Andrew Womack
 

Department of Mathematics, Washington University
 

Title: The Posterior Predictive Information Criterion
 

Abstract: Model comparison is a necessary part of any data analysis. There are three tools that have become standards for model comparison and averaging: Bayes Factors, Information Criteria, and Cross Validation. The first suffers from issues of indeterminacy whenever improper priors are used, asymmetric treatment of models in the intrinsic case, and subjectivity of proper priors. The second suffers from choice of model focus and necessity of asymptotics. The third suffers from issues of data-splitting and poor performance. I introduce a new information theoretic model selection criterion that is a correction to Aitkin's Posterior Bayes Factor to control for over-fitting. Motivations and properties of the method will be discussed. Consistency will be exhibited in some examples and comparisons to popular methods will be made.