Prior Choice in Model Selection Problems

Speaker: James Lawrence (University of Cambridge)

Abstract

“Bayesian Model Selection and Averaging are common tools to deal with model uncertainty, not just in the ecological community but in statistics at large. It is usual in such circumstances to choose an “uninformative” prior over the model space, but we demonstrate how it is important to exercise care in choosing the model parameter priors as well, so as not to bias the prior towards certain models. We present an example in modeling density dependence in the American Wigeon and show how applying some simple rules and a little thought can lead to improved posterior inference and noticeably better predictive power.