A general Bayesian procedure for model comparisons with non-informative priors - a different resolution of Lindley's paradox Seminar

This talk gives a brief overview of the Bayesian model comparison approach using the posterior distribution of the likelihoods for each model for non-nested models, and for their likelihood ratio for nested models. This approach requires only the same diffuse or non-informative priors used for posterior parameter inference, and avoids completely the Lindley paradox problem afflicting Bayes factors.

The approach is illustrated with some simple examples, and a complex example of the number of components in a finite mixture of normals. A book-length treatment is in my 2010 book Statistical Inference: an Integrated Bayesian/Likelihood Approach (Chapman and Hall/CRC Press).