Bayesian Models for Categorical Data
So Moon Tong · Journal of the Royal Statistical Society Series A (Statistics in Society) · 2006
This book is aimed at students and researchers, including those with primary disciplinary interests outside statistics. It is also intended to be helpful for courses in Bayesian data analysis and statistical computing. The author claims that the book is to make modern Bayesian methods accessible via practically oriented exposition, with statistical computing and applied data analysis at the forefront. Its format is that of an anthology of journal papers—condensed, technical with limited explanations, that may not be suitable for the targeted audience. It has been designed to pinpoint the computational aspects of Bayesian modelling by using WinBUGS, but very little space is given over to introduce the preliminary concepts of that package. Instead, a set of examples with documented WinBUGS codes that can be downloaded from the publisher's Web site is included. This is a good idea, but it would have been useful to list all the program names, with some interpretations of the computational results. The chapter themes include model selection (Chapter 2), regression models for metric data (Chapter 3) and for count and binomial data (Chapters 4 and 5). Multinomial data, including random-effects and latent variable models, are described in Chapter 6, followed by ordinal data models (Chapter 7), discrete spatial data (Chapter 8), time series models for discrete and clustered variables (multilevel and panel models) (Chapters 9 and 10) and missing data models (Chapter 11). The coverage is at least adequate for Bayesian categorical modelling. Many readers, particularly non-Bayesian statistics users, would feel more comfortable if more space were given to introducing the key concepts of the Bayesian paradigm, such as the Bayes factor and its variants. Partial Bayes factors (fractional or intrinsic and the like) deserve some discussion, especially when the author mentions (page 32) ‘the evaluation of marginal likelihoods may become unstable when vague (non-informative) priors are used on parameters’. In summary, I do not feel that the book succeeds in bringing its material to students and researchers whose primary disciplinary interests lie outside statistics. Rather, it is more a reference than a text-book for readers who already have some knowl-edge of Bayesian modelling. The bibliographies after each chapter are useful for directing readers to other sources, but it would have been more helpful if they had been annotated.