Bayesian Statistical Modelling

Rating:
Summary: nice recent text on Bayesian methods with many applications
Review: Congdon presents a very nice and modern treatment of Bayesian methods and models emphasizing implementation using BUGS or WINBUGS. The book covers Bayesian models for regression including linear, log-linear, robust and nonparametric regression. Covers association and classification, mixture models, latent variables, problems of missing data, survival analysis, hierarchical models for pooling information, time series and other correlated data methods (e.g. spatial processes), multivariate analysis, growth curves and model assessment criteria.

The book is loaded with techniques and applications covering a wide variety of topics with reasonable depth.

It also has a very large bibliography with many very relevant and useful references. But there is also a negative side to the bibliography. It was not carefully proofread and there are some annoyances as you will see the same reference listed two, three or more times in the bibliography. Also for such a nice reference text it should have included an author index as well as an ordinary index.

Gibbs sampling is one of the primary estimation techniques in the book but the details are put off until section 10.1 where we get a nice introduction to Gibbs sampling and also the Metropolis algorithm with several excellent references.

This is a good book to start implementing Bayesian methods through the MCMC technique. It contains mostly medical applications which is a nice feature for biostatisticians.

Rating:
Summary: nice recent text on Bayesian methods with many applications
Review: Congdon presents a very nice and modern treatment of Bayesian methods and models emphasizing implementation using BUGS or WINBUGS. The book covers Bayesian models for regression including linear, log-linear, robust and nonparametric regression. Covers association and classification, mixture models, latent variables, problems of missing data, survival analysis, hierarchical models for pooling information, time series and other correlated data methods (e.g. spatial processes), multivariate analysis, growth curves and model assessment criteria.

The book is loaded with techniques and applications covering a wide variety of topics with reasonable depth.

It also has a very large bibliography with many very relevant and useful references. But there is also a negative side to the bibliography. It was not carefully proofread and there are some annoyances as you will see the same reference listed two, three or more times in the bibliography. Also for such a nice reference text it should have included an author index as well as an ordinary index.

Gibbs sampling is one of the primary estimation techniques in the book but the details are put off until section 10.1 where we get a nice introduction to Gibbs sampling and also the Metropolis algorithm with several excellent references.

This is a good book to start implementing Bayesian methods through the MCMC technique. It contains mostly medical applications which is a nice feature for biostatisticians.

Rating:
Summary: Use it against spam?
Review: One specific application of Bayesian approaches has recently become hot. It has been claimed by some that a way to attack spam in email is to use Bayesian filters. If that is your inclination, you may want to check out this book.

It has a solid, technical explanation of Bayesians and is replete with several examples. Some parts may be heavy going, depending on your mathematical background. But worth it eventually.

The problem is, if you do develop a Bayesian filter for spam and use it, you may find that there are fundamental limitations, due to broadening, and to spammers actively counterattacking Bayesians.

The examples in the book of applications all involve cases where the system being modelled does not change significantly; otherwise the Bayesian will have to be retrained on new data. But, and more importantly, the examples do not treat the case where the system can change in a way to deliberately defeat the Bayesian.