Some Recent Developments in Bayesian Model Selection
James O. Berger
Duke University and SAMSI
Abstract
We
review two fairly recent developments in Bayesian model
selection
- A generalization of BIC has recently been developed,
for contexts common in the social sciences, that appropriately assesses
the dimension of a model and the effective sample size for each
parameter in a model. The generalization allows for the model dimension
to grow with the sample size.
- When the space of models is large, Search Strategies
need to be carefully developed for exploration of the space. One
successful strategy in variable selection is to perform a stochastic
search that (roughly) adds or removes variables based on their current
estimated posterior inclusion probabilities. This approach, and related
diagnostics, will be illustrated. An interesting phenomenon that seems
to be frequently encountered is that no model receives significant
posterior probability, so that the meaning of model selection and the
questions we pose concerning models may need to be reconsidered.
|