|
Abstract:
|
Bayesian model selection has a number of desirable properties, including consistency, generality of application, automatic Occam's razor effect, and appropriate handling of multiplicities. Unlike inference under a single model, in the presence of model uncertainty, the choice of priors has a profound effect. However, in these situations, subjective elicitation is usually unfeasible. In this roundtable, we can debate 'automatic' or objective choice of 'good' priors for model selection (or model averaging). Questions to be considered could include (but are not limited to) the following: Which 'default' or 'automatic' priors should not be used? Which properties should an objective model selection procedure have? How do we numerically compute Bayes factors? When is BIC a sensible approximation? Are there generalizations for increasing parameter dimensionality?
|
