Let me point out immediately that this is not a book you should choose as your first text on Bayesian statistics or as a gentle introduction to Bayesian analysis. It is unfortunate that the author makes no reference to whom the book is aimed and the proper level of reader background in computing. My guess is the audience should be MS level statistics students with at least one intro course in Bayesian analysis plus at least one course in mathematical statistics.
The author is not shy about programming steps, and the annotations are usually straight forward. Software is recommended for implementing the Markov chain Monte Carlo technique (MCMC), which is often used for sampling from a probability distribution based on constructing a Markov chain that has the desired distribution as its equilibrium distribution. The baseline for these sampling schemes is the Metropolis-Hastings (M-H) algorithm. These are used through much of the text. The software programs suggested are: BUGS, Winbugs, R-INLA and JAGS, although SAS and R are mentioned as having a growing arsenal of Bayesian routines.
I present major chapter headings to give the reader a better flavor of the book’s scope:
- Bayesian methods and Bayesian estimation
- Hierarchical models for related units
- Models for time series
- Analysis of panel data
- Models for spatial outcomes and geographical association
- Latent variable and structural equation models
- Survival and event history models
The text is heavily mathematical throughout and, although much code is given, many more completely worked examples would be desired. The book’s great strength lies in two areas: the first is Peter Congdon’s generation of an excellent bibliography of the most modern techniques available, and the other is his (slightly) more straightforward explanations of the strengths and weaknesses of these techniques and suggestions for optimizing the results.
Particularly helpful are his chapter subheadings for model choice and model assessment. Here, the more inexperienced analyst will find clearer explanations of criteria for choosing models, limitations of existing models with acceptable substitutes, and comparing models by predictions versus actual data.
Another positive includes the book’s treatment of meta-analysis. There are many instances of incomplete or small data sets that, when analyzed separately, yield little or no information, but when tied together may be more illuminating. Unfortunately, this is rarely a case of just throwing them all together and analyzing them as a more or less homogeneous group. The author considers univariate and multivariate meta-analysis, as well as multilevel models, with a few examples from clinical studies and a large dose of mathematics to justify the choices.
In summary, this is not a text for the novice. However, for those math/statistics aficionados, there is much to be had. A nice guidebook to intermediate and advanced Bayesian models.
Availability
Applied Bayesian Modelling (2nd Ed.), Peter Congdon. Wiley, Somerset, NJ. 464 pages, July 2014, $90. ISBN: 978-1-119-95151-3
John Wass is a statistician based in Chicago, IL. He may be reached at editor@ScientificComputing.com.