A0157
Title: Near-Bayesian methods
Authors: Steven MacEachern - The Ohio State University (United States) [presenting]
Abstract: Bayesian methods have proven to be extremely successful when the class of models contains the mechanism that generates the data. Their performance suffers when the data-generating mechanism lies outside the support of the prior distribution; that is when the model is misspecified. Several methods have been proposed to handle model misspecification from a Bayesian perspective. These include methods that are formally Bayesian, those that are arguably Bayesian, and those that have a Bayesian motivation but depart from Bayes. This suggests a need for mechanisms to assess how close a method is to a formal Bayesian method. Several approaches are surveyed, briefly highlighting their strengths and weaknesses and contrasting their performance.
