretation and estimation of mixtures is the semiparametric perspective. This approach considers that since very few phenomena obey probability laws corresponding to the most standard distributions, mixtures such as (6.2) can be seen as a good trade-oﬀ between fair representation of the phenomenon and eﬃcient estimation of the underlying distribution. If k is large enough, there is theoretical support for the argument that (6.2) provides a good approximation (in some functional sense) to most distributions. Hence, a mixture distribution can be perceived as a type of (functional) basis approximation of unknown distributions, in a spirit similar to wavelets and splines, but with a more intuitive ﬂavor (for a statistician at least). However, this chapter mostly focuses on the “parametric” case, that is, on situations when the partition of the sample into subsamples with diﬀerent distributions fj does make sense from the dataset or modelling point of view (even though the computational processing is the same in both cases). In other words, we consider settings where clustering the sample into strata or subpopulations is of interest.

6.3 Mixture Likelihoods and Posteriors Let us consider an iid sample x = (x1 , . . . , xn ) from model (6.2). The likelihood is such that n k (θ, p|x) = pj f (xi |θj ) . i=1 j=1 n

This likelihood contains k terms when the inner sums are expanded. While this expansion is not necessary for computing the likelihood at a given value (θ, p), a computation that is feasible in O(nk) operations as demonstrated by the representation in Fig. 6.2, it remains a necessary step in the understanding of the mixture structure. Alas, the computational diﬃculty in using the expanded version precludes analytic solutions for either maximum likelihood or Bayes estimators. Example 6.1. Consider the simple case of a two-component normal mixture p N (μ1 , 1) + (1 − p) N (μ2 , 1) ,

(6.3)

where the weight p = 0.5 is known. The likelihood surface can be computed by an R code as in the following plotmix function, which relies on the image function and a discretization of the (μ1 , μ2 ) space into pixels. Given a sample

178

6 Mixture Models

sampl that is generated in the ﬁrst lines of the function, the log-likelihood surface is computed by pbar=1-p mu1=mu2=seq(min(sampl),max(sampl),.1) mo1=mu1%*%t(rep(1,length(mu2))) mo2=rep(1,length(mu2))%*%t(mu2) ca1=-0.5*mo1*mo1 ca2=-0.5*mo2*mo2 like=0*mo1 for (i in 1:n) like=like+log(p*exp(ca1+sampl[i]*mo1)+ pbar*exp(ca2+sampl[i]*mo2)) like=like+.1*(ca1+ca2) and plotted by image(mu1,mu2,like,xlab=expression(mu[1]), ylab=expression(mu[2]),col=heat.colors(250)) contour(mu1,mu2,like,levels=seq(min(like),max(like),lengthl), add=TRUE,drawlabels=FALSE) We note that the outcome of the plotmix function is the list list(sample=sampl,like=like), used in subsequent analyses of the data. For instance, this outcome, including the level sets obtained by contour, is provided in Fig. 6.2. In this case, the parameters are identiﬁable: μ1 cannot be confused with μ2 when p is diﬀerent from 0.5. Nonetheless, the log-likelihood surface in this ﬁgure exhibits two modes, one being close to the true value of the parameters used to simulate the dataset and one corresponding to an inverse separation of the dataset into two groups.4 For any prior π (θ, p), the posterior distribution of (θ, p) is available up to a multiplicative constant: ⎤ ⎡ n k pj f (xi |θj )⎦ π (θ, p) . (6.4) π(θ, p|x) ∝ ⎣ i=1 j=1

While π(θ, p|x) can thus be computed for a given value of (θ, p) at a cost of order O(kn), we now explain why the derivation of the posterior characteristics, and in particular of posterior expectations of quantities of interest, is only possible in an exponential time of order O(k n ). To explain this diﬃculty in more detail, we consider the rather intuitive missing-variable representation of mixture models: With each xi is associated 4

To get a better understanding of this second mode, consider the limiting setting when p = 0.5. In that case, there are two equivalent modes of the likelihood, (μ1 , μ2 ) and (μ2 , μ1 ). As p moves away from 0.5, this second mode gets lower and lower compared with the other mode, but it still remains.

6.3

Mixture Likelihoods and Posteriors

179

4

μ2

2

0

−2

−2

0

2

4

μ1

Fig. 6.2. R image representation of the log-likelihood of the mixture (6.3) for a simulated dataset of 500 observations and a true value (μ1 , μ2 , p) = (2.5, 0, 0.7). Besides a mode (represented by a diamond ) in the neighborhood of the true value, the R contour function exhibits an additional mode on the likelihood surface

a missing variable zi that indicates “its” component, i.e. the index zi of the distribution from which it was generated. Formally, this means that we have a hierarchical structure associated with the model: zi |p ∼ Mk (p1 , . . . , pk ) and xi |zi , θ ∼ f (·|θzi ) . The completed likelihood corresponding to the missing structure is such that (θ, p|x, z) =

n

pzi f (xi |θzi )

i=1

and

9 π(θ, p|x, z) ∝

n

: pzi f (xi |θzi ) π (θ, p) ,

i=1

where z = (z1 , . . . , zn ). If we denote by Z = {1, . . . , k}n the set of the k n possible values of the vector z, we can decompose Z into a partition of subsets Z = ∪rj=1 Zj

180

6 Mixture Models

as follows (see Exercise 6.2 for the value of r: For a given allocation size vector (n1 , . . . , nk ), where n1 + . . . + nk , i.e. a given number of observations allocated to each component, we deﬁne the partition sets Zj =

z:

n

Izi =1 , . . . ,

i=1

n

Izi =k

,

i=1

which consist of all allocations with the given allocation size (n1 , . . . , nk ). We label those partition sets with j = j(n1 , . . . , nk ) by using, for instance, the lexicographical ordering on the (n1 , . . . , nk )’s. (This means that j = 1 corresponds to (n1 , . . . , nk ) = (n, 0, . . . , 0), j = 2 to (n1 , . . . , nk ) = (n − 1, 1, . . . , 0), j = 3 to (n1 , . . . , nk ) = (n − 1, 0, 1, . . . , 0), and so on). Using this partition, the posterior distribution of (θ, p) can be written in closed form as π (θ, p|x) =

π (θ, p|x, z) =

r

ω (z) π (θ, p|x, z) ,

(6.5)

i=1 z∈Zi

z∈Z

where ω (z) represents the marginal posterior probability of the allocation z conditional on the observations x (derived by integrating out the parameters θ and p). With this representation, a Bayes estimator of (θ, p) can also be written in closed form as Eπ [θ, p|x] =

r

ω (z) Eπ [θ, p|x, z] .

i=1 z∈Zi

Continuation of Example 6.1. In the special case of model (6.3), if we take two diﬀerent independent normal priors on both means, μ1 ∼ N (0, 4) ,

μ2 ∼ N (2, 4) ,

the posterior weight of a given allocation vector z is * ω (z) ∝ (n1 + 1/4)(n − n1 + 1/4) pn1 (n1 − p)n−l × exp −[(n1 + 1/4)ˆ s1 (z) + n1 {¯ x1 (z)}2 /4]/2 s2 (z) + (n − n1 ){¯ x2 (z) − 2}2 /4]/2 , × exp −[(n − n1 + 1/4)ˆ x ¯1 (z) = sˆ1 (z) =

n i=1

n 1 Iz =1 xi , n1 i=1 i

x¯2 (z) = 2

Izi =1 (xi − x ¯1 (z)) ,

n 1 Iz =2 xi , n − n1 i=1 i

sˆ2 (z) =

n

2

Izi =2 (xi − x ¯2 (z))

i=1

(if we set x ¯1 (z) = 0 when n1 = 0 and x¯2 (z) = 0 when n − n1 = 0). Implementing this derivation in R is quite straightforward:

6.3

Mixture Likelihoods and Posteriors

181

omega=function(z,x,p){ n=length(x) n1=sum(z==1);n2=n-n1 if (n1==0) xbar1=0 else xbar1=sum((z==1)*x)/n1 if (n2==0) xbar2=0 else xbar2=sum((z==2)*x)/n2 ss1=sum((z==1)*(x-xbar1)^2) ss2=sum((z==2)*(x-xbar2)^2) return(sqrt((n1+.25)*(n2+.25))*p^n1*(1-p)^n2* exp(-((n1+.25)*ss1+(n2+.25)*ss2)/2)* exp(-(n1*xbar1^2+n2*xbar2)/8)) } leading for instance to > omega(z=sample(1:2,4,rep=TRUE), + x=plotmix(n=4,plot=FALSE)$samp,p=.8) [1] 0.0001781843 > omega(z=sample(1:2,4,rep=TRUE), + x=plotmix(n=4,plot=FALSE)$sample,p=.8) [1] 5.152284e-09 Note that the omega function is not and cannot be normalized, so the values must be interpreted on a relative scale. The decomposition (6.5) makes a lot of sense from an inferential point of view. The posterior distribution simply considers each possible partition z of the dataset, then allocates a posterior probability ω (z) to this partition, and at last constructs a posterior distribution for the parameters conditional on this allocation. Unfortunately, the computational burden resulting from this decomposition is simply too intensive because there are k n terms in the sum. However, there exists a solution that overcomes this computational problem. It uses an MCMC approach that takes advantage of the missing-variable structure and removes the requirement to explore the k n possible values of z by only looking at the most likely ones. Although this is beyond the scope of the book, let us point out here that there also exists in the statistical literature a technique that predates MCMC simulation algorithms but still relates to the same missing-data structure and completion mechanism. It is called the EM Algorithm5 and consists of an iterative but deterministic sequence of “E” (for expectation) and “M” (for maximization) steps that converge to a local maximum of the likelihood. At iteration t, the “E” step corresponds to the computation of the function Q{(θ(t) , p(t) ), (θ, p)} = E(θ (t) ,p(t) ) [log (θ, p|x, z)|x] , 5 In non-Bayesian statistics, the EM algorithm is certainly the most ubiquitous numerical method, even though it only applies to (real or artiﬁcial) missing variable models.

182

6 Mixture Models

where the likelihood (θ, p|x, z) is the joint distribution of x and z, while the expectation is computed under the conditional distribution of z given x and the value (θ (t) , p(t) ) for the parameter. The “M” step corresponds to the maximization of Q((θ (t) , p(t) ), (θ, p)) in (θ, p), with solution (θ (t+1) , p(t+1) ). As we will see in Sect. 6.4, the Gibbs sampler takes advantage of exactly the same conditional distribution. Further details on EM and its Monte Carlo versions (namely, when the “E” step is not analytically feasible) are given in Robert and Casella (2004, Chap. 5; 2009, Chap. 5).

6.4 MCMC Solutions For the joint distribution (6.4), the full conditional distribution of z given x and the parameters is always available as π(z|x, θ, p) ∝

n

pzi f (xi |θzi )

i=1

and can thus be computed at a cost of O(n). Since, for standard distributions f (·|θ), the full posterior conditionals are also easily simulated when using conjugate priors, this implies that the Gibbs sampler can be derived in this setting.6 If p and θ are independent a priori, then, given z, the vectors p and x are independent; that is, π(p|z, x) ∝ π(p)f (z|p)f (x|z) ∝ π(p)f (z|p) ∝ π(p|z) . Moreover, in that case, θ is also independent a posteriori from p given z and x, with density π(θ|z, x). If we apply the Gibbs sampler in this problem, it involves the successive simulation of z and (p, θ) conditional on one another and on the data:

6

Historically, missing-variable models constituted one of the ﬁrst instances where the Gibbs sampler was used by completing the missing variables by simulation under the name of data augmentation (see Tanner, 1996, and Robert and Casella, 2004, Chaps. 9 and 10).

6.4

MCMC Solutions

183

Algorithm 6.11 Mixture Gibbs Sampler Initialization: Choose p(0) and θ (0) arbitrarily. Iteration t (t ≥ 1): (t) 1. For i = 1, . . . , n, generate zi such that (t−1) (t−1) . f xi |θj P (zi = j|θ, p) ∝ pj 2. Generate p(t) according to π(p|z(t) ). 3. Generate θ(t) according to π(θ|z(t) , x). The simulation of the pj ’s is also generally obvious since there exists a conjugate prior (as detailed below). In contrast, the complexity in the simulation of the θj ’s will depend on the type of sampling density f (·|θ) as well as the prior π. The marginal (sampling) distribution of the zi ’s is a multinomial distribution Mk (p1 , . . . , pk ), which allows for a conjugate prior on p, namely the Dirichlet distribution p ∼ D (γ1 , . . . , γk ), with density Γ (γ1 + . . . + γk ) γ1 p · · · pγkk Γ (γ1 ) · · · Γ (γk ) 1 on the simplex of Rk , Sk =

⎧ ⎨ ⎩

(p1 , . . . , pk ) ∈ [0, 1]k ;

In this case, denoting nj =

k

pj = 1

j=1 n

⎫ ⎬ ⎭

.

Izl =j (1 ≤ j ≤ k) the allocation sizes, the

l=1

posterior distribution of p given z is p|z ∼ D (n1 + γ1 , . . . , nk + γk ) . It is rather peculiar that, despite its importance for Bayesian statistics, the Dirichlet distribution is not available in R (at least in the standard stat package). It is however fairly straightforward to code, using a representation based on gamma variates, as shown below. rdirichlet=function(n=1,par=rep(1,2)){ k=length(par) mat=matrix(0,n,k) for (i in 1:n){

184

6 Mixture Models

sim=rgamma(k,shape=par,scale=1) mat[i,]=sim/sum(sim) } mat } When the density f (·|θ) also allows for conjugate priors, the simulation of θ can be speciﬁed further since an independent conjugate prior on each θj leads to independent and conjugate posterior distributions on the θj ’s, given z and x. Continuation of Example 6.1. For the mixture (6.3), under independent normal priors N (δ, 1/λ) (both δ ∈ R and λ > 0 are ﬁxed hyperparameters) on both μ1 and μ2 , the parameters μ1 and μ2 are independent given (z, x), with conditional distributions λδ + n1 x λδ + (n − n1 )¯ ¯1 (z) x2 (z) 1 1 N , , and N , λ + n1 λ + n1 λ + n − n1 λ + n − n1 respectively. Similarly, the conditional posterior distribution of the zi ’s given (μ1 , μ2 ) is (i = 1, . . . , n) P (zi = 1|μ1 , xi ) ∝ p exp −0.5 (xi − μ1 )2 . We can thus construct an R function like the following one to generate a sample from the posterior distribution: assuming δ = 0 and λ = 1, gibbsmean=function(p,datha,niter=10^4){ n=length(datha) z=rep(0,n); ssiz=rep(0,2) nxj=rep(0,2) mug=matrix(mean(datha),nrow=niter+1,ncol=2) for (i in 2:(niter+1)){ for (t in 1:n){ prob=c(p,1-p)*dnorm(datha[t],mean=mug[i-1,]) z[t]=sample(c(1,2),size=1,prob=prob) } for (j in 1:2){ ssiz[j]=1+sum(z==j) nxj[j]=sum(as.numeric(z==j)*datha) } mug[i,]=rnorm(2,mean=nxj/ssiz,sd=sqrt(1/ssiz)) } mug }

6.4

MCMC Solutions

185

which can be used as > dat=plotmix()$sample > simu=gibbsmean(0.7,dat) > points(simu,pch=".") to produce Figs. 6.3 and 6.4. This R code illustrates two possible behaviors of this algorithm if we use a simulated dataset of 500 points from the mixture 0.7N (0, 1) + 0.3N (2.5, 1), which corresponds to the level sets on both pictures. The starting point in both cases is located at the saddle point between the two modes, i.e. at an instable equilibrium. Depending on the very ﬁrst (random) iterations of the algorithm, the ﬁnal sample may end up located on the upper or on the lower mode. For instance, in Fig. 6.3, the Gibbs sample based on 10,000 iterations is in agreement with the likelihood surface, since the second mode discussed in Example 6.1 is much lower than the mode where the simulation output concentrates. In Fig. 6.4, the Gibbs sample ends up being trapped by this lower mode. Example 6.2. If we consider the more general case of a mixture of two normal distributions with all parameters unknown, pN (μ1 , σ12 ) + (1 − p)N (μ2 , σ22 ) , and for the conjugate prior distribution (j = 1, 2)

4

μ2

2

0

−2

−2

0

2

4

μ1

Fig. 6.3. Log-likelihood surface and the corresponding Gibbs sample for the model (6.3), based on 10,000 iterations

186

6 Mixture Models

4

μ2

2

0

−2

−2

0

μ1

2

4

Fig. 6.4. Same legend as Fig. 6.3, with the same starting point located at the saddle point. In this instance, the Gibbs sample ends up around a lower mode

μj |σj ∼ N (ξj , σj2 /lj ) ,

σj2 ∼ I G (νj /2, s2j /2) ,

p ∼ Be(α, β) ,

the same decomposition conditional on z and straightforward (if dreary) algebra imply that p|x, z ∼ Be(α + n1 , β + n2 ), σj2 , σj2 |x, z ∼ I G ((νj + nj )/2, sj (z)/2) , μj |σj , x, z ∼ N ξ1 (z), nj + l j ¯j (z) and sˆ2j (z) are the empirical where nj is the number of zi equal to j, x mean and variance (biased) for the subsample with zi equal to j, and ξj (z) =

lj ξj + nj x¯j (z) , l j + nj

sj (z) = s2j + nj sˆ2j (z) +

l j nj (ξj − x¯j (z))2 . l j + nj

The modiﬁcation of the above R code is also straightforward and we do not reproduce it here to save space. The extension to more than two components is equally straightforward, as described below for License.

6.4

MCMC Solutions

187

If we model License by a k = 3 component normal mixture model, we start by deriving the prior distribution from the scale of the problem. Namely, we choose a D3 (1/2, 1/2, 1/2) prior for the weights (although picking parameters less than 1 in the Dirichlet prior has the potential drawback that it may allow very small weights for some components), a N (x, σi2 ) distribution on ˆ 2 ) distribution on the precisions σi−2 , where x the means μi , and a G a(10, σ 2 and σ ˆ are the empirical mean and variance of License, respectively. (This empirical choice of a prior is debatable on principle, as it depends on the dataset, but this is relatively harmless since it is equivalent to standardizing the dataset so that the empirical mean and variance are equal to 0 and 1, respectively.) If we deﬁne the parameter vector mix as a list, > mix=list(k=k,p=p,mu=mu,sig=sig) our R function gibbsnorm=function(niter,mix) is made of an initialization step: n=length(datha);k=mix$k z=rep(0,n) #missing data nxj=rep(0,k) ssiz=ssum=rep(0,k) mug=sigg=prog=matrix(0,nrow=niter,ncol=k) lopost=rep(0,niter) #log-posterior lik=matrix(0,n,k) prog[1,]=rep(1,k)/k;mug[1,]=rep(mix$mu,k) sigg[1,]=rep(mix$sig,k) #current log-likelihood for (j in 1:k) lik[,j]=prog[1,j]*dnorm(x=datha,mean=mug[1,j], sd=sqrt(sigg[1,j])) lopost[1]=sum(log(apply(lik,1,sum)))+ sum(dnorm(mug[1,],mean(datha),sqrt(sigg[1,]),log=TRUE))(10+1)*sum(log(sigg[1,]))-sum(var(datha)/sigg[1,])+ .5*sum(log(prog[1,])) and of the main loop for data completion and conditional parameter simulation: for (i in 1:(niter-1)){ for (t in 1:n){ #missing data completion prob=prog[i,]*dnorm(datha[t],mug[i,],sqrt(sigg[i,])) if (sum(prob)==0) prob=rep(1,k)/k z[t]=sample(1:k,1,prob=prob) }

188

6 Mixture Models

#conditional parameter simulation for (j in 1:k){ ssiz[j]=sum(z==j) nxj[j]=sum(as.numeric(z==j)*datha) } mug[i+1,]=rnorm(k,(mean(datha)+nxj)/(ssiz+1), sqrt(sigg[i,]/(ssiz+1))) for (j in 1:k) ssum[j]=sum(as.numeric(z==j)*(datha-nxj[j]/ssiz[j])^2) sigg[i+1,]=1/rgamma(k,shape=.5*(20+ssiz),rate=var(datha)+ .5*ssum+.5*ssiz/(ssiz+1)*(mean(datha)-nxj/ssiz)^2) prog[i+1,]=rdirichlet(1,par=ssiz+0.5) #current log-likelihood for (j in 1:k) lik[,j]=prog[i+1,j]*dnorm(x=datha,mean=mug[i+1,j], sd=sqrt(sigg[i+1,j])) lopost[i+1]=sum(log(apply(lik,1,sum)))+ sum(dnorm(mug[i+1,],mean(datha),sqrt(sigg[i+1,]),log=TRUE))(10+1)*sum(log(sigg[i+1,]))-sum(var(datha)/sigg[i+1,])+ .5*sum(log(prog[i+1,])) } returning all simulated values as a list list(k=k,mu=mug,sig=sigg,p=prog,lopost=lopost) The output of this R function, represented in Fig. 6.5 as an overlay of the License histogram is then produced by the R code mix=list(k=3,mu=mean(datha),sig=var(datha)) simu=gibbsnorm(1000,mix) hist(datha,prob=TRUE,main="",xlab="",ylab="",nclass=100) x=y=seq(min(datha),max(datha),length=150) yy=matrix(0,ncol=150,nrow=1000) for (i in 1:150){ yy[,i]=apply(simu$p*dnorm(x[i],mean=simu$mu, sd=sqrt(simu$sig)),1,sum) y[i]=mean(yy[,i]) } for (t in 501:1000) lines(x,yy[t,],col="gold") lines(x,y,lwd=2.3,col="sienna2") This output demonstrates that this crude prior modeling is suﬃcient to capture the modal features of the histogram as well as the tail behavior in a surprisingly small number of Gibbs iterations, despite the large sample size of

6.4

MCMC Solutions

189

2,625 points. The range of the simulated densities represented in Fig. 6.5 reﬂects the variability of the posterior distribution, while the estimate of the density is obtained by averaging the simulated densities over the 500 iterations.7

0.20 0.10 0.00 −10

−5

0

5

10

Fig. 6.5. Dataset License: Representation of 500 Gibbs iterations for the mixture estimation. (The accumulated lines correspond to the estimated mixtures at each iteration and the overlaid curve to the density estimate obtained by summation.)

The experiment produced in Example 6.1, page 184, gives a false sense of security about the performance of the Gibbs sampler because it hides the structural dependence of the sampler on its initial conditions. The fundamental feature of Gibbs sampling—its derivation from conditional distributions—implies that it is often restricted in the width of its moves and that, in some situations, this restriction may even jeopardize convergence. In the case of mixtures of distributions, conditioning on z implies that the proposals for (θ, p) are quite concentrated and do not allow drastic changes in the allocations at the next step. To obtain a signiﬁcant modiﬁcation of z requires a considerable number of iterations once a stable position has been reached.8 Figure 6.4 illustrates this phenomenon for the very same sample as in Fig. 6.3: A Gibbs sampler initialized at the saddlepoint may get close to the second mode in the very ﬁrst iterations and is then unable to escape its (fatal) attraction, even after a large number of iterations, for the reason given above. It is quite interesting to see that this Gibbs sampler suﬀers from the same pathology as the EM algorithm. However, this is not immensely surprising given that it is based on a similar principle. In general, there is very little one can do about improving the Gibbs sampler since its components are given by the joint distribution. The solutions are (a) to change the parameterization and thus the conditioning (see 7

That this is a natural estimate of the model, compared with the “plug-in” density using the estimates of the parameters, will be explained more clearly in Sect. 6.5. 8 In practice, the Gibbs sampler never leaves the vicinity of a given mode if the attraction of this mode is strong enough, for instance in the case of many observations.

190

6 Mixture Models

Exercise 6.6), (b) to use tempering to facilitate exploration (see Sect. 6.7), or (c) to mix the Gibbs sampler with another MCMC algorithm. To look for alternative MCMC algorithms is not a diﬃculty in this setting, given that the likelihood of mixture models is available in closed form, being computable in O(kn) time, and the posterior distribution is thus known up to a multiplicative constant. We can therefore use any Metropolis–Hastings algorithm, as long as the proposal distribution q provides a correct exploration of the posterior surface, since the acceptance ratio π(θ , p |x) q(θ, p|θ , p ) ∧1 π(θ, p|x) q(θ , p |θ, p) can be computed in O(kn) time. For instance, we can use a random walk Metropolis–Hastings algorithm where each parameter is the mean of the proposal distribution for the new value, that is, (t−1) + uj , ξ;j = ξj

where uj ∼ N (0, ζ 2 ) and ζ is chosen to achieve a reasonable acceptance rate. Continuation of Example 6.1. For the posterior associated with (6.3), the Gaussian random walk proposal is (t−1) 2 (t−1) 2 μ

+ > > > + + >

simu=gibbsnorm(1000,mix) plot(simu$mu[,1],ylim=range(simu$mu), ylab=expression(mu[i]),xlab="n",type="l",col="sienna3") lines(simu$mu[,2],col="gold4") lines(simu$mu[,3],col="steelblue") plot(simu$mu[,2],simu$p[,2],col="sienna3", xlim=range(simu$mu),ylim=range(simu$p), xlab=expression(mu[i]),ylab=expression(p[i])) points(simu$mu[,3],simu$p[,3],col="steelblue")

194

6 Mixture Models

In Fig. 6.7, we see that each component is thus identiﬁed by its mean, and the posterior distributions of the means are very clearly distinct. Although this result has the appeal of providing distinct estimates for the three components, it suﬀers from the severe drawback that the Gibbs sampler has not explored the whole parameter space after 1,000 iterations. Running the algorithm for a much longer period does not solve this problem since the Gibbs sampler cannot simultaneously switch enough component allocations in this highly peaked setup. In other words, the algorithm is unable to explore more than one of the 3! = 6 equivalent modes of the posterior distribution. Therefore, it is diﬃcult to trust the estimates derived from such an output.

2.0

l

pi

μi

1.0 0.0

0.4

ll l l ll lll l l l l l l l ll l l l ll l l l l l l l l l l l l l l ll l ll l l l l l l l l l l l l l l l l l l ll l l l l l l l l ll l l l l l l l l l l l ll l l l l l l l l l ll l l l l l l l l l l l l l l l l l l l l l l l l l l l l l l ll l l ll l l l l l l l l l l l l l l l l l l l l l l l l l l l l l ll l l l l l l l l l l l l l l l ll l l l l l l l l l l l l l l l l l l l l l l l l l l l l l l l l l l l l l ll

0.2

−1.5

0

200

400

600

800

−1.5 −1.0 −0.5

1000

l l l lll lll l l l l lll ll l l l l l l ll ll l l lll ll l l ll ll l l l l ll l l l l l l ll l l l ll l l l l l l ll l l l l l l l l l l l l l l l l ll l l l l l l l l ll l l l l ll l l l l l ll ll l l l l l l l l l l l l l l l l l l l l l l l l l l l l l l l l l l l l l l l l l l l l l l l l l l l l l l l ll l l l l l l l l l l l l l l l l ll l l ll l l l l l l l l l l l l l l l l l ll l l ll l l l l l l l l l l ll l l ll l l l l l l l l l l l l l l l l l l l l l l l l l l l l l l l ll l l l l l l l l l l l l l l l l l l l l l l l l l l l l ll l l l l l l l l l l l l l l l l l l l l l l l l l l ll l l ll l l l ll l l l l l l ll l ll l l l l l l l l l l ll l l l ll l l l l l l l l l l l ll l l l l l ll l l l l l l l

0.0

5

σ2i

pi

0.4

3 2 1

0.2

0

400

600

800

0.2

1000

0.3

2.0

4

1.0

3

μi

σi2

5

2

l l l l l l l l l l l l l l l l l l l l l l l l ll l ll l ll l ll l l l l l l l l l l l l l l l l l l l l l l l l l l l l ll l l l l l l l l ll l l ll l ll lll ll l l ll l l l l l l l l l l l l l l l l l l ll l ll l l l l l ll l l l l l l l l l l l ll ll l l l

0.0

1

l lll l l l l l l l l l l l l l ll l l l ll l l l l l l l l ll l l l l l l l l l l l l l l l l l l l l l l l l l l l l l l l l l l l l l l l l

−1.5

0

200

400

600 t

2.0

0.4

0.5

pi

t

0

1.5

l ll ll lll ll l l l lllll l lll ll l l l l l l l l l ll ll l l ll l l ll l l l l l l l l l l ll l l ll ll l l l l ll l l l l l l l ll l l l l l l l l l l l l l l l l l l l l l l l l l l l ll l ll l l l l l l l l l l l l l l l l l l l l l l l l l l l l l l l l l l ll l l l l ll l l l l l l l l l l l l ll ll l l ll l l l l l l l l l l ll l l l l l l l l l l l l l l l l l l l l l l l ll l l l l l l l l l l l l ll l l l l l l ll l l ll l l l l ll l l l l l ll l l l l l l l l l l l l l l l l l l l l l l l ll l l l l l l l l l l l l l l l l l l l l l l ll l l l l l l l l l l l l l l l l l l l l l l l l l l l l l l l l l l l l l l l l l l l l l l l ll l l l l l l l l l l l l l l l l l l ll l l ll l l l l l l l l l l ll l l l l l l l l l l l l l l l l l l l l l l l l l l l l l l l l l l l l l l ll ll l l l l l l l l l l ll l l l l l l l l l l l l l l l l l l l l l l l l l lll l llll ll l l l l l l l l l lll ll l l l l l l ll l l ll l l l l l l l l l l l l l l l l l l l l l l l l ll l l l l l l l l l l ll lll l ll l ll l l l l l ll ll l l l l ll l l l l l l l l ll l ll l l l l l l l l l ll l l l l ll l l l l ll l l lll l l l l l l l l l lll ll l l l l l l ll l l l ll l ll ll l l l lll l l l l l llll l lll lll l l l l l l l ll l l ll l l l l l l ll l l ll l l l l l l l l l l ll l l ll l l l l l ll l l ll l l ll l l ll l l l l l llll ll l ll l l l l l l l l l l l l l l l ll l l l ll ll ll l l l ll l l l l l ll l l l lll l l l l l l ll ll ll ll l l l l l l l l l l l l ll ll l ll l lll ll l l l ll l l ll l l l l ll l ll l l l l l l ll l ll l l l l l l l l ll l l l l ll l ll l ll l l l l l lll l l lll l l ll l l ll l l l l l l l l l l l ll l ll l l l l l l l l l ll ll ll ll ll ll l l l l l l l l ll ll ll l l l l l l l ll l l l l ll l l l l l l l l l l ll l l l l l l l ll ll l l l l l l l l l l l l l l l ll l ll ll l l l l ll l l ll l l l lll l ll l l ll l l ll ll l ll l ll l l l l l l l l l l l l l ll l l l l l l l l ll l l l l l l ll lll l l ll l l ll ll l lll l l ll l ll l ll ll l l l l l l l ll l ll ll l ll l l l ll l l ll ll l l l l ll l l l lll l l l l l ll lll l l ll l l l l l l ll l l l l l l ll ll l l l l l ll l l l l l l l l l l ll l ll ll l l l l l l

4

200

1.0

μi

t

0

0.5

l lll l l l l ll l l l l l ll ll l l l l l l l l l l l l l l l l l l ll l l l l l ll l l l l l l l l l l l l l l l l l l l l l l l l ll l l l l ll l l l ll l l l l l l l l l l l l l l l l l l l l l l l l l l l l l l l l l ll l l l l l l l ll l l l l l l l l l l l l l l l l ll l ll l l l l l l l l l l l l l l l l l l l l l l l l l l l l ll l l l ll l l l l l l l l l l l l lll l l l l ll l l l l l l l l l l l l l l l l l l l l l l lll l l l l l l l l ll l l l l ll l l l l l l l l l l l l l l l l l l l l l l l l l l lll l l l l l l l ll l l l l l l l l l l l l l l l l l l l l l ll l l l l l l l l l l l l l l l l l l l l l l l l l l l l l l l l l l ll l l l l l l l l l l l l l l l l l l l l l l llll l l l l l l l l l l ll l l l l l lll ll ll ll lll l llll l l lll l ll l l l l l ll l l l ll ll ll l l l l l ll ll

800

1000

0

l l ll l l ll ll l lll ll ll l ll l ll l l l l l l l ll l l l l l l l l l ll ll l l l l l l l l l ll l l l l l l l l l l l l l l l l l l l l l ll l l l l l l l l l l l l l l l l l ll lllllll l l l l l l l l l l l l l l l l l l l l l l l l l l l l l l l ll l l l l l l l l l ll l l l l l l l l l l l l l l l l ll l l l l l l l l l l l l l l l ll ll ll l l l l l l l l ll l l ll l l l l l l l l l l l l l l l ll l l l l l l l l l ll l ll l l l l ll l l l l l l l l l l l l l l l ll ll l l ll l l l l l l l l l l l l l l l l l l l l l l l l l l l l l ll lll l l l l ll l l l l l l l l l l l l l l l l l l l l l l l l l l l l ll l l l ll l l l l ll l l l ll l l l l l l l l l l l l l l l l l l l l l l l l l ll l l l l l l l l l l l l l l lll l l l l l l l l l l l ll l l l l ll ll l l l ll l l l l l l l l l l ll l l ll ll l ll lll l l ll l

lll l l ll ll l l l l

1

l

l

l

2

3

4

5

σ2i

Fig. 6.7. Dataset License: (left) Convergence of the three types of parameters of the normal mixture, each component being identiﬁed by a diﬀerent grey level /color ; (right) 2 × 2 plot of the Gibbs sample for the three types of parameters of a normal mixture

This identiﬁability problem related to the exchangeability of the posterior distribution, often called “label switching,” thus requires either an alternative prior modeling or a more tailored inferential approach. A na¨ıve answer to the problem is to impose an identiﬁability constraint on the parameters, for instance deﬁning the components by ordering the means (or the variances or

6.5

Label Switching Diﬃculty

195

the weights) in a normal mixture (see Exercise 6.3). From a Bayesian point of view, this amounts to truncating the original prior distribution, going from π (θ, p) to π (θ, p) Iμ1 ≤...≤μk for instance. While this seems innocuous (given that the sampling distribution is the same with or without this indicator function), the introduction of an identiﬁability constraint has severe consequences on the resulting inference, both from a prior and from a computational point of view. When reducing the parameter space to its constrained part, the imposed truncation has no reason to respect the topology of either the prior or the likelihood. Instead of singling out one mode of the posterior, the constrained parameter space may then well include parts of several modes and the resulting posterior mean could, for instance, lie in a very low probability region between the modes, while the high posterior probability zones are located at the boundaries of this space. In addition, the constraint may radically modify the prior modeling and come close to contradicting the prior information. For large values of k, the introduction of a constraint also has a consequence on posterior inference: With many components, the ordering of components in terms of one of the parameters of the mixture is unrealistic. Some components will be close in mean while others will be close in variance or in weight. This may even lead to very poor estimates of the parameters if the inappropriate ordering is chosen. Note that while imposing a constraint that is not directly related to the modal regions of the target distribution may considerably reduce the eﬃciency of an MCMC algorithm, it must be stressed that the constraint does not need to be imposed during the simulation but can instead be imposed after simulation by reordering the MCMC output according to the constraint. For instance, if the constraint imposes an ordering of the means, once the simulation is over, the components can be relabeled for each MCMC iteration according to this constraint; that is, deﬁning the ﬁrst component as the one associated with the smallest simulated mean and so on. From this perspective, identiﬁability constraints have nothing to do with (or against) simulation. An empirical resolution of the label switching problem that avoids imposing the constraints altogether consists of arbitrarily selecting one of the k! modal regions of the posterior distribution once the simulation step is over and only then operate the relabeling in terms of proximity to this region. Given an MCMC sample of size M , we can ﬁnd a Monte Carlo approxi∗ ∗ mation of the maximum a posteriori (MAP) estimator by taking θ(i ) , p(i ) such that ( ) i∗ = arg max π (θ, p)(i) |x ; i=1,...,M

196

6 Mixture Models

that is, the simulated value that gives the maximal posterior density. (Note that π does not need its normalizing constant for this computation.) This value is quite likely to be in the vicinity of one of the k! modes, especially if we run many simulations. The approximate MAP estimate will thus act as a pivot in the sense that it gives a good approximation to a mode and we can reorder the other iterations with respect to this mode. Rather than selecting the reordering based on a Euclidean distance in the parameter space, we use a distance in the space of allocation probabilities. Indeed, the components of the parameter vary in diﬀerent spaces, from the real line for the means to the simplex for the weights. Let Sk be the k-permutation set and τ ∈ Sk . We suggest to minimize in τ an entropy dis (i ) = j|θ , p(i ) )’s and tance summing(the relative entropies between the P(z t ) the P(zt = j|τ (θ (i) , p(i) ) )’s, namely

h(i, τ ) =

n k

P(zt = j|θ(i ) , p(i ) )

t=1 j=1

) (

× log P(zt = j|θ (i ) , p(i ) ) P(zt = j|τ (θ(i) , p(i) ) ) . The selection of the permutations reordering the MCMC output thus reads as follows: Algorithm 6.12 Pivotal Reordering At iteration i ∈ {1, . . . , M }: 1. Compute τi = arg min h(i, τ ) , τ ∈Sk

2. Set (θ , p(i) ) = τi {(θ , p(i) )}. (i)

(i)

Thanks to this reordering, most iteration labels get switched to the same mode (when n gets large, this is almost a certainty), and the identiﬁability problem is thus solved. Therefore, after this reordering step, the Monte Carlo estimate of the posterior expectation Eπ [θi |x], M j=1

(θi )(j) M ,

6.5

Label Switching Diﬃculty

197

can be used as in a standard setting because the reordering automatically gives diﬀerent meanings to diﬀerent components. Obviously, Eπ [θi |x] (or its ∗ approximation) should also be compared with θ(i ) to check convergence.9 Using the Gibbs output simu of License (which is the datha of the following code) as in the previous illustration, the corresponding R code involves the determination of the MAP approximation indimap=order(simu$lopost,decreasing=TRUE)[1] map=list(mu=simu$mu[indimap,],sig=simu$sig[indimap,], p=simu$p[indimap,]) that is easily derived by storing the values of the log-likelihood densities in the Gibbs sampling function gibbsnorm. The corresponding (MAP) allocation probabilities for the data are then lili=alloc=matrix(0,length(datha),3) for (t in 1:length(datha)){ lili[t,]=map$p*dnorm(datha[t],mean=map$mu, sd=sqrt(map$sig)) lili[t,]=lili[t,]/sum(lili[t,]) } They are used as reference for the reordering: ormu=orsig=orp=matrix(0,ncol=3,nrow=1000) library(combinat) perma=permn(3) for (t in 1:1000){ entropies=rep(0,factorial(3)) for (j in 1:n){ alloc[j,]=simu$p[t,]*dnorm(datha[j],mean=simu$mu[t,], sd=sqrt(simu$sig[t,])) alloc[j,]=alloc[j,]/sum(alloc[j,]) for (i in 1:factorial(3)) entropies[i]=entropies[i]+ sum(lili[j,]*log(alloc[j,perma[[i]]])) } best=order(entropies,decreasing=TRUE)[1] ormu[t,]=simu$mu[t,perma[[best]]] orsig[t,]=simu$sig[t,perma[[best]]] orp[t,]=simu$p[t,perma[[best]]] } 9

While this resolution seems intuitive enough, there is still a lot of debate in academic circles on whether or not label switching should be observed on an MCMC output and, in case it should, on which substitute to the posterior mean should be used.

198

6 Mixture Models

An output comparing the original MCMC sample and the one corresponding to this reordering for the License dataset is then constructed. However, since the Gibbs sampler does not switch between the k! modes in this case, the above reordering does not modify the labelling and we thus abstain from producing the corresponding graph as it is identical to Fig. 6.7.

6.6 Prior Selection After10 insisting in Chap. 2 that conjugate priors are not the only possibility for prior modeling, we seem to be using them quite extensively in this chapter! The fundamental reason for this is that, as explained below, it is not possible to use the standard alternative of noninformative priors on the components. Nonconjugate priors can be used as well (with Metropolis–Hastings steps) but are diﬃcult to fathom when the components have no speciﬁc “real” meaning (as, for instance, when the mixture is used as a nonparametric proxy). The representation (6.2) of a mixture model precludes the use of independent improper priors, π (θ) =

k

πj (θj ) ,

j=1

since if, for any 1 ≤ j ≤ k,

πj (θi )dθj = ∞ ,

then, for every n,

π(θ, p|x)dθdp = ∞ .

The reason for this inconvenient behavior is that among the k n terms in the expansion (6.5) of π(θ, p|x), there are (k − 1)n terms without any observation allocated to the ith component and thus there are (k − 1)n terms with a conditional posterior π(θi |x, z) that is equal to the prior πi (θi ). The inability to use improper priors may be seen by some as a marginalia, a fact of little importance, since they argue that proper priors with large variances can be used instead. However, since mixtures are ill-posed problems,11 this diﬃculty with improper priors is more of an issue, given that the 10 This section may be skipped by most readers, as it only addresses the very speciﬁc issue of handling improper priors in mixture estimation. 11 By nature, ill-posed problems are not precisely deﬁned. They cover classes of models such as inverse problems, where the complexity of getting back from the data to the parameters is huge. They are not to be confused with nonidentiﬁable problems, though.

6.7

Tempering

199

inﬂuence of a particular proper prior, no matter how large its variance, cannot be truly assessed. In other words, the prior gives a speciﬁc meaning to what distinguishes one component from another.

Prior distributions must always be chosen

with the utmost care when dealing with mixtures and their bearing on the resulting inference assessed by a sensitivity study. The fact that some noninformative priors are associated with undeﬁned posteriors, no matter what the sample size, is a clear indicator of the complex nature of Bayesian inference for those models.

6.7 Tempering The notion of tempering can be found in diﬀerent areas under many diﬀerent denominations, but it always comes down to the same intuition that governs simulated annealing (Chap. 8), namely that when you ﬂatten a posterior surface, it is easier to move around, while if you sharpen it, it gets harder to do so except around peaks. More formally, given a density π(x), we can deﬁne an associated density πα (x) ∝ π(x)α for α > 0 large enough (if α is too small, π(x)α does not integrate). An important property of this family of distributions is that they all share the same modes. When α > 1, the surface of πα is more contrasted than the surface of π: Peaks are higher and valleys are lower. Increasing α to inﬁnity results in a Dirac mass at the modes of π, and this is the principle behind simulated annealing. Conversely, lowering α to values less than 1 makes the surface smoother by lowering peaks and raising valleys. In a compact space, lowering α to 0 ends up with the uniform distribution. This rather straightforward intuition can be exploited in several directions for simulation. For instance, a tempered version of π, πα , can be simulated in a preliminary step to determine where the modal regions of π are. (Diﬀerent values of α can be used in parallel to compare the results.) This preliminary exploration can then be used to build a more appropriate proposal. Alternatively, these simulations may be pursued and associated with appropriate importance weights. Note also that a regular Metropolis–Hastings algorithm may be used with πα just as well as with π since the acceptance ratio is transformed into α q(θ, p|θ , p ) π(θ , p |x) ∧1 (6.7) π(θ, p|x) q(θ , p |θ, p) in the case of the mixture parameters, with the same irrelevance of the normalizing constants. Continuation of Example 6.1. If we consider once more the posterior associated with (6.3), we can check in Fig. 6.8 the cumulative eﬀect of a small

200

6 Mixture Models

2

2

2 μ2

4

μ2

4

μ2

4

0

0

0

−2

−2

−2

−2

0

2 μ1

4

−2

0

2 μ1

4

−2

0

2

4

μ1

Fig. 6.8. Comparison of Metropolis–Hastings samples of 104 points started in the vicinity of the spurious mode for the target distributions πα when α = 1, 0.1, 0.01 (from left to right), π is the same as in Fig. 6.6, and the proposal is a random walk with variance 0.1 (the shape of log-likelihood does not changed)

variance for the random walk proposal (chosen here as 0.1) and a decrease in the power α. The R function used to produce this ﬁgure is hmmeantemp=function(dat,niter=100,var=.1,alpha=1){ mu=matrix(0,niter,2) mu[1,]=c(1,3) for (i in 2:niter){ muprop=rnorm(2,mu[i-1,],sqrt(var)) bound=lpost(dat,muprop)-lpost(dat,mu[i-1,]) if (runif(1) > > > + + + + + >

simu=gibbsnorm(1000,mix) lopos=order(simu$lopost)[1000] lnum1=simu$lolik[lopos] lnum2=sum(dnorm(simu$mu[lopos,], mean=mean(datha),sd=simu$sig[lopos,],log=TRUE)+ dgamma(1/simu$sig[lopos,],10,var(datha),log=TRUE)2*log(simu$sig[lopos,]))+ sum((rep(0.5,k)-1)*log(simu$p[lopos,]))+ lgamma(sum(rep(0.5,k)))-sum(lgamma(rep(0.5,k))) lchibapprox2=lnum1+lnum2-log(simu$deno)

we obtain Table 6.2 which gives the approximations of the marginal likelihoods from k = 2 to k = 8. For the License dataset, the favored number of components is thus k = 4.

k 2 3 4 5 6 ρˆk (x) −5373.445 −5315.351 −5308.79 −5336.23 −5341.524 Table 6.2. Dataset License: estimations of the marginal log-likelihoods by the symmetrized Chib’s approximation

6.9 Exercises 6.1 Show that a mixture of Bernoulli distributions is again a Bernoulli distribution. Extend this to the case of multinomial distributions. 6.2 Show that the number of nonnegative integer solutions of the decomposition of n into k parts such that n1 + . . . + nk is equal to n+k−1 r= . n

6.9

Exercises

207

Deduce that the number of partition sets is of order O(nk−1 ). (Hint: This is a classical combinatoric problem.) 6.3 For a mixture of two normal distributions with all parameters unknown, pN (μ1 , σ12 ) + (1 − p)N (μ2 , σ22 ) , and for the prior distribution (j = 1, 2) μj |σj ∼ N (ξj , σj2 /nj ) ,

σj2 ∼ I G (νj /2, s2j /2) ,

p ∼ Be(α, β) ,

show that μj |σj , x, z ∼ N

p|x, z ∼ Be(α + 1 , β + 2 ), σj2 , σj2 |x, z ∼ I G ((νj + j )/2, sj (z)/2) , ξ1 (z), nj + j

¯j (z) and sˆ2j (z) are the empirical mean and where j is the number of zi equal to j, x variance for the subsample with zi equal to j, and ξj (z) =

nj ξj + j x ¯j (z) , nj + j

sj (z) = s2j + j sˆ2j (z) +

nj j (ξj − x ¯j (z))2 . nj + j

Compute the corresponding weight ω(z). 6.4 For the normal mixture model of Exercise 6.3, compute the function Q(θ0 , θ) and derive both steps of the EM algorithm. Apply this algorithm to a simulated dataset and test the inﬂuence of the starting point θ0 . 6.5 In the mixture model with independent priors on the θj ’s, show that the θj ’s are dependent on each other given (only) x by summing out the z’s. 6.6 Construct and test the Gibbs sampler associated with the (ξ, μ0 ) parameterization of (6.3), when μ1 = μ0 − ξ and μ2 = μ0 + ξ. 6.7 Show that, if an exchangeable prior π is used on the vector of weights (p1 , . . . , pk ), then, necessarily, Eπ [pj ] = 1/k and, if the prior on the other parameters (θ1 , . . . , θk ) is also exchangeable, then Eπ [pj |x1 , . . . , xn ] = 1/k for all j’s. 6.8 Show that running an MCMC algorithm with target π(θ|x)γ will increase the proximity to the MAP estimate when γ > 1 is large. (Note: This is a crude version of the simulated annealing algorithm. See also Chap. 8.) Discuss the modiﬁcations required in Algorithm 6.11 to achieve simulation from π(θ|x)γ when γ ∈ N∗ is an integer. 6.9 Show that the ratio (6.7) goes to 1 when α goes to 0 when the proposal q is a random walk. Describe the average behavior of this ratio in the case of an independent proposal. 6.10 If one needs to use importance sampling weights, show that the simultaneous choice of several powers α requires the computation of the normalizing constant of πα . 6.11 In the setting of the mean mixture (6.3), run an MCMC simulation experiment to compare the inﬂuence of a N (0, 100) and of a N (0, 10000) prior on (μ1 , μ2 ) on a sample of 500 observations. 6.12 Show that, for a normal mixture 0.5 N (0, 1) + 0.5 N (μ, σ 2 ), the likelihood is unbounded. Exhibit this feature by plotting the likelihood of a simulated sample using the R image procedure.

7 Time Series

Rebus was intrigued by the long gaps in the chronology. —Ian Rankin, The Falls.—

Roadmap At one point or another, everyone has to face modeling time series datasets, by which we mean series of dependent observations that are indexed by time (like both series in the picture above!). As in the previous chapters, the diﬃculty in modeling such datasets is to balance the complexity of the representation of the dependence structure against the estimation of the corresponding model—and thus the modeling most often involves model choice or model comparison. We cover here the Bayesian processing of some of the most standard time series models, namely the autoregressive and moving average models, as well as extensions that are more complex to handle like stochastic volatility models used in ﬁnance. This chapter also covers the more complex dependence structure found in hidden Markov models, while spatial dependence in considered in Chap. 8. The reader should be aware that, due to mathematical constraints related to the long-term stability of the series, this chapter contains more advanced material, although we restrained from introducing complex simulation procedures on variable-dimension spaces. J.-M. Marin and C.P. Robert, Bayesian Essentials with R, Springer Texts in Statistics, DOI 10.1007/978-1-4614-8687-9 7, © Springer Science+Business Media New York 2014

209

210

7 Time Series

7.1 Time-Indexed Data While we started with independent (and even iid) observations, for the obvious reason that they are easier to process, we soon departed from this setup, gathering more complexity either through heterogeneity as in the linear and generalized linear models (Chaps. 3 and 4), or through some dependence structure as in the open capture–recapture models of Chap. 5 that pertain to the generic notion of hidden Markov models covered in Sect. 7.5. 7.1.1 Setting This chapter concentrates on time-series (or dynamic) models, which somehow appear to be simpler because they are unidimensional in their dependence, being indexed only by time. Their mathematical validation and estimation are however not so simple, while they are some of the most commonly used models in applications, ranging from ﬁnance and economics to reliability, to medical experiments, and ecology. This is the case, for instance, for series of pollution data, such as ozone concentration levels, or stock market prices, whose value at time t depends at least on the previous value at time t − 1. The dataset we use in this chapter is a collection of four time series connected with the stock market. Figure 7.1 plots the successive values from January 1, 1998, to November 9, 2003, of those four stocks1 which are the ﬁrst ones (in alphabetical order) to appear in the ﬁnancial index Eurostoxx50, a ﬁnancial reference for the euro zone2 made of 50 major stocks. These four series constitute the Eurostoxx50 dataset. A perusal of these graphs is suﬃcient for rejecting the assumption of independence of these series: High values are followed by high values and small values by small values, even though the variability (or volatility) of the stocks varies from share to share. The simplest mathematical structure for a time series is when the series (xt ) is Markov. We recall that a stochastic process (xt )t∈T , that is, a sequence of random variables indexed by the t’s in T (where, here, T is equal to N or Z) is a Markov chain when the distribution of xt conditional on the past values (for instance, x0:(t−1) = (x0 . . . , xt−1 ) when T = N) only depends on xt−1 . This process is homogeneous if the distribution of xt conditional on the past

1 The four stocks are as follows. ABN Amro is an international bank from Holland. Aegon is a Dutch insurance company. Ahold Kon., namely Koninklijke Ahold N.V., is also a Dutch company, dealing in retail and food-service businesses. Air Liquide is a French company specializing in industrial and medical gases. 2 At the present time, the euro zone is made up of the following countries: Austria, Belgium, Finland, France, Germany, Greece, Holland, Ireland, Italy, Portugal, and Spain.

7.1

Time-Indexed Data

211

AEGON

ABN AMRO HOLDING 30 50 25

40 30

20

20

15

10 10 0

500

1000

1500

0

500

t AHOLD KON.

1000

1500

1000

1500

t AIR LIQUIDE 160 150

30

140 130

20 120 110

10

100

0

500

1000

1500

0

500

t

t

Fig. 7.1. Dataset Eurostoxx50: Evolution of the ﬁrst four stocks over the period January 1, 1998 to November 9, 2003

is constant in t ∈ T . Thus, given an observed sequence x0:T = (x0 , . . . , xT ) from a homogeneous Markov chain, the associated likelihood is given by (θ|x0:T ) = f0 (x0 |θ)

T

f (xt |xt−1 , θ) ,

t=1

where f0 is the distribution of the starting value x0 . From a Bayesian point of view, this likelihood can be processed almost as in an iid model once a prior distribution on θ is chosen. However, a generic time series may be represented in formally the same way, namely through the full conditionals as in (θ|x0:T ) = f0 (x0 |θ)

T

ft (xt |x0:(t−1) , θ) .

(7.1)

t=1

When this function can be obtained in a closed form, a Bayesian analysis is equally possible.

212

7 Time Series

Note that general time-series models can often be represented as Markov models via the inclusion of missing variables and an increase in the dimension of the model. This is called a state-space representation. 7.1.2 Stability of Time Series While we pointed out above that, once the likelihood function is written down, the Bayesian processing of the model is the same as in the iid case,3 there exists a major diﬀerence that leads to a more delicate determination of the corresponding prior distributions in that the new properties of stationarity and causality constraints must often be accounted for. We cannot embark here on a mathematically rigorous coverage of stationarity for stochastic processes or even for time series (see Brockwell and Davis, 1996), thus simply mention (and motivate) below the constraints found in the time series literature. A stochastic process (xt )t∈T is stationary4 if the joint distributions of (x1 , . . . , xk ) and (x1+h , . . . , xk+h ) are the same for all indices k and k + h in T . Formally, this property is called strict stationarity because there exists an alternative version of stationarity, called second-order stationarity. This alternative imposes invariance in time only on ﬁrst and second moments of the process. If we deﬁne the autocovariance function γx (·, ·) of the process (xt )t∈T by γx (r, s) = E[{xr − E(xr )}{xs − E(xs )}],

r, s ∈ T ,

namely the covariance between xr and xs , cov(xr , xs ), assuming that the variance V(xt ) is ﬁnite, a process (xt )t∈T with ﬁnite second moments is secondorder stationary if E(xt ) = μ

and γx (r, s) = γx (r + t, s + t)

for all r, s, t ∈ T . If (xt )t∈T is second-order stationary, then γx (r, s) = γx (|r − s|, 0) for all r, s ∈ T . It is therefore convenient to redeﬁne the autocovariance function of a second-order stationary process as a function of just one variable; i.e., with a slight abuse of notation, γx (h) ≡ γx (h, 0),

h∈T .

3 In the sense that, once a closed form of the posterior is available as in (7.1), there exist generic simulation techniques that do not take into account the dynamic structure of the model. 4 The connection with the stationarity requirement of MCMC methods is that these methods produce a Markov kernel such that, when the Markov chain is started at time t = 0 from the target distribution π, the whole sequence (xt )t∈N is stationary with marginal distribution π.

7.1

Time-Indexed Data

213

The function γx (·) is called the autocovariance function of (xt )t∈T , and γx (h) is said to be the autocovariance “at lag” h. The autocorrelation function is implemented in R as acf(), already used in Chaps. 4 and 5 for computing the eﬀective sample size of an MCMC sample. By default, the function acf() returns 10 log10 (m) autocorrelations when applied to a series (vector) of size m, the autocovariances being obtained with the option type="covariance", and it also produces a graph of those autocorrelations unless the option plot=FALSE is activated. An illustration of acf() for the ABN Amro stock series is given by > > > > > > > >

data(Eurostoxx50) abnamro=Eurostoxx50[,2] abnamro=ts(abnamro,freq=365-55*2,start=1998) par(mfrow=c(2,2),mar=c(4,4,1,1)) plot.ts(abnamro,col="steelblue") acf(abnamro,lag=365-55*2) plot.ts(diff(abnamro),col="steelblue") acf(diff(abnamro))

whose graphical output is given in Fig. 7.2. The ts function turns the vector of ABN Amro stocks into a time series, which explains for the years on the ﬁrst axis in plot.ts and the relative values on the ﬁrst axis in acf, where 1.0 corresponds to a whole year. (The range of a year is computed by adding six bank holidays per year to the weekend breaks.) The second row corresponds to the time-series representation of the ﬁrst diﬀerence (xt+1 − xt ), a standard approach used to remove the clear lack of stationarity of the original series. The diﬀerence in the autocorrelation graphs is striking: in particular, the complete lack of signiﬁcant autocorrelation in the ﬁrst diﬀerence is indicative of a random walk behavior for the original series. Obviously, strict stationarity is stronger than second-order stationarity, and this feature somehow seems more logical from a Bayesian viewpoint as it is a property of the whole model.5 For a process (xt )t∈N , this property relates to the distribution f0 of the starting values. From a Bayesian point of view, to impose the stationarity condition on a model (or rather on its parameters) is however objectionable on the grounds that the data themselves should indicate whether or not the underlying model is stationary. In addition, since the datasets we consider are always ﬁnite, the stationarity requirement is at best artiﬁcial in practice. For instance, the series in Fig. 7.1 are clearly not stationary on the temporal scale against which they are plotted. However, for reasons ranging from asymptotics (Bayes estimators are not necessarily convergent in nonstationary settings) to causality, 5

Nonetheless, there exists a huge amount of literature on the study of time series based only on second-moment assumptions.

214

7 Time Series

Fig. 7.2. Representations (left) of the time series ABN Amro (top) and of its ﬁrst diﬀerence (bottom), along with the corresponding acf graphs (right)

to identiﬁability (see below), and to common practice, it is customary to impose stationarity constraints, possibly on transformed data, even though a Bayesian inference on a nonstationary process could be conducted in principle. The practical diﬃculty is that, for complex models, the stationarity constraints may get quite involved and may even be unknown in some cases, as for some threshold or changepoint models. We will expose (and solve) this diﬃculty in the following sections.

7.2 Autoregressive (AR) Models In this section, we consider one of the most common (linear) time series models, the AR(p) model, along with its Bayesian analyses and its Markov connections (which can be exploited in some MCMC implementations).

7.2

Autoregressive (AR) Models

215

7.2.1 The Models An AR(1) process (xt )t∈Z (where AR stands for autoregressive) is deﬁned by the conditional relation (t ∈ Z), xt = μ + (xt−1 − μ) + t ,

(7.2)

where (t )t∈Z is an iid sequence of random variables with mean 0 and variance σ 2 (that is, a so-called white noise). Unless otherwise speciﬁed, we will only consider the t ’s to be iid N (0, σ 2 ) variables.6 If || < 1, (xt )t∈Z can be written as xt = μ +

∞

j t−j ,

(7.3)

j=0

and it is easy to see that this is a unique second-order stationary representation. More surprisingly, if || > 1, the unique second-order stationary representation of (7.2) is ∞ −j t+j . xt = μ − j=1

This stationary solution is frequently criticized as being artiﬁcial because it implies that xt is correlated with the future white noises (t )s>t , a property not shared by (7.3) when || < 1. While mathematically correct, the fact that xt appears as a weighted sum of random variables that are generated after time t is indeed quite peculiar, and it is thus customary to restrict the deﬁnition of AR(1) processes to the case || < 1 so that xt has a representation in terms of the past realizations (t )s≤t . Formally, this restriction corresponds to so-called causal or future-independent autoregressive processes.7 Notice that the causal constraint for the AR(1) model can be naturally associated with a uniform prior on (−1, 1). Note that, when we replace the above normal sequence (t ) with another white noise sequence, it is possible to express an AR(1) process with || > 1 as an AR(1) process with || < 1. However, this modiﬁcation is not helpful from a Bayesian point of view because of the complex distribution of the transformed white noise.

6 Once again, there exists a statistical approach that leaves the distribution of the t ’s unspeciﬁed and only works with ﬁrst and second moments. But this perspective is clearly inappropriate within the Bayesian framework, which cannot really work with half-speciﬁed models. 7 Both stationary solutions above exclude the case || = 1. This is because the process (7.2) is then a random walk with no stationary solution.

216

7 Time Series

A natural generalization of the AR(1) model is obtained by increasing the lag dependence on the past values. An AR(p) process is thus deﬁned by the conditional (against the past) representation (t ∈ Z), xt = μ +

p

i (xt−i − μ) + t ,

(7.4)

i=1

where (t )t∈Z is a white noise. As above, we will assume implicitly that the white noise is normally distributed. This natural generalization assumes that the p most recent values of the process inﬂuence (linearly) the current value of the process. As for the AR(1) model, stationarity and causality constraints can be imposed on this model. A lack of stationarity of a time series theoretically implies that the series ultimately diverges to ±∞. An illustration of this property is provided by the following R code, which produces four AR(10) series of 260 points based on the same t ’s when the coeﬃcients i are uniform over (−.5, .5). The ﬁrst and the last series either have coeﬃcients that satisfy the stationarity conditions or have not yet exhibited a divergent trend. Both remaining series clearly exhibit divergence. > > > > > + + + + +

p=10 T=260 dat=seqz=rnorm(T) par(mfrow=c(2,2),mar=c(2,2,1,1)) for (i in 1:4){ coef=runif(p,min=-.5,max=.5) for (t in ((p+1):T)) seqz[t]=sum(coef*seqz[(t-p):(t-1)])+dat[t] plot(seqz,ty="l",col="sienna",lwd=2,ylab="") }

As shown in Brockwell and Davis (1996, Theorem 3.1.1), the AR(p) process (7.4) is both causal and second-order stationary if and only if the roots of the polynomial p i u i (7.5) P(u) = 1 − i=1

are all outside the unit circle in the complex plane. (Remember that polynomials of degree p always have p roots, but that some of those roots may be complex numbers.) While this necessary and suﬃcient condition on the parameters i is clearly deﬁned, it also imposes an implicit constraint on the vector = (1 , . . . , p ). Indeed, in order to verify that a given vector satisﬁes this condition, one needs ﬁrst to ﬁnd the roots of the pth degree polynomial P and then to check that these roots all are of modulus larger than 1. In other words, there is no clearly deﬁned boundary on the parameter space to deﬁne

7.2

Autoregressive (AR) Models

217

Fig. 7.3. Four simulation of an AR(10) series of 260 points when based on the same standard normal perturbations t and when the coeﬃcients i are uniform over (−.5, .5)

the ’s that satisfy (or do not satisfy) this constraint, and this creates a major diﬃculty for simulation applications, given that simulated values of need to be tested one at a time. For instance, the R code > maxi=0 > for (i in (1:10^6)) maxi=maxi+ + (max(Mod(polyroot(c(1,runif(10,-.5,.5)))))>1) > maxi/10^6 [1] 1 shows that no simulation out of one million simulated coeﬃcients for the AR(10) model that satisfy the constraint. It is therefore very likely that all series in Fig. 7.3 are non-stationary. Note that the general AR(p) model is Markov, just like the AR(1) model, because the distribution of xt+1 only depends on a ﬁxed number of past

218

7 Time Series

values. It can thus be expressed as a regular Markov chain when considering the vector, for t ≥ p − 1, T

zt = (xt , xt−1 , . . . , xt+1−p ) = xt:(t+1−p) . Indeed, we can write zt+1 = μ1p + B(zt − μ1p ) + εt+1 ,

(7.6)

where ⎛

1p = (1, . . . , 1)T ∈ Rp ,

1 2 3 . . . p−2 p−1 ⎜1 0 ... ⎜ ⎜ 0 B = ⎜ 0 1 0 ... 0 ⎜ .. ⎝. ...

0

0

1

⎞ p 0⎟ ⎟ 0⎟ ⎟, .. ⎟ .⎠ 0

and εt = (t , 0, . . . , 0)T . If we now consider the likelihood associated with a series x0:T of observations from a Gaussian AR(p) process, it depends on the unobserved values x−p , . . . , x−1 since (μ, 1 , . . . , p , σ|x0:T , x−p:−1 ) ∝ ⎫ ⎧ 9 :2 p T ⎬ ⎨

exp − xt − μ − i (xt−i − μ) σ −T −1 2σ 2 . ⎭ ⎩ t=0

i=1

These unobserved initial values can be processed in various ways that we now describe. First, they can all be set equal to μ, but this is a purely computational convenience with no justiﬁcation. Second, if the stationarity and causality constraints hold, the process (xt )t∈Z has a stationary distribution and one can assume that x−p:−1 is distributed from the corresponding stationary distribution, namely a Np (μ1p , A) distribution. We can then integrate those initial values out to obtain the marginal likelihood ⎧ 2 ⎫ p T ⎬ ⎨ −1 σ −T −1 exp −μ− (x −μ) x f (x−p:−1 |μ, A) dx−p:−1 , t i t−i ⎭ ⎩ 2σ 2 t=0

i=1

based on the argument that they are not directly observed. This likelihood can be dealt with analytically but is more easily processed via a Gibbs sampler that simulates the initial values. An alternative and equally coherent approach is to consider instead the likelihood conditional on the initial observed values x0:(p−1) ; that is,

7.2

Autoregressive (AR) Models

219

c (μ, 1 , . . . , p , σ|xp:T , x0:(p−1) ) ∝ ⎫ ⎧ 9 :2 p T ⎬ ⎨

2σ 2 . (7.7) exp − xt − μ − i (xt−i − μ) σ −T +p−1 ⎭ ⎩ t=p

i=1

Unless speciﬁed otherwise, we will adopt this approach. In this case, if we do not restrict the parameter space through stationarity conditions, a natural conjugate prior can be found for the parameter θ = (μ, , σ 2 ), made up of a normal distribution on (μ, ) and an inverse gamma distribution on σ 2 . Instead of the Jeﬀreys prior, which is controversial in this setting (see Robert, 2007, Note 4.7.2), we can also propose a more traditional noninformative prior such as π(θ) = 1/σ 2 . 7.2.2 Exploring the Parameter Space by MCMC Algorithms If we do impose the causal stationarity constraint on that all the roots of P in (7.5) be outside the unit circle, the set of acceptable ’s becomes quite involved and we cannot, for instance, use as prior distribution a normal distribution restricted to this set, if only because we lack a simple algorithm to properly describe the set. While a feasible solution is based on the partial autocorrelations of the AR(p) process (see Robert, 2007, Sect. 4.5.2), we cover here a diﬀerent and somehow simpler reparameterization approach using the inverses of the real and complex roots of the polynomial P, which are within the unit interval (−1, 1) and the unit sphere, respectively. Because of this unusual structure of the parameter space, involving two subsets of completely diﬀerent natures, we introduce an MCMC algorithm that could be related with birth and death processes and simulation in variable dimension spaces. If we represent the polynomial (7.5) in its factorized form P(x) =

p

(1 − λi x) ,

i=1

the inverse roots, λi (i = 1, . . . , p), are either real numbers or complex conjugates.8 Under the causal stationarity constraint, a natural prior is then to use uniform priors for these roots, taking a uniform distribution on the number rp of conjugate complex roots and uniform distributions on [−1, 1] and on the unit sphere S = {λ ∈ C; |λ| ≤ 1} for the real and nonconjugate complex roots, respectively. In other words, π(λ) =

1 1 1 I|λi |1){ for (i in 2:pr) Psi[2:(i+1),i]=Psi[2:(i+1),i-1]-lr[i]*Psi[1:i,i-1] } } if (pc>0){ if (pr>0){ Psi[2,pr+2]=-2*lc[1]+Psi[2,pr] Psi[3:(pr+3),pr+2]=(lc[1]^2+lc[2]^2)*Psi[1:(pr+1),pr] -2*lc[1]*Psi[2:(pr+2),pr]+Psi[3:(pr+3),pr] }else{ Psi[2,2]=-2*lc[1]; Psi[3,2]=(lc[1]^2+lc[2]^2); } if (pc>2){ for (i in seq(4,pc,2)){ pri=pr+i prim=pri-2 Psi[2,pri]=-2*lc[i-1]+Psi[2,prim]

7.2

Autoregressive (AR) Models

221

Psi[3:(pri+1),pri]=(lc[i-1]^2+lc[i]^2)*Psi[1:(pri-1), prim]-2*lc[i-1]*Psi[2:pri,prim]+Psi[3:(pri+1),prim] } } } Rho=Psi[1:(p+1),p] where the ρi ’s are the opposites of the components of Rho[2:p]. The log-likelihood (7.7) is then derived in a straightforward manner: x=x-mu loglike=0 for (i in (p+1):T) loglike=loglike-(t(Rho)%*%x[i:(i-p)])^2 loglike=(loglike/sig2-(T-p)*log(sig2))/2 x=x+mu Since the conditional likelihood function (7.7) is a standard Gaussian likelihood in both μ and σ, we can directly use a Gibbs sampler on those parameters and opt for a Metropolis-within-Gibbs step on the remaining (inverse) roots of P, λ = (λ1 , . . . , λp ). A potentially ineﬃcient9 if straightforward Metropolis–Hastings implementation is to use the prior distribution π(λ) itself as a proposal on λ. This means selecting ﬁrst one or several roots of P, λ1 , . . . , λq (1 ≤ q ≤ p), and then proposing new values for these roots that are simulated from the prior, λ1 , . . . , λq ∼ π(λ). (Reordering the roots so that the modiﬁed values are the ﬁrst ones is not restrictive since both the prior and the likelihood are permutation invariant.) The acceptance ratio then simpliﬁes into the likelihood ratio by virtue of Bayes’ theorem: (μ, λ , σ|xp:T , x0:(p−1) ) (μ, λ , σ|xp:T , x0:(p−1) ) π(μ, λ , σ) π(λ) = (μ, λ, σ|xp:T , x0:(p−1) ) π(μ, λ, σ) π(λ ) (μ, λ, σ|xp:T , x0:(p−1) ) The main diﬃculty with this scheme is that one must take care to modify complex roots by (conjugate) pairs. This means, for instance, that to create a complex root (and its conjugate) either another complex root (and its conjugate) or two real roots must be chosen and modiﬁed. Formally, this is automatically satisﬁed by simulations from the prior (7.8). One possible algorithmic representation is therefore:

9 Simulating from the prior distribution when aiming at the posterior distribution is inevitably leading to a waste of simulations if the data is informative about the parameters. The solution is of course unavailable when the prior is improper.

222

7 Time Series

Algorithm 7.13 Metropolis–Hastings AR(p) Sampler Initialization: Choose λ(0) , μ(0) , and σ (0) . Iteration t (t ≥ 1): 1. Select one root at random. If the root is real, generate a new real root from the prior distribution. Otherwise, generate a new complex root from the prior distribution and update the conjugate root. Replace λ(t−1) with λ using these new values. Calculate the corresponding = (1 , . . . , p ). Take ξ = λ with probability c (μ(t−1) , , σ (t−1) |xp:T , x0:(p−1) ) ∧1, c (μ(t−1) , (t−1) , σ (t−1) |xp:T , x0:(p−1) ) and ξ = λ(t−1) otherwise. 2. Select two real roots or two complex conjugate roots at random. If the roots are real, generate a new complex root from the prior distribution and compute the conjugate root. Otherwise, generate two new real roots from the prior distribution. Replace ξ with λ using these new values. Calculate the corresponding = (1 , . . . , p ). Accept λ(t) = λ with probability c (μ(t−1) , , σ (t−1) |xp:T , x0:(p−1) ) ∧1, c (μ(t−1) , (t−1) , σ (t−1) |xp:T , x0:(p−1) ) and set λ(t) = ξ otherwise. 3. Generate μ by a random walk proposal. Accept μ(t) = μ with probability c (μ , (t) , σ (t−1) |xp:T , x0:(p−1) ) ∧1, c (μ(t−1) , (t) , σ (t−1) |xp:T , x0:(p−1) ) and set μ(t) = μ(t−1) otherwise. 4. Generate σ by a log-random walk proposal. Accept σ (t) = σ with probability c (μ(t) , (t) , σ |xp:T , x0:(p−1) ) c (μ(t) , (t) , σ (t−1) |xp:T , x0:(p−1) ) and set σ (t) = σ (t−1) otherwise.

∧1,

7.2

Autoregressive (AR) Models

223

While the whole R code is too long (300 lines) to be reproduced here, the core part about the modiﬁcation of the roots can be implemented as follows: “down” moves removing one pair of complex roots are chosen with probability 0.1 while “up” moves creating one pair of complex roots are chosen with probability 0.9, in order to compensate for the inherently higher diﬃculty in accepting complex proposals from the prior. Those uneven weights must then be accounted for in the acceptance probability, along with the changes in the masses of the uniform priors on the real and complex roots in (7.8). if (runif(1)0){ lambpropcomp=lambdacomp[((1:pcomp)[-(ind:(ind+1))] }else{ #no complex root lambpropcomp=0 #dummy necessary for ARllog function } lambpropreal=c(lambdareal,2*runif(2)-1) coef=9*(1+(preal0){ lambpropreal=lambdareal[(1:preal)[-ind]] }else{ lambpropreal=0 #dummy necessary for ARllog function } theta=2*pi*runif(1); rho=sqrt(runif(1)) lambpropcomp=c(lambdacomp,rho*cos(theta),rho*sin(theta)) coef=(4/pi)*(1+(ppropcomp ar(x = Eurostoxx50[, 4], method = "ml") Coefficients: 1 2 3 4 5 6 1.042 -0.080 -0.038 0.080 -0.049 0.006 Order selected 8 sigma^2 estimated as 0.3228

7 0.080

8 -0.043

Figure 7.4 summarizes the MCMC output for 50,000 iterations. The top left graph shows that jumps between 2 and 0 complex roots occur with high frequency and therefore that the MCMC algorithm mixes well between both (sub)models. Both following graphs on the ﬁrst row relate to the hyperparameters μ and σ, which are updated outside the reversible jump steps. The parameter μ appears to be mixing better than σ, which is certainly due to the choice of the same scaling factor in both cases. The middle rows correspond to the ﬁrst three coeﬃcients of the autoregressive model, 1 , 2 , 3 . Their stability is a good indicator of the convergence of the reversible jump algorithm. Note also that, except for 1 , the other coeﬃcients are close to 0 (since their posterior means are approximately 0.052, −0.0001, 2.99 × 10−5 , and −2.66 × 10−7 , respectively). The ﬁnal row is an assessment of the ﬁt of the model and the convergence of the MCMC algorithm. The ﬁrst graph provides the sequence of corresponding log-likelihoods, which remain stable almost from the start, the second the distribution of the complex (inverse) roots, and the last one the connection between the actual series and its one-step-ahead prediction E[Xt+1 |xt , xt−1 , . . .]: On this scale, both series are well-related. While the above algorithm is a regular Metropolis–Hastings algorithm on a parameter space with a ﬁxed number of parameters, 1 , . . . , p , the

7.2

Autoregressive (AR) Models

225

0.34 2

30

0.32 0.30

0

0

1

25

2000

2

σ

μ

3

4000

4

6000

35

0.36

8000

real–complex dichotomy gives us the opportunity to mention a new class of MCMC algorithms, variable dimension MCMC algorithms. The class of variable dimension models is made of models characterized by a collection of submodels, Mk , often nested, that are considered simultaneously and associated with diﬀerent parameter spaces. The number of submodels can be inﬁnite, and the “parameter” is deﬁned conditionally on the index of the submodel, θ = (k, θk ), with a dimension that generally depends on k. It naturally occurs in settings like Bayesian model choice and Bayesian model assessment. Inference on such structures is obviously more complicated than on single models, especially when there are an inﬁnite number of submodels, and it can

−1

0

1

2

3

4

0

5

2000

4000

6000

8000

10000

2000

4000

6000

8000

10000

8000

10000

Iterations

Iterations

0.01 0.00

0.1 0 2000

4000

6000

8000

ρ −0.02

− 0.0 5

−1.10

0.0 0

−0.01

ρ

0.0 5

−1.05

ρ

−1.00

0.15

p

10000

2000

4000

8000

10000

2000

4000

6000

Iterations

30

0

x −0.15

− 3000

10

−0.05

20

Im ( λ )

0.05

−1000 − 2000

log−likelihood

6000

Iterations 0.15

Iterations

0

2000

4000

6000

Iterations

8000

10000

−0.2

0.0

0.2

0.4

Re(λ)

0.6

0.8

1.0

0

500

1000

1500

t

Fig. 7.4. Dataset Eurostoxx50: Output of the MCMC algorithm for the Ahold Kon. series and an AR(5) model: (top row, left) histogram and sequence of numbers of complex roots (ranging from 0 to 4), (top row, middle and right) sequence of μ and σ 2 , (middle row) sequences of i (i = 1, 2, 3), (bottom row, left) sequence of observed log-likelihood, (bottom row, middle) representation of the cloud of complex roots, with a part of the boundary of the unit circle on the right, (bottom row, right) comparison of the series and the one-step-ahead prediction

226

7 Time Series

be tackled from two diﬀerent (or even opposite) perspectives. The ﬁrst approach is to consider the variable dimension model as a whole and to estimate quantities that are meaningful for the whole model (such as moments or predictives) as well as quantities that only make sense for submodels (such as posterior probabilities of submodels and posterior moments of θk ). From a Bayesian perspective, once a prior is deﬁned on θ, the only diﬃculty is in ﬁnding an eﬃcient way to explore the complex parameter space in order to produce these estimators. The second perspective on variable dimension models is to resort to testing, rather than estimation, by adopting a model choice stance. This requires choosing among all possible submodels the “best one” in terms of an appropriate criterion. The drawbacks of this second approach are far from benign. The computational burden may be overwhelming when the number of models is inﬁnite, the interpretation of the selected model is delicate and the variability of the resulting inference is underestimated since it is impossible to include the eﬀect of the selection of the model in the assessment of the variability of the estimators built in later stages. Nonetheless, this is an approach often used in linear and generalized linear models (Chaps. 3 and 4) where subgroups of covariates are compared against a given dataset. It is obviously the recommended approach when the number of models is small, as in the mixture case (Chap. 6) or in the selection of the order p of an AR(p) model, provided the Bayes factors can be approximated. MCMC algorithms that can handle such variable-dimension structures are facing measure theoretic diﬃculties and, while a universal and elegant solution through reversible jump algorithms exists (Green, 1995), we have made the choice of not covering these in this book. An introductory coverage can be found in the earlier edition (Marin and Robert, 2007, Sect. 6.7), as well as in Robert and Casella (2004). Nonetheless, we want to point out that the above MCMC algorithm happens to be a special case of the birth-and-death MCMC algorithm (and of its generalization, the reversible jump algorithm) where, in nested models, additional components are generated from the prior distribution and the move to a larger model is accepted with a probability equal to the ratio of the likelihoods (with the proper reweighting to account for the multiplicity of possible moves). For instance, extending the above algorithm to the case of the unknown order p is straightforward.

7.3 Moving Average (MA) Models A second type of time series model that still enjoys linear dependence and closed-form expression is the MA(q) model, where MA stands for moving average. It appears as a dual version of the AR(p) model. An MA(1) process (xt )t∈Z is such that, conditionally on the past xt = μ + t − ϑt−1 ,

t∈T ,

(7.9)

7.3

Moving Average (MA) Models

227

where (t )t∈T is a white noise sequence. For the same reasons as above, we will assume the white noise is normally distributed unless otherwise speciﬁed. Thus, E[xt ] = μ , V(xt ) = (1 + ϑ2 )σ 2 , γx (1) = ϑσ 2 ,

and γx (h) = 0 (h > 1) .

An important feature of (7.9) is that the model is not identiﬁable per se. Indeed, we can also rewrite xt as xt = μ + ˜t−1 −

1 t , ˜ ϑ

˜ ∼ N (0, ϑ2 σ 2 ) .

Therefore, both pairs (ϑ, σ) and (1/ϑ, ϑσ) are equivalent representations of the same model. To achieve identiﬁability, it is therefore customary in (nonBayesian environments) to restrict the parameter space of MA(1) processes by |ϑ| < 1 , and we will follow suit. Such processes are called invertible. As with causality, the property of inversibility is not a property of the sole process (xt )t∈Z but of the connection between the two processes (xt )t∈T and (t )t∈T . A natural extension of the MA(1) model is to increase the dependence on the past innovations, namely to introduce the MA(q) process as the process (xt )t∈T deﬁned by q ϑi t−i , (7.10) xt = μ + t − i=1

where (t )t∈T is a white noise (once again assumed to be normal unless otherwise speciﬁed). The corresponding identiﬁability condition in this model is that the roots of the polynomial Q(u) = 1 −

q

ϑi ui

i=1

are all outside the unit circle in the complex plane (see Brockwell and Davis, 1996, Theorem 3.1.2, for a proof). Thus, we end up with exactly the same parameter space as in the AR(q) case! The intuition behind the MA(q) representation is however less straightforward than the regression structure underlying the AR(p) model. This representation assumes that the dependence between observables stems from a dependence between the (unobserved) noises rather than directly through the observables. Furthermore, in contrast with the AR(p) models, where the covariance between the terms of the series is exponentially decreasing to zero but always diﬀerent from 0, the autocovariance function for the MA(q) model is such that γx (s) is equal to 0 for |s| > q, meaning that xt+s and xt are independent. In addition, the MA(q) process is obviously (second-order and strictly) stationary, whatever the vector (ϑ1 , . . . , ϑq ), since the white noise is

228

7 Time Series

iid and the distribution of (7.10) is thus independent of t. A major diﬀerence between the MA(q) and the AR(p) models, though, is that the MA(q) dependence structure is not Markov (even though it can be represented as a Markov process through a state-space representation, introduced below). While, in the Gaussian case, the whole (observed) vector x1:T is a realization of a normal random variable, with constant mean μ and covariance matrix Σ, and thus provides a formally explicit likelihood function, both the computation and the integration (or maximization) of this likelihood are quite costly since they involve inverting the huge matrix Σ.10 A more manageable representation of the MA(q) likelihood is to use the likelihood of x1:T conditional on the past white noises 0 , . . . , −q+1 , c (μ, ϑ1 , . . . , ϑq , σ|x1:T , (−q+1):0 ) ∝ ⎧ ⎛ ⎫ ⎞2 ⎪ ⎪ q T ⎨ ⎬ σ −T exp − ⎝xt − μ + ϑj ˆt−j ⎠ 2σ 2 , ⎪ ⎪ ⎩ ⎭ t=1 j=1

(7.11)

where ˆ0 = 0 ,. . ., ˆ1−q = 1−q and (t > 0) t = xt − μ + ˆ

q

ϑj ˆt−j .

j=1

This recursive deﬁnition of the likelihood is still costly since it involves T sums of q terms. Nonetheless, even though the problem of handling the conditioning values (−q+1):0 must be treated separately via an MCMC step, the complexity O(T q) of this representation is much more manageable than the normal exact representation mentioned above. Since the transform of the roots into the coeﬃcients is exactly the same as with the AR(q) model, the expression of the log-likelihood function conditional on the past white noises eps is quite straightforward. Taking for Psi the subvector Psi[2:(p+1),p], the computation goes as follows: x=x-mu # construction of the epsilonhats heps=rep(0,T+q) heps[1:q]=eps # past noises for (i in 1:T) heps[p+i]=x[i]+sum(rev(Psi)*heps[i:(q+i-1)]) # completed loglikelihood (includes negative epsilons) loglike=-((sum(heps^2)/sig2)+(T+q)*log(sig2))/2 x=x+mu 10 Obviously, taking advantage of the block diagonal structure of Σ—due to the fact that γx (s) = 0 for |s| > q— may reduce the computational cost, but this requires advanced programming abilities!

7.3

Moving Average (MA) Models

229

Given both x1:T and the past noises (−q+1):0 , the conditional posterior distribution of the parameters (μ, ϑ1 , . . . , ϑq , σ) is formally very close to the posterior associated with an AR(q) posterior distribution. This proximity is such that we can recycle the code of Algorithm 7.13 to some extent since the simulation of the (inverse) roots of the polynomial Q is identical once we modify the likelihood according to the above changes. The past noises −i (i = 1, . . . , q) are simulated conditional both on the xt ’s and on the parameters μ, σ and ϑ = (ϑ1 , . . . , ϑq ). While the exact distribution f ((−q+1):0 |x1:T , μ, σ, ϑ) ∝

0 i=−q+1

2

e−i /2σ

2

T

2

2

e−ˆt /2σ ,

(7.12)

t=1

where the ˆt ’s are deﬁned as above, is exactly a normal distribution on the vector (−q+1):0 (Exercise 7.13), its computation is too costly to be available for realistic values of T . We therefore implement a hybrid Gibbs algorithm where the missing noise (−q+1):0 is simulated from a proposal based either on the previous simulated value of (−q+1):0 (in which case we use a simple termwise random walk) or on the ﬁrst part of (7.12) (in which case we can use normal proposals).11 More speciﬁcally, one can express ˆt (1 ≤ t ≤ q) in terms of the −t ’s and derive the corresponding (conditional) normal distribution on either each −t or on the whole vector 12 (see Exercise 7.14). The additional step, when compared with the AR(p) function, is the conditional simulation of the past noises (−q+1):0 . For 1 ≤ i ≤ q (the indices are drifted to start at 1 rather than −q), the corresponding part of our R code is as follows. Unfortunately, the derivation of the Metropolis–Hastings acceptance probability does require computing the inverse ˆ−t ’s as they are functions of the proposed noises. x=x-mu heps[1:q]=eps # simulated ones for (j in (q+1):(2*p+1)) # epsilon hat heps[j]=x[j]+sum(rev(Psi)*heps[(j-q):(j-1)]) heps[i]=0 for (j in 1:(q-i+1)) keps[j]=x[j]+sum(rev(Psi)*heps[j:(j+q-1)]) x=x+mu 11 In the following output analysis, we actually used a more hybrid proposal with the innovations ˆ t ’s (1 ≤ t ≤ q) ﬁxed at their previous values. This approximation remains valid when accounted for in the Metropolis–Hastings acceptance ratio, which requires computing the ˆt ’s associated with the proposed −i . 12 Using the horizon t = q is perfectly sensible in this setting given that x1 , . . . , xq are the only observations correlated with the −t ’s, even though (7.11) gives the impression of the opposite, since all ˆt ’s depend on the −t ’s.

230

7 Time Series

epsvar=1/sum(c(1,Psi[i:q]^2)) epsmean=sum(Psi[i:q]*keps[1:(q-i+1)])*epsvar epsmean=epsmean/epsvar epsvar=sig2*epsvar propeps=rnorm(1,mean=epsmean,sd=sqrt(epsvar)) epspr=eps epspr[i]=propeps lloprop=MAllog(pr=preal,pc=pcomp,lr=lambdareal, lc=lambdacomp,mu=mu,sig2=sig2,compsi=FALSE,pepsi=Psi, eps=epspr) propsal1=dnorm(propeps,mean=epsmean,sd=sqrt(epsvar),log=TRUE) x=x-mu heps[i]=propeps for (j in (q+1):(2*q+1)) heps[j]=x[j]+sum(rev(Psi)*heps[(j-q):(j-1)]) heps[i]=0 for (j in 1:(q-i+1)) keps[j]=x[j]+sum(rev(Psi)*heps[j:(j+q-1)]) x=x+mu epsvar=1/sum(c(1,Psi[i:q]^2)) epsmean=sum(Psi[i:q]*keps[1:(q-i+1)]) epsmean=epsmean*epsvar epsvar=sig2*epsvar propsal0=dnorm(eps[i],mean=epsmean,sd=sqrt(epsvar),log=TRUE) if (log(runif(1)) arima(x = Eurostoxx50[1:350, 5], order = c(0, 0, 9)) Coefficients: ma1 ma2 ma3 ma4 ma5 ma6 ma7 1.0605 0.9949 0.9652 0.8542 0.8148 0.7486 0.5574 s.e. 0.0531 0.0760 0.0881 0.0930 0.0886 0.0827 0.0774 ma8 ma9 intercept 0.3386 0.1300 114.3146 s.e. 0.0664 0.0516 1.1281 sigma^2 estimated as 8.15: log likelihood = -864.97 The favored number of complex roots is 6, and the smaller values 0 and 2 are not visited after the initial warmup. The mixing over the σ parameter is again lower than over the mean μ, despite the use of three diﬀerent proposals. The ﬁrst one is based on the inverted gamma distribution associated with > −(q−1):q , the second one is based on a (log) random walk with scale 0.1ˆ σx , and the third one is an independent inverted gamma distribution with scale σ ˆx /(1 + ϑ21 + 2 1/2 . . . + ϑq ) . Note also that, except for ϑ9 , the other coeﬃcients ϑi are quite diﬀerent from 0 (since their posterior means are approximately 1.0206, 0.8403, 0.8149, 0.6869, 0.6969, 0.5693, 0.2889, and 0.0895, respectively). This is also the case for the estimates above obtained in R arima. The prediction being of little interest for MA models (Exercise 7.15), we represent instead the range of simulated t ’s in the bottom right ﬁgure. The range is compatible with the N (0, σ 2 ) distribution.

0

30 σ2 20 10

15

μ

112 110

114

2 0

2000

4

4000

25

6

116

6000

35

118

7 Time Series 8

232

0

2

4

6

8

2000

4000

p

6000

8000

10000

2000

6000

8000

10000

8000

10000

8000

10000

−0.4 θ

10000

−0.8 −1.2 2000

4000

6000

8000

10000

2000

4000

Iterations

6000 Iterations

10 5 −5

0

ε

0.0

−15

−550

−0.5

−500

Im(λ)

−450

0.5

15

−400

20

Iterations

log−likelihood

8000

0.0

0.0 −0.4 θ −0.8 −1.2 4000

6000 Iterations

−0.6 −0.8 θ −1.0 −1.2 2000

4000

Iterations

2000

4000

6000

Iterations

8000

10000

−0.5

0.0

0.5

Re(λ)

2000

4000

6000

sube(start:W)

Fig. 7.5. Dataset Eurostoxx50: Output of the MCMC algorithm for the Air Liquide series and an MA(9) model: (top row, left) histogram and sequence of numbers of complex roots (ranging from 0 to 8); (top row, middle and right) sequence of μ and σ 2 ; (middle row) sequences of ϑi (i = 1, 2, 3); (bottom row, left) sequence of observed likelihood; (bottom row, middle) representation of the cloud of complex roots, with the boundary of the unit circle; and (bottom row, right) evolution of the simulated −t ’s

7.4 ARMA Models and Other Extensions An alternative approach that is of considerable interest for the representation and analysis of the MA(q) model and its generalizations is the so-called statespace representation, which relies on missing variables to recover both the Markov structure and the linear framework.13 The general idea is to represent a time series (xt ) as a system of two equations, 13

It is also inspired from the Kalman ﬁlter, ubiquitous for prediction, smoothing, and ﬁltering in time series.

7.4

ARMA Models and Other Extensions

233

xt = Gyt + εt ,

(7.13)

yt+1 = F yt + ξt ,

(7.14)

where εt and ξ t are multivariate normal vectors14 with general covariance matrices that may depend on t and E[εT u ξv ] = 0 for all (u, v)’s. Equation (7.13) is called the observation equation, while (7.14) is called the state equation. This representation embeds the process of interest (xt ) into a larger space, the state space, where the missing process (yt ) is Markov and linear. For instance, (7.6) is a state-space representation of the AR(p) model (see Exercise 7.16). The MA(q) model can be written that way by deﬁning yt as yt = (t−q , . . . , t−1 , t )T . Then the state equation is ⎛

yt+1

01 ⎜0 0 ⎜ =⎜ ⎜ ⎝0 0 00

0 ... 1 ... ... 0 ... 0 ...

⎛ ⎞ ⎞ 0 0 ⎜0⎟ ⎟ 0⎟ ⎜ ⎟ .⎟ ⎟ yt + t+1 ⎜ ⎜ .. ⎟ , ⎟ ⎟ ⎜ 1⎠ ⎝0⎠ 0 1

(7.15)

while the observation equation is xt = xt = μ − ϑq ϑq−1 . . . ϑ1 −1 yt , with no perturbation εt . The state-space decomposition of the MA(q) model thus involves no vector εt in the observation equation, while ξ t is degenerate in the state equation. The degeneracy phenomenon is quite common in state-space representations, but this is not a hindrance in conditional uses of the model, as in MCMC implementations. Notice also that the state-space representation of a model is not unique, again a harmless feature for MCMC uses. For instance, for the MA(1) model, the observation equation can also be chosen as xt = μ+(1 0)yt with yt = (y1t , y2t )T directed by the state equation 01 1 yt+1 = yt + t+1 . 00 −ϑ1 Note that, while the state-space representation is wide-ranging and convenient, it does not mean that the derived MCMC strategies are necessarily eﬃcient. In particular, when the hidden state xt is too large, a na¨ıve completion may prove itself disastrous. Alternative solutions based in sequential importance sampling (SMC) have been shown to be usually more eﬃcient. (See Del Moral et al., 2006.) 14

Notice the diﬀerent fonts that distinguish the εt ’s used in the state-space representation from the t ’s used in the AR and MA models.

234

7 Time Series

A straightforward extension of both previous AR and MA models are the (normal) ARMA(p, q) models, where xt (t ∈ Z) is conditionally deﬁned by xt = μ −

p i=1

i (xt−i − μ) + t −

q

ϑj t−j ,

t ∼ N (0, σ 2 ) ,

(7.16)

j=1

the (t )’s being independent. The role of such models, as compared with both AR and MA models, is to aim toward parsimony; that is, to resort to much smaller values of p and q than in a pure AR(p) or a pure MA(q) modeling. The causality and inversibility conditions on the parameters of (7.16) still correspond to the roots of both polynomials P and Q being outside the unit circle, respectively, with a further condition that both polynomials have no common root. (But this almost surely never happens under a continuous prior on the parameters.) The root reparameterization can therefore be implemented for both the ϑj ’s and the i ’s, still calling for MCMC techniques owing to the complexity of the posterior distribution. State-space representations also exist for ARMA(p, q) models, one possibility being xt = xt = μ − ϑr−1 ϑr−2 . . . ϑ1 −1 yt for the observation equation and ⎛ 0 1 0 ⎜0 0 1 ⎜ yt+1 = ⎜ ⎜ ⎝0 0 0 r r−1 r−2

... ... ... ... ...

⎛ ⎞ ⎞ 0 0 ⎜0⎟ ⎟ 0⎟ ⎜ ⎟ .⎟ ⎟ yt + t+1 ⎜ ⎜ .. ⎟ ⎟ ⎜ ⎟ 1⎠ ⎝0⎠ 1 1

(7.17)

for the state equation, with r = max(p, q + 1) and the convention that m = 0 if m > p and ϑm = 0 if m > q. Similarly to the MA(q) case, this state-space representation is handy in devising MCMC algorithms that converge to the posterior distribution of the parameters of the ARMA(p, q) model. A straightforward MCMC processing of the ARMA model is to take advantage of the AR and MA algorithms that have been constructed above by using both algorithms sequentially. Indeed, conditionally on the AR parameters, the ARMA model can be expressed as an MA model and, conversely, conditionally on the MA parameters, the ARMA model can be expressed almost as an AR model. This is quite obvious for the MA part since, if we deﬁne (t > p) p i (xt−i − μ) , x˜t = xt − μ + i=1

the likelihood is formally equal to a standard MA(q) likelihood on the x ˜t ’s. The reconstitution of the AR(p) likelihood is more involved: If we now deﬁne q the residuals ˜t = j=1 ϑj t−j , the log-likelihood conditional on x0:(p−1) is

7.4

−

T

⎛ ⎝xt − μ −

t=p

ARMA Models and Other Extensions

p

⎞2 j [xt−j − μ] − ˜t ⎠

235

2σ 2 ,

j=1

which is obviously close to an AR(p) log-likelihood, except for the ˜t ’s. The original AR(p) MCMC code can then be recycled modulo this modiﬁcation in the likelihood. Another extension of the AR model is the ARCH model, used to represent processes, particularly in ﬁnance, with independent errors but time-dependent variances, as in the ARCH(p) process15 (t ∈ Z) xt = σt t ,

iid

t ∼ N (0, 1) ,

σt2 = α +

p

βi x2t−i .

i=1

The ARCH(p) process deﬁnes a Markov chain since xt only depends on xt−p:t−1 . It can be shown that a stationarity condition for the ARCH(1) model is that E[log(β1 2t )] < 0, which is equivalent to β1 < 3.4. This condition becomes much more involved for larger values of p. Contrary to the stochastic volatility model deﬁned below, the ARCH(p) model enjoys a closed-form likelihood when conditioning on the initial values x1 , . . . , xp . However, because of the nonlinearities in the variance terms, approximate methods based on MCMC algorithms must be used for their analysis. State-space models are special cases of hidden Markov models (detailed below in Sect. 7.5) in the sense that (7.13) and (7.14) are a special occurrence of the generic representation xt = G(yt , t ) , yt = F (yt−1 , ζt ) .

(7.18)

Note, however, that it is not necessarily appealing to resort to this hidden Markov representation, in comparison with state-space models, because the complexity of the functions F or G may hinder the processing of this representation to unbearable levels (while, for state-space models, the linearity of the relations always allows for a generic if not necessarily eﬃcient processing based on, e.g., Gibbs sampling steps). Stochastic volatility models are quite popular in ﬁnancial applications, especially in describing series with sudden and correlated changes in the magnitude of variation of the observed values. These models use a hidden chain (yt )t∈N , called the stochastic volatility, to model the variance of the observables (xt )t∈N in the following way: Let y0 ∼ N (0, σ 2 ) and, for t = 1, . . . , T , deﬁne 15

The acronym ARCH stands for autoregressive conditional heteroscedasticity, heteroscedasticity being a term favored by econometricians to describe heterogeneous variances. Gouri´eroux (1996) provides a general reference on these models, as well as classical inferential methods of estimation.

236

7 Time Series

yt = ϕyt−1 + σ∗t−1 , xt = βeyt /2 t ,

(7.19)

where both t and ∗t are iid N (0, 1) random variables. In this simple version, the observable is thus a white noise, except that the variance of this noise enjoys a particular AR(1) structure on the logarithmic scale. Quite obviously, this structure makes the computation of the (observed) likelihood a formidable challenge! Figure 7.6 gives the sequence {log(xt ) − log(xt−1 )} when (xt ) is the Aegon stock sequence plotted in Fig. 7.1. While this real-life sequence is not necessarily a stochastic volatility process, it presents some features that are common with those processes, including an overall stationary structure and periods in the magnitude of the variation of the sequence.

xt–xt–1

–0.15

–0.10

–0.05

0.00

0.05

0.10

0.15

0.20

AEGON

0

500

1000

1500

t

Fig. 7.6. Dataset Eurostoxx50: First-order diﬀerence {log(xt ) − log(xt−1 )} of the Aegon stock sequence regarded as a potential stochastic volatility process (7.19)

When comparing ARMA with the hidden Markov models of the following Section, it may appear that the former are more general in the sense that they allow a diﬀerent dependence on the past values. Resorting to the statespace representation (7.18) shows that this is not the case. Diﬀerent horizons p of dependence can also be included for hidden Markov models simply by (a) using a vector xt = (xt−p+1 , . . . , xt ) for the observables or by (b) using a vector yt = (yt−q+1 , . . . , yt ) for the latent process in (7.18).

7.5 Hidden Markov Models Hidden Markov models are a generalization of the mixture models of Chap. 6. Their appeal within this chapter is that they constitute an interesting case of non-Markov time series, besides being extremely useful in modeling, e.g.,

7.5

Hidden Markov Models

237

for ﬁnancial, telecommunication, and genetic data. We refer the reader to McDonald and Zucchini (1997) for a deeper introduction to these models and to Capp´e et al. (2004) and Fr¨ uhwirth-Schnatter (2006) for a complete coverage of their statistical processing. 7.5.1 Basics The family of hidden Markov models (abbreviated to HMM) consists of a bivariate process (xt , yt )t∈N , where the unobserved subprocess (yt )t∈N is a homogeneous Markov chain on a state space Y and, conditional on (yt )t∈N , (xt )t∈N is a series of random variables on X such that the conditional distribution of xt given yt and the past (xj , yj )j 0 , and compute the corresponding suﬃcient statistics. Iteration m (m ≥ 1): 1. Generate (pi1 , . . . , piκ ) ∼ D(1 + ni1 , . . . , 1 + niκ ) , (q1i , . . . , qki ) ∼ D(1 + mi1 , . . . , 1 + mik ) , and correct for the missing initial probability by a Metropolis–Hastings step with acceptance probability y0 /y0 . 2. Generate successively each yt (0 ≤ t ≤ T ) by if t = 0 , pii qxi 1 piy1 P(yt = i|xt , yt−1 , yt+1 ) ∝ pyt−1 i qxi t piyt+1 if t > 0 , and compute the corresponding suﬃcient statistics.

In the initialization step of Algorithm 7.15, any distribution on (yt )t∈N is obviously valid, but this particular choice is of interest since it is related to the true conditional distribution, simply omitting the dependence on the next value. The main loop in the Gibbs sampler is then of the form (for κ = 2 and k = 4 as in Dnadataset) # Beta/Dirichlet simulations for P a=1/(1+rgamma(1,nab+1)/rgamma(1,naa+1)) b=1/(1+rgamma(1,nba+1)/rgamma(1,nbb+1)) P=matrix(c(a,1-a,1-b,b),ncol=2,byrow=T) q1=rgamma(4,ma+1) # and Q q2=rgamma(4,mb+1) q1=q1/sum(q1); q2=q2/sum(q2) # (hidden) Markov conditioning x[1]=sample(1:2,1,prob=c(a*P[1,x[2]]*q1[y[1]],b*P[2,x[2]]

7.5

Hidden Markov Models

241

*q2[y[1]])) for (m in 2:(T-1)) x[m]=sample(1:2,1,prob=c(P[x[m-1],1]*P[1,x[m+1]]*q1[y[m]], P[x[m-1],2]*P[2,x[m+1]]*q2[y[m]])) x[T]=sample(1:2,1,prob=c(P[x[T-1],1]*q1[y[T]],P[x[T-1],2] *q2[y[T]])) # Sufficient statistics for next iteration naa=sum((x[1:(T-1)]==1)*(x[2:T]==1)) nab=sum((x[1:(T-1)]==1)*(x[2:T]==2)) nba=sum((x[1:(T-1)]==2)*(x[2:T]==1)) nbb=sum((x[1:(T-1)]==2)*(x[2:T]==2)) ya=y[x==1] ma=c(sum(ya==1),sum(ya==2),sum(ya==3),sum(ya==4)) yb=y[x==2] mb=c(sum(yb==1),sum(yb==2),sum(yb==3),sum(yb==4)) We ran several Gibbs samplers for 1,000 iterations, starting from small, medium and high values for p11 and p22 , and got very similar results in both ﬁrst and both last cases for the approximations to the Bayes posterior means, as shown by Table 7.1. The raw output also gives a sense of stability, as shown by Fig. 7.9. For the third case, started at small values of both p11 and p22 , the simulated chain had not visited the same region of the posterior distribution after those 1,000 iterations, and it produced an estimate with a smaller loglikelihood17 value of −13, 160. However, running the Gibbs sampler longer (for 4,000 more iterations) did produce a similar estimate, as shown by the third replication in Table 7.1. This phenomenon is slightly related to the phenomenon, discussed in the context of Figs. 6.4 and 6.3, that the Gibbs sampler tends to “stick” to lower modes for lack of suﬃcient energy. In the current situation, the energy required to leave the lower mode appears to be available. Note that we have reordered the output to compensate for a possible switch between hidden states 1 and 2 among experiments. This is quite natural, given the lack of identiﬁability of the hidden states (Exercise 7.17). Flipping the indices 1 and 2 does not modify the likelihood, and thus all these experiments explore the same mode of the posterior.

7.5.2 Forward–Backward Representation When the state space of the hidden Markov chain Y is ﬁnite, that is, when Y = {1, . . . , κ} , 17

The log-posterior is proportional to the log-likelihood in that special case, and the log-likelihood is computed using a technique described below in Sect. 7.5.2.

242

7 Time Series 0.9 0.8 0.7 0.6 0.5

0.85 0.80 0.75 0.70 0.65 0

200

400

600

800

1000

0

200

400

p11

600

800

1000

600

800

1000

600

800

1000

600

800

1000

p22 0.25 0.20 0.15 0.10 0.05 0.00

0.40 0.35 0.30 0

200

400

600

800

1000

0

200

400

q11

q21 0.40 0.35 0.30 0.25 0.20 0.15 0.10

0.45 0.40 0.35 0.30 0.25 0

200

400

600

800

1000

0

200

400

q31

q12

0.5 0.4 0.3 0.2 0.1

0.25 0.20 0.15 0.10 0.05 0.00 0

200

400

600

800

1000

0

200

q22

400 q23

Fig. 7.9. Dataset Dnadataset: Convergence of a Gibbs sequence to the region of interest on the posterior surface for the hidden Markov model (this is replication 2 in Table 7.1). The row-wise order of the parameters is the same as in Table 7.1 Table 7.1. Dataset Dnadataset: Five runs of the Gibbs sampling approximations to the Bayes estimates of the parameters for the hidden Markov model along with ﬁnal log-likelihood (starting values are indicated on the line below in parentheses) based on M = 1000 iterations (except for replication 3, based on 5,000 iterations) Run p11 1 0.720 (0.844) 0.662 2 (0.628) 0.696 3 (0.055) 0.704 4 (0.915) 0.694 5 (0.600)

p22 0.581 (0.885) 0.620 (0.621) 0.609 (0.150) 0.580 (0.610) 0.585 (0.516)

q11 0.381 (0.260) 0.374 (0.203) 0.376 (0.293) 0.377 (0.237) 0.376 (0.296)

q21 0.032 (0.281) 0.016 (0.352) 0.023 (0.200) 0.024 (0.219) 0.0218 (0.255)

q31 0.396 (0.279) 0.423 (0.199) 0.401 (0.232) 0.407 (0.228) 0.410 (0.288)

q12 0.306 (0.087) 0.317 (0.066) 0.318 (0.150) 0.313 (0.079) 0.315 (0.110)

q22 0.406 (0.094) 0.381 (0.114) 0.389 (0.102) 0.403 (0.073) 0.395 (0.095)

q32 0.018 (0.0937) 0.034 (0.0645) 0.030 (0.119) 0.020 (0.076) 0.0245 (0.107)

Log-like. –13,121 –13,123 –13,118 –13,121 –13,119

the likelihood function18 of the observed process (xt )1≤t≤T can be computed in a manageable O(T × κ2 ) time by a recurrence relation called the forward–

18

To lighten notation, we will not use the parameters appearing in the various distributions of the HMM, even though they are obviously of central interest.

7.5

Hidden Markov Models

243

backward or Baum–Welch formulas.19 We now explain how those formulas are derived. As illustrated in Fig. 7.7, a generic feature of HMMs is that (t = 2, . . . , T ) p(yt |yt−1 , x0:T ) = p(yt |yt−1 , xt:T ) . In other words, knowledge of the past observations is redundant for the distribution of the hidden Markov chain when we condition on its previous value. Therefore, when Y is ﬁnite, we can write that p(yT |yT −1 , x0:T ) ∝ pyT −1 yT f (xT |yT ) ≡ pT (yT |yT −1 , x0:T ) , meaning that we deﬁne pT (yT |yT −1 , x0:T ) as the unnormalized version of the density p(yT |yT −1 , x0:T ). Then we can process backward the deﬁnition of the previous conditionals, so that (1 < t < T ) p(yt |yt−1 , x0:T ) = ∝ =

κ i=1 κ i=1 κ

p(yt , yt+1 = i|yt−1 , xt:T ) p(yt , yt+1 = i, xt:T |yt−1 ) p(yt |yt−1 )f (xt |yt )p(yt+1 = i, x(t+1):T |yt )

i=1

∝ pyt−1 yt f (xt |yt ) ∝ pyt−1 yt f (xt |yt )

κ i=1 κ

p(yt+1 = i|yt , x(t+1):T ) pt+1 (i|yt , x1:T ) ≡ pt (yt |yt−1 , x1:T ) .

i=1

At last, the conditional distribution of the ﬁrst hidden value y0 is p(y0 |x0:T ) ∝ y0 f (x0 |y0 )

κ

p1 (i|y0 , x0:t ) ≡ p0 (y0 |x0:T ) ,

i=1

where (k )k is the stationary distribution associated with the Markov transition matrix P. (This is unless the ﬁrst hidden value y0 is automatically set equal to 1 for identiﬁability reasons.) While this construction amounts to a straightforward conditioning argument, the use of the unnormalized functions pt+1 (yt+1 = i|yt , x(1:T ) is crucial for deriving the joint conditional distribution of y1:T since resorting to the normalized conditionals instead would result in a useless identity. 19 This recurrence relation has been known for quite a while in the signal processing literature and is also used in the corresponding EM algorithm; see Capp´e et al. (2004) for details.

244

7 Time Series

Notice that, as stated above, the derivation of the pt ’s indeed has a cost of O(T ×κ2 ) since, for each t and each of the κ values of yt , a sum of κ terms has to be computed. So, in terms of raw computational time, computing the observed likelihood does not take less time than simulating the sequence (yt )t∈N in the Gibbs sampler. However, the gain in using this forward–backward formula may impact in subtler ways a resulting Metropolis–Hastings algorithm, such as a better mixing of the chain of the parameters, given that we are simulating the whole vector at once. Once we have all the conditioning functions (or backward equations), it is possible to simulate sequentially the hidden sequence y0:T given x0:T by generating ﬁrst y0 from p(y0 |x0:T ), second y1 from p(y1 |y0 , x0:T ) and so on. However, there is (much) more to be done. Indeed, when considering the joint conditional distribution of y0:T given x0:T , we have p(y0:T |x0:T ) = p(y0 |x0:T )

T

p(yt |yt−1 , x0:T )

t=1

κ T π(y1 )f (x0 |y0 ) pyt−1 yt f (xt |yt ) i=1 pt+1 (i|yt , x1:T ) = κ κ i=1 p0 (i|x0:T ) t=1 i=1 pt (i|yt−1 , x(1:T ) T κ pyt−1 yt f (xt |yt ) p1 (i|x0:T ) = π(y0 )f (x0 |y0 ) t=1

i=1

since all the other sums cancel. This joint conditional distribution immediately leads to the derivation of the observed likelihood since, by Bayes’ formula, κ

f (x0:T ) =

f (x0:T |y1:T ) p(y0:T ) = p1 (i|x0:T ) , p(y0:T |x0:T ) i=1

which is the normalizing constant of the initial conditional distribution! Therefore, working with the unnormalized densities has this supplementary advantage to provide an approximation to the observed likelihood. (Keep in mind that all the expressions above implicitly depend on the model parameters.) A forward derivation of the likelihood can similarly be constructed. Besides the obvious construction that is symmetrical to the previous one, consider the so-called prediction ﬁlter ϕt (i) = P(yt = i|x1:t−1 ) , with ϕ1 (j) = π(j) (where the term prediction refers to the conditioning on the observations prior to time t). The forward equations are then given by (t = 1, . . . , T ) κ 1 f (xt |yt = i)ϕt (i)pij , ϕt+1 (j) = ct i=1

7.5

where ct =

κ

Hidden Markov Models

245

f (xt |yt = k)ϕt (k)

k=1

is the normalizing constant. (This formula uses exactly the same principle as the backward equations.) Exploiting the Markov nature of the joint process (xt , yt )t , we can then derive the log-likelihood as 9 κ : t log p(x1:t ) = log p(xt , yt = i|x1:(r−1) ) r=1

=

t r=1

i=1

log

9 κ

: f (xr |yt = i)ϕr (i) ,

i=1

which also requires a O(T × κ2 ) computational time. The resulting R function for computing the (observed) likelihood is therefore (for κ = 2 and k = 4 as in Dnadataset) likej=function(vec,log=TRUE){ # vec is the aggregated parameter vector P=matrix(c(vec[1],1-vec[1],1-vec[2],vec[2]),ncol=2,byrow= TRUE) Q1=vec[3:6]; Q2=vec[7:10] pxy=c(P[1,1],P[2,2]) pxy=pxy/sum(pxy) # stationary distribution of P pyy=rep(1,T) pyy[1]=pxy[1]*Q1[y[1]]+pxy[2]*Q2[y[1]] for (t in 2:T){ pxy=pxy[1]*Q1[y[t-1]]*P[1,]+pxy[2]*Q2[y[t-1]]*P[2,] pxy=pxy/sum(pxy) pyy[t]=(pxy[1]*Q1[y[t]]+pxy[2]*Q2[y[t]]) } if (log){ ute=sum(log(pyy)) } else{ ute=prod(pyy) } ute } Obviously, to be able to handle directly the observed likelihood when T is reasonable opens new avenues for simulation methods. For instance,

246

7 Time Series

the completion step (of simulating the hidden Markov chain) is no longer necessary, and Metropolis–Hastings alternatives such as random-walk proposals can be used. Returning to Dnadataset, we can compute the log-likelihood (and hence the posterior up to a normalizing constant) associated with a given parameter using, for instance, the prediction ﬁlter. In that case, 9 k : T i log p(x1:T ) = log qxt ϕt (i) , t=1

i=1

where ϕt (j) ∝ 2i=1 qxi t ϕt (i)pij . This representation of the log-likelihood is used in the computation given above for the Gibbs sampler. Furthermore, given that all parameters to be simulated are probabilities, using a normal random walk proposal in the Metropolis–Hastings algorithm is not adequate. Instead, a more appropriate proposal is based on Dirichlet distributions centered at the current value, with scale factor α > 0; that is (j = 1, 2), p˜jj ∼ Be(αpjj , α(1 − pjj )) q˜j ∼ D(αq1j , . . . , αq4j ) . The Metropolis–Hastings acceptance probability is then the ratio of the likelihoods over the ratio of the proposals, f (θ|θ )/f (θ |θ). Since larger values of α produce more local moves, we could test a range of values to determine the “proper” scale. However, this requires a long calibration step. Instead, the algorithm can take advantage of the diﬀerent scales by picking at random for each iteration a value of α from among 1, 10, 100, 10,000 or 100,000. (The randomness in α can then be either ignored in the computation of the proposal density f or integrated by a Rao–Blackwell argument.) For Dnadataset, this range of α’s was wide enough since the average probability of acceptance is 0.25 and a chain (θm )m started at random does converge to the same values as the Gibbs chains simulated above, as shown by Fig. 7.10, which also indicates that more iterations would be necessary to achieve complete stability. We can note in particular that the maximum log-posterior value found along the iterations of the Metropolis–Hastings algorithm is −13,116, which is larger than the values found in Table 7.1 for the Gibbs sampler, for parameter values of (0.70, 0.58, 0.37, 0.011, 0.42, 0.19, 0.32, 0.42, 0.003, 0.26). When the state space Y is ﬁnite, it may be of interest to estimate the order of the hidden Markov chain. For instance, in the case of Dnadataset, it is relevant to infer on how many hidden coding states there are. A possible approach, not covered here, is to use a reversible jump MCMC algorithm that resemble very much the reversible jump algorithm for the mixture model. The reference in this direction is Capp´e et al. (2004, Chap. 16) where the authors construct a reversible jump algorithm in this setting. However, the availability

7.5

Hidden Markov Models

247

0

1000

0.5 0.3 0.1

0.7 0.5 0.3 0

1000

2000

3000

4000

5000

2000

p11

3000

4000

5000

3000

4000

5000

3000

4000

5000

3000

4000

5000

p22

0.40 0.15

0.30 0.20

0.00 0

1000

2000

3000

4000

5000

0

1000

2000

q11

q21 0.25

0.35 0.20

0.10 0

1000

2000

3000

4000

5000

0

1000

2000

q13

q21 0.12

0.5 0.3

0.06

0.1

0.00 0

1000

2000

3000 q22

4000

5000

0

1000

2000

q23

Fig. 7.10. Dataset Dnadataset: Convergence of a Metropolis–Hastings sequence for the hidden Markov model based on 5,000 iterations. The overlayed curve in the background is the sequence of log-posterior values

of the (observed) likelihood means that the marginal solution of Chib (1995), exposed in Chap. 6 (Sect. 6.8) for the mixtures of distributions also applies in the current setting (Exercise 7.19).

Fig. 7.11. DAG representation of the dependence structure of a Markov-switching model where (xt )t is the observable process and (yt )t is the hidden chain

The model ﬁrst introduced for Dnadataset is overly simplistic in that, at least within the coding regime, the xt ’s are not independent. A more realistic modeling thus assumes that the xt ’s constitute a Markov chain within each state of the hidden chain, resulting in the dependence graph of Fig. 7.11. To distinguish this case from the earlier one, it is often called Markov-switching.

248

7 Time Series

This extension is much more versatile than the model of Fig. 7.7, and we can hope to capture the time dependence better. However, it is far from parsimonious, as the use of diﬀerent Markov transition matrices for each hidden state induces an explosion in the number of parameters. For instance, if there are two hidden states, the number of parameters is 26; if there are four hidden states, the number jumps to 60.

7.6 Exercises 7.1 Consider the process (xt )t∈Z deﬁned by xt = a + bt + yt , where (yt )t∈Z is an iid sequence of random variables with mean 0 and variance σ 2 , and where a and b are constants. Deﬁne wt = (2q + 1)−1 qj=−q xt+j . Compute the mean and the autocovariance function of (wt )t∈Z . Show that (wt )t∈Z is not stationary but that its autocovariance function γw (t + h, t) does not depend on t. 7.2 Suppose that the process (xt )t∈N is such that x0 ∼ N (0, τ 2 ) and, for all t ∈ N, xt+1 |x0:t ∼ N (xt /2, σ 2 ) ,

σ > 0.

Give a necessary condition on τ 2 for (xt )t∈N to be a (strictly) stationary process. 7.3 Suppose that (xt )t∈N is a Gaussian random walk on R: x0 ∼ N (0, τ 2 ) and, for all t ∈ N, xt+1 |x0:t ∼ N (xt , σ 2 ) ,

σ > 0.

2

Show that, whatever the value of τ is, (xt )t∈N is not a (strictly) stationary process. 7.4 Give the necessary and suﬃcient condition under which an AR(2) process with autoregressive polynomial P(u) = 1 − 1 u − 2 u2 (with 2 = 0) is causal. 7.5 Consider the process (xt )t∈N such that x0 = 0 and, for all t ∈ N, xt+1 |x0:t ∼ N ( xt , σ 2 ) . Suppose that π(, σ) = 1/σ and that there is no constraint on . Show that the conditional posterior distribution of , conditional on the observations x0:T and on σ 2 , is a N (μT , ωT2 ) distribution with μT =

T

t=1

xt−1 xt

T t=1

x2t−1

and

ωT2 = σ 2

T

x2t−1 .

t=1

Show that the marginal posterior distribution of is a Student T (T − 1, μT , νT2 ) distribution with

7.6 νT2 =

1 T −1

T

x2t

T −1

t=1

Exercises

249

x2t − μ2T

.

t=0

Apply this modeling to the Aegon series in Eurostoxx50 and evaluate its predictive abilities. 7.6 For Algorithm 7.13, show that, if the proposal on σ 2 is a log-normal distribu2 ), τ 2 ) and if the prior distribution on σ 2 is the noninformative prior tion L N (log(σt−1 2 2 π(σ ) = 1/σ , the acceptance ratio also reduces to the likelihood ratio because of the Jacobian. 7.7 Write down the joint distribution of (yt , xt )t∈N in (7.19) and deduce that the (observed) likelihood is not available in closed form. 7.8 Show that the stationary distribution of x−p:−1 in an AR(p) model is a Np (μ1p , A) distribution, and give a ﬁxed point equation satisﬁed by the covariance matrix A. 7.9 Show that the posterior distribution on θ associated with the prior π(θ) = 1/σ 2 and an AR(p) model is well-deﬁned for T > p observations. 7.10 Show that the coeﬃcients of the polynomial P in (7.5) associated with an AR(p) model can be derived in O(p2 ) time from the inverse roots λi using the recurrence relations (i = 1, . . . , p, j = 0, . . . , p) ψ0i = 1 ,

i−1 ψji = ψji−1 − λi ψj−1 ,

where ψ00 = 1 and ψji = 0 for j > i, and setting j = −ψjp (j = 1, . . . , p). 7.11 Given the polynomial P in (7.5), the fact that all the roots are outside the unit circle can be determined without deriving the roots, thanks to the Schur–Cohn test. If Ap = P, a recursive deﬁnition of decreasing degree polynomials is (k = p, . . . , 1) uAk−1 (u) = Ak−1 (u) − ϕk A k (u) , where A k denotes the reciprocal polynomial A k (u) = uk Ak−1 (1/u). 1. Given the expression of ϕk in terms of the coeﬃcients of Ak . 2. Show that the degree of Ak is at most k. 3. If am,k denotes the m-th degree coeﬃcient in Ak , show that ak,k = 0 for k = 0, . . . , p if, and only if, a0,k = ak,k for all k’s. 4. Check by simulation that, in cases when ak,k = 0 for k = 0, . . . , p, the roots are outside the unit circle if, and only if, all the coeﬃcients ak,k are positive. 7.12 For an MA(q) process, show that (s ≤ q) q−|s|

γx (s) = σ 2

ϑi ϑi+|s| .

i=0

7.13 Show that the conditional distribution of (0 , . . . , −q+1 ) given both x1:T and the parameters is a normal distribution. Evaluate the complexity of computing the mean and covariance matrix of this distribution. i ’s. 7.14 Give the conditional distribution of −t given the other −i ’s, x1:T , and the ˆ Show that this distribution only depends on the other −i ’s, x1:q−t+1 , and ˆ1:q−t+1 .

250

7 Time Series

7.15 Show that the (useful) predictive horizon for the MA(q) model is restricted to the ﬁrst q future observations xt+i . 7.16 Show that the system of equations given by (7.13) and (7.14) induces a Markov chain on the completed variable (xt , yt ). Deduce that state-space models are special cases of hidden Markov models. 7.17 Show that, for a hidden Markov model, when the support Y is ﬁnite and when (yt )t∈N is stationary, the marginal distribution of xt is the same mixture distribution for all t’s. Deduce that the same identiﬁability problem as in mixture models occurs in this setting. 7.18 Given a hidden Markov chain (xt , yt ) with both xt and yt taking a ﬁnite number of possible values, k and κ, show that the time required for the simulation of T consecutive observations is in O(kκT ). 7.19 Implement Chib’s method of Sect. 6.8 in the case of a doubly ﬁnite hidden Markov chain. First, show that an equivalent to the approximation (6.9)) is available for the denominator of (6.8). Second, discuss whether or not the label switching issue also rises in this framework. Third, apply this approximation to Dnadataset. 7.20 Show that the counterpart of the prediction ﬁlter in the Markov-switching case is given by κ t

log f (xr |xr−1 , yr = i)ϕr (i) , log p(x1:t ) = r=1

i=1

where ϕr (i) = P(yr = i|x1:r−1 ) is given by the recursive formula ϕr (i) ∝

κ

j=1

pji f (xr−1 |xr−2 , yr−1 = j)ϕr−1 (j) .

8 Image Analysis

“Reduce it to binary, Siobhan,” she told herself. —Ian Rankin, Resurrection Men.—

Roadmap This ﬁnal chapter covers the analysis of pixelized images through Markov random ﬁeld models, towards pattern detection and image correction. We start with the statistical analysis of Markov random ﬁelds, which are extensions of Markov chains to the spatial domain, as they are instrumental in this chapter. This is also the perfect opportunity to cover the ABC method, as these models do not allow for a closed form likelihood. Image analysis has been a very active area for both Bayesian statistics and computational methods in the past 30 years, so we feel it well deserves a chapter of its own for its speciﬁc features.

J.-M. Marin and C.P. Robert, Bayesian Essentials with R, Springer Texts in Statistics, DOI 10.1007/978-1-4614-8687-9 8, © Springer Science+Business Media New York 2014

251

252

8 Image Analysis

8.1 Image Analysis as a Statistical Problem If we think of a computer image as a (large) collection of colored pixels disposed on a grid, there does not seem to be any randomness involved nor any need for statistical analysis! Nonetheless, image analysis seen as a statistical analysis is a thriving ﬁeld that saw the emergence of several major statistical advances, including, for instance, the Gibbs sampler. (Moreover, this ﬁeld has predominantly adopted a Bayesian perspective both because this was a natural thing to do and because the analytical power of this approach was higher than with other methods.) The reason for this apparent paradox is that, while pixels usually are deterministic objects, the complexity and size of images require one to represent those pixels as the random output of a distribution governed by an object of much smaller dimension. For instance, this is the case in computer vision, where speciﬁc objects need to be extracted out of a much richer (or noisier) background. In this spirit of extracting information from huge dimensional structure, we thus build in Sect. 8.2 a speciﬁc family of distributions inspired from particle physics, the Potts model, in order to structure images and other spatial structures in terms of local homogeneity. Unfortunately, this is a mostly theoretical section with very few illustrations. In Sect. 8.3, we address the fundamental issue of handling the missing normalizing constant in these models by introducing a new computational technique called ABC that operates on intractable likelihoods (with the penalty of producing an approximative answer). In Sect. 8.4, we impose a strong spatial dimension on the prior associated with an image in order to gather homogeneous structures out of a complex or blurry image.

8.2 Spatial Dependence 8.2.1 Grids and Lattices An image (in the sense of a computer generated image) is a special case of a lattice, in the sense that it is a random object whose elements are indexed by the location of the pixels and are therefore related by the geographical proximity of those locations. In full generality, a lattice is a mathematical multidimensional object on which a neighbourhood relation can be deﬁned. Even though the original analysis of lattice models by Besag (1974) focussed on plant ecology and agricultural experiments, the neighbourhood relation is only constrained to be a symmetric relation and it does not necessarily have a connection with a geographical proximity, nor with an image. For instance, the relation can describe social interactions between Amazon tribes or words in a manuscript sharing a linguistic root. (The neighbourhood relation between two points of the lattice is generally translated in statistical terms into a probabilistic dependence between those points.) The lattice associated with

8.2

Spatial Dependence

253

an image is a regular n × m array made of (i, j)’s (1 ≤ i ≤ n, 1 ≤ j ≤ m), whose nearest (but not necessarily only) neighbors are made of the four entries (i, j − 1), (i, j + 1), (i − 1, j) and (i + 1, j). In order to properly describe a dependence structure in images or in other spatial objects indexed by a lattice, we need to expand the notion of Markov chain on those structures. Since a lattice is a multidimensional object—as opposed to the unidimensional line corresponding to the times of observation of the Markov chain—, a ﬁrst requirement for the generalization is to deﬁne a proper neighbourhood structure. In order to illustrate this notion, we consider a small dataset1 depicting the presence of tufted sedges2 in a part of a wetland. This dataset, called Laichedata, is simply a 25×25 matrix of zeroes and ones. The corresponding lattice is the 25 × 25 array (Fig. 8.1).

Fig. 8.1. Presence/absence of the tufted sedge plant (Carex elata) on a rectangular patch

Given a lattice I of sites i ∈ I on a map or of pixels in an image,3 a neighbourhood relation on I is denoted by ∼, i ∼ j meaning that i and j are neighbors. If we associate a probability distribution on a vector x indexed by the lattice, x = (xi )i∈I , with this relation, meaning that two components xi and xj are correlated if the sites i and j are neighbors, a fundamental 1

Taken from Gaetan and Guyon (2010), kindly provided by the authors. Wikipedia: “Carex is a genus of plants in the family Cyperaceae, commonly known as sedges. Most (but not all) sedges are found in wetlands, where they are often the dominant vegetation.” Laˆıche is the French for sedge. 3 We will indiscriminately use site and pixel in the remainder of the chapter. 2

254

8 Image Analysis

requirement for the existence of this distribution is that the neighbourhood relation is symmetric (Cressie, 1993): if i is a neighbor of j (written as i ∼ j), then j is a neighbor of i. (By convention, i is not a neighbor of itself.) Figure 8.2 illustrates this notion for three types of neighborhoods on a regular grid. For instance, Laichedata could be associated with a northwest-southeast neighbourhood to account for dominant winds: an entry (i, j) would have as neighbors (i − 1, j − 1) and (i + 1, j + 1).

Fig. 8.2. Some common neighbourhood structures used in imaging, with four (upper left), eight (upper right), or twelve neighbors (lower)

8.2.2 Markov Random Fields A random ﬁeld on I is a random structure indexed by the lattice I, a collection of random variables {xi ; i ∈ I} where each xi takes values in a ﬁnite set χ. Obviously, the interesting case is when the xi ’s are dependent random variables in relation with the neighbourhood structure on I. If n(i) is the set of neighbors of i ∈ I and if xA = {xi ; i ∈ A} denotes the subset of x for indices in a subset A ⊂ I, then xn(i) is the set of values taken by the neighbors of i. The extension from a Markov chain to a Markov random ﬁeld then assumes only dependence on the neighbors.4 More precisely, / A} the coordinates that are not in if, as before, we denote by x−A = {xi ; i ∈ 4

This dependence immediately forces the neighbourhood relation to be symmetric.

8.2

Spatial Dependence

255

a given subset A ⊂ I, a random ﬁeld is a Markov random ﬁeld (MRF) if the conditional distribution of any pixel given the other pixels only depends on the values of the neighbors of that pixel; i.e., for i ∈ I, π(xi |x−i ) = π(xi |xn(i) ) . Markov random ﬁelds have been used for quite a while in imaging, not necessarily because images obey Markov laws but rather because these dependence structures oﬀer highly stabilizing properties in modeling. Indeed, constructing the joint prior distribution of an image is a daunting task because there is no immediate way of describing the global properties of an image via a probability distribution. Just as for the directed acyclic graphs (DAG) models at the core of the BUGS software, using the full conditional distributions breaks the problem down to a sequence of local problems and this is therefore more manageable in the sense that we may be able to express more clearly how we think xi behaves when the conﬁguration of its neighbors is known.5 Before launching into the use of speciﬁc MRFs to describe prior assumptions on a given lattice, we need to worry6 about the very existence of MRFs! Indeed, deﬁning a set of full conditionals does not guarantee that there is a joint distribution behind them (Exercise 8.1). In our case, this means that general forms of neighborhoods and general types of dependences on the neighbors do not usually correspond to a joint distribution on x. We ﬁrst obtain a representation that can be used for testing the existence of a joint distribution. Starting from a complete set of full conditionals on a lattice I, if there indeed exists a corresponding joint distribution, π(x), it is completely deﬁned by the ratio π(x)/π(x∗ ) for a given ﬁxed value x∗ since the normalizing constant is automatically determined. Now, if I = {1, . . . , n}, it is simple to exhibit a full conditional density within the joint density by writing the natural decomposition π(x) = π(x1 |x−1 )π(x−1 ) and then to introduce x∗ by the simple divide-and-multiply trick π(x) =

π(x1 |x−1 ) π(x∗1 , x−1 ) . π(x∗1 |x−1 )

If we iterate this trick for all terms in the lattice (assuming we never divide by 0), we eventually get to the representation 5 It is no surprise that computational techniques such as the Gibbs sampler stemmed from this area, as the use of conditional distributions is deeply ingrained in the imaging community. 6 For those that do not want nor do not need to worry, the end of this section can be skipped, it being of a more theoretical nature and not used in the rest of the chapter.

256

8 Image Analysis n−1 π(xi+1 |x∗ , . . . , x∗ , xi+2 , . . . , xn ) π(x) 1 i = . ∗ |x∗ , . . . , x∗ , x ∗ π(x ) π(x i+2 , . . . , xn ) i+1 1 i i=0

(8.1)

Hence, we can truly write the joint density as a product of ratios of its full conditionals modulo one renormalization.7 This result can also be used toward our purpose of checking for compatibility of the full conditional distributions: if there exists a joint density such that the full conditionals never cancel, then (8.1) must hold for every representation of I = {1, . . . , n}; that is, for every ordering of the indices, and for every choice of reference value x∗ . Although we cannot provide here the reasoning behind the result, there exists a necessary and suﬃcient condition for the existence of an MRF. This condition relies on the notion of clique: Given a lattice I and a neighbourhood relation ∼, a clique is a maximal subset of I made of sites that are all neighbors. The corresponding existence result (Cressie, 1993) is that an MRF associated with I and the neighbourhood relation ∼ necessarily is of the form ΦC (xC ) , (8.2) π(x) ∝ exp − C∈C

where C is the collection of all cliques. This result amounts to saying that the joint distribution must separate in terms of its system of cliques. We now embark on the description of two speciﬁc MRFs that are appropriate for image analysis, namely the Ising model used for binary images and its extension, the Potts model, used for images with more than two colors. 8.2.3 The Ising Model If pixels of the image x under study can only take two colors (black and white, say, as in Fig. 8.1), x is binary. We typically refer to each pixel xi as being foreground if xi = 1 (black) and background if xi = 0 (white). The conditional distribution of a pixel is then Bernoulli, with the corresponding probability parameter depending on the other pixels. A simpliﬁcation step is to assume that it is a function of the number of black neighboring pixels, using for instance a logit link as (j = 0, 1)

π(xi = j|x−i ) ∝ exp(βni,j ) ,

β > 0,

(8.3)

where ni,j = ∈n(i) Ix =j is the number of neighbors of xi with color j. The Ising model is then deﬁned via these full conditionals π(xi = 1|x−i ) =

exp(βni,1 ) , exp(βni,0 ) + exp(βni,1 )

7 This representation is by no means limited to MRFs: it holds for every joint distribution such that the full conditionals never cancel. It is called the Hammersley– Cliﬀord theorem, and a two-dimensional version of it was introduced in Exercise 3.10.

8.2

Spatial Dependence

and the joint distribution therefore satisﬁes ⎞ ⎛ π(x) ∝ exp ⎝β Ixj =xi ⎠ ,

257

(8.4)

j∼i

where the summation is taken over all pairs (i, j) of neighbors (Exercise 8.17). When inferring on β and thus simulating the posterior distribution β, we will be faced with a major obstacle, namely that the normalizing constant of (8.4), Z(β), is intractable except for very small lattices I, while depending on β. Therefore the likelihood function cannot be computed. We will introduce in Sect. 8.3 a computational technique called ABC that is intended to ﬁght this very problem. At this early stage, however, we consider β to be known and focus on the simulation of x in preparation for the inference on both β and x given a noisy version of the image, y, as presented in Sect. 8.4. The computational conundrum of Ising models goes deeper as, due to the convoluted correlation structure of the Ising model, a direct simulation of x is not possible, expect in very speciﬁc cases. Faced with this diﬃculty, the image community very early developed computational tools which eventually led in 1984 to the proposal of the Gibbs sampler (Sect. 3.5.1).8 The speciﬁcation of Markov random ﬁelds and in particular of the Ising model implies the full conditional distributions of those models are available in closed form. The local structure of Markov random ﬁelds thus provides an immediate site-bysite update for the Gibbs sampler: Algorithm 8.16 Ising Gibbs Sampler Initialization: For i ∈ I, generate independently (0)

xi

∼ B(1/2) .

Iteration t (t ≥ 1): 1. Generate u = (ui )i∈I , a random ordering of the elements of I. (t) (t) 2. For 1 ≤ ≤ |I|, update nu ,0 and nu ,1 , and generate x(t) u

∼B

(t)

exp(βnu ,1 ) (t)

(t)

.

exp(βnu ,0 ) + exp(βnu ,1 )

In this implementation, the order of the updates of the pixels of I is random in order to overcome possible bottlenecks in the exploration of the distribu8

The very name “Gibbs sampling” was proposed in reference to Gibbs random ﬁelds, related to the physicist Willard Gibbs. Interestingly, both of the major MCMC algorithms are thus named after physicists and were originally developed for problems that were beyond the boundaries of (standard) statistical inference.

258

8 Image Analysis

tion, although this is not a necessary condition for the algorithm to converge. In fact, when considering two pixels x1 and x2 that are m pixels apart, the inﬂuence of a change in x1 is not felt in x2 before at least m iterations of the basic Gibbs sampler. Of course, if m is large, the dependence between x1 and x2 is quite moderate, but this slow propagation of changes is indicative of slow mixing in the Markov chain. For instance, to see a change of color of a relatively large homogeneous region is an event of very low probability, even though the distribution of the colors is exchangeable (Exercise 8.18).

If β is large, the Ising distribution (8.4) is very peaked around both single color conﬁgurations. In such settings, the Gibbs sampler will face enormous diﬃculties to simply change the value of a single pixel.

Running Algorithm 8.16 in R is straightforward: opting for a four-neighbor relation, if we use the following function for the number of neighbors at (a, b), xneig4=function(x,a,b,col){ n=dim(x)[1];m=dim(x)[2] nei=c(x[a-1,b]==col,x[a,b-1]==col) if (a!=n) nei=c(nei,x[a+1,b]==col) if (b!=m) nei=c(nei,x[a,b+1]==col) sum(nei) } the above Gibbs sampler can be written as isingibbs=function(niter,n,m=n,beta){ # initialization x=sample(c(0,1),n*m,prob=c(0.5,0.5),rep=TRUE) x=matrix(x,n,m) for (i in 1:niter){ sampl1=sample(1:n) sampl2=sample(1:m) for (k in 1:n){ for (l in 1:m){ n0=xneig4(x,sampl1[k],sampl2[l],0) n1=xneig4(x,sampl1[k],sampl2[l],1) x[sampl1[k],sampl2[l]]=sample(c(0,1),1, prob=exp(beta*c(n0,n1))) }}} x } where niter is the number of times the whole matrix x is modiﬁed. (It should therefore be scaled against n*m, the size of x.) Figure 8.3 presents the output of simulations from Algorithm 8.16

8.2 100

100

60

60

20

20

20

40

60

80

100

100

100

60

60

20

20

20

40

60

80

100

100

100

60

60

20

20

20

40

60

80

100

100

100

60

60

20

20

20

40

60

80

100

100

100

60

60

20

20

20

40

60

80

100

Spatial Dependence

259

20

40

60

80

100

20

40

60

80

100

20

40

60

80

100

20

40

60

80

100

20

40

60

80

100

Fig. 8.3. Simulations from the Ising model with a four-neighbor neighbourhood structure on a 100 × 100 array after 1,000 iterations of the Gibbs sampler: β varies in steps of 0.1 from 0.3 to 1.2 (ﬁrst column, then second column)

> image(1:100,1:100,isingibbs(10^3,100,100,beta)) for diﬀerent values of β. Although we cannot discuss here convergence assessment for the Gibbs sampler (see Robert and Casella, 2009, Chap. 8), the images thus produced are representative of the Ising distributions: the larger

260

8 Image Analysis

β, the more homogeneous the image (and also the slower the Gibbs sampler).9 When looking at the result associated with the larger values of β, we can start to see the motivations for using such representations to model images like the Menteith dataset, discussed in Sect. 8.4. Along with the slow dynamic induced by the single-site updating, we can point out another ineﬃciency of this algorithm, namely that many updates will not modify the current value of x simply because the new value of xl is equal to its previous value! It is, however, straightforward to modify the algorithm so that it only proposes changes of values. The update of each pixel l is then a Metropolis–Hastings step with acceptance probability ρ = exp(βnl,1−xl )/ exp(βnl,xl ) ∧ 1 , with the corresponding R function isinghm=function(niter,n,m=n,beta){ x=sample(c(0,1),n*m,prob=c(0.5,0.5),rep=TRUE) x=matrix(x,n,m) for (i in 1:niter){ sampl1=sample(1:n) sampl2=sample(1:m) for (k in 1:n){ for (l in 1:m){ n0=xneig4(x,sampl1[k],sampl2[l],x[sampl1[k],sampl2[l]]) n1=xneig4(x,sampl1[k],sampl2[l],1-x[sampl1[k],sampl2[l]]) if (runif(1) βc , the Markov chain converges to one of two diﬀerent stationary distributions, depending on the starting point. In other words, the chain is no longer irreducible. In particle physics, this phenomenon is called phase transition.

8.2

ni,g =

Spatial Dependence

261

Ixj =g ,

j∼i

the full conditional distribution of xi is chosen as π(xi = g|x−i ) ∝ exp(βni,g ) . This choice corresponds to a (true) joint probability model, the Potts model, whose density is given by (Exercise 8.6) ⎛ ⎞ π(x) ∝ exp ⎝β Ixj =xi ⎠ . (8.5) j∼i

This model is a clear generalization of the Ising model and it suﬀers from the same drawback, namely that the normalizing constant of this density—which is a function of β—is not available in closed form and thus hinders inference and the computation of the likelihood function. Once again, we face the hindrance that, when simulating x from a Potts model with a large β, the single-site Gibbs sampler may be quite slow. More eﬃcient alternatives are available, including the Swendsen–Wang algorithm (Exercise 8.7). For instance, Algorithm 8.17 below is again a Metropolis– Hastings algorithm that forces moves on the current values. Note the special feature that, while this Metropolis–Hastings proposal is not a random walk, using instead a uniform proposal on the G − 1 other possible values still leads to an acceptance probability that is equal to the ratio of the target densities. Algorithm 8.17 Potts Metropolis–Hastings Sampler Initialization: For i ∈ I, generate independently (0)

xi

∼ U ({1, . . . , G}) .

Iteration t (t ≥ 1): 1. Generate u = (ui )i∈I a random ordering of the elements of I. 2. For 1 ≤ ≤ |I|, generate x ˜u ∼ U ({1, 2, . . . , x(t−1) − 1, x(t−1) + 1, . . . , G}) , u u (t)

compute the nul ,g and ( ) ρl = exp(βnu ,˜xu )/ exp(βn(t) u ,xu ) ∧ 1 ,

(t)

and set xu equal to x ˜u with probability ρl .

262

8 Image Analysis

Figure 8.4 illustrates the result of a simulation using Algorithm 8.17 in a situation where there are G = 4 colors, using the following R function pottshm=function(ncol=2,niter=10^4,n,m=n,beta=0){ x=matrix(sample(1:ncol,n*m,rep=TRUE),n,m) for (i in 1:niter){ sampl=sample(1:(n*m)) for (k in 1:(n*m)){ xcur=x[sampl[k]] a=(sampl[k]-1)%%n+1 b=(sampl[k]-1)%/%n+1 xtilde=sample((1:ncol)[-xcur],1) acpt=beta*(xneig4(x,a,b,xtilde)-xneig4(x,a,b,xcur)) if (log(runif(1)) xbar=mean(normaldata) > s2=(n-1)*var(normaldata) > Nsim=10^6 #simulations from the prior > indem=sample(c(0,1),Nsim,rep=TRUE) > ssigma=1/rexp(Nsim) > smu=rnorm(Nsim)*sqrt(ssigma)*(indem==1) > ss2=s2/(ssigma*rchisq(Nsim,n-1)) > sobs=n*(rnorm(Nsim,smu,sqrt(ssigma/n))-xbar)^2+ + ss2-1-log(ss2) > epsi=quantile(sobs,.001) #bound and selection > prob=sum(indem[sobs (1-prob)/prob [1] 0.1574074 producing a numerical value to be compared with the exact Bayes factor −1/2

(n + 1)

n¯ x2 + s2 + 2 n¯ x2 /(n + 1) + s2 + 2

n+2/2

(deduced from the derivation on page 45 by modifying for the exponential prior), which is equal to 0.1369 for normaldata. Figure 8.7 represents the variability of the ABC approximation compared with the true value.

8.3.3 Inference on Potts Models If we consider the speciﬁc case of the posterior distribution associated with (8.6) and a uniform prior, Algorithm 8.18 simulates values of β uniformly over 0.05 0.10 0.15 0.20 0.25 0.30

Fig. 8.7. Dataset normaldata: Boxplot representation of the ABC approximation to the Bayes factor, which true value is represented by an horizontal line, based on 105 proposals, a 1 % acceptance rate, and 500 replications

8.3

Handling the Normalizing Constant

271

(0, 2) and then values x from the Potts model (8.6). Simulating a data set x is unfortunately non-trivial for Markov random ﬁelds and in particular for Potts models, as we already discussed. While there exist developments towards this goal in the special case of the Ising model—in the sense that they produce exact simulations, at a high computing cost—, we settle for using a certain number of steps of an MCMC sampler (for instance, Algorithm 8.17) updating one clique at a time conditional on the others. Obviously, this solution brings a further degree of approximation into the picture in that running a ﬁxed number of iterations of the MCMC sampler does not produce an exact simulation from (8.6). There is however little we can do about this if we want to use ABC. (And we can further argue that ABC involves such a signiﬁcant departure from the exact posterior that an imperfect MCMC simulation does not matter so much!) Since, for every new value of β, the algorithm runs a full MCMC simulation, we need to discuss the choice of the starting value as well. There are (at least) three natural solutions: – start completely at random; – start from the previously simulated x. – always start from the observed value x0 ; The ﬁrst one is the closest to the MCMC idea and it produces independent outcomes. The second solution is less compelling as the continuity it creates between draws is not statistically meaningful, given that the simulated β’s change (independently or not) from one step to the other. The third solution oﬀers the appealing feature of connecting with the observed value x0 , thus favoring proximity between the simulated and the observed values, but this feature could confuse the issues in that this proximity may be due to a poor mixing of the chain rather than to a proper choice for β. (For instance, in the extreme case the MCMC chain does not move from x0 , x = x0 does not mean that the simulated β is at all interesting for π(β|x0 ). . . ) The distance used in step 3 of Algorithm 8.18 is the (natural) absolute diﬀerence between the suﬃcient statistics S(x) and S(x0 ), with Ixi =xj . S(x) = i∼j

For the four-neighbour relation, the statistic can be computed directly without loops as sum(x[-1,]==x[-n,])+sum(x[,-1]==x[,-m]) and the whole R code corresponding to a random start of the Metropolis– Hastings algorithm is as follows: > ncol=4; nrow=10; Nsim=2*10^4; Nmc=10^2 > suf0=sum(x0[-1,]==x0[-nrow,])+sum(x0[,-1]==x0[,-nrow]) > outa=dista=rep(0,Nsim)

272

8 Image Analysis

> for (tt in 1:Nsim){ + beta=runif(1,max=2) + xprop=pottshm(ncol,nit=Nmc,n=nrow,beta=beta) + dista[tt]=abs(suf0-(sum(xprop[-1,]==xprop[-nrow,])+ + sum(xprop[,-1]==xprop[,-ncol]))) + outa[tt]=beta + } betas=outa[order(dista)