Understanding Inference from Stochastic Search Variable Selection with Different Priors: How Much Evidence Is There?

• Activity Number: 225
• Type: Contributed
• Date/Time: Monday, August 3, 2009 : 2:00 PM to 3:50 PM
• Sponsor: Section on Bayesian Statistical Science
• Author(s): Michael D. Swartz*+
• Companies: The University of Texas M.D. Anderson Cancer Center
• Keywords: Stochastic Search ; Model Selection ; Bayesian Model Averaging ; Bayes Factor ; Epidemiologic applications

Abstract

Stochastic Search Variable Selection (SSVS) has been shown to be a useful technique for case-control designs, using both probit and logistic regression likelihoods. But unlike Bayes Factors, SSVS does not yet have a rule of thumb or decision rule for selecting variables. Although Barbieri and Berger's (2004) median model provides guidance for selecting a model with optimal predictive properties, it can be sensitive to the prior. Using simulated and real data, we investigate the influence of different priors on the median model decision rule. We fit logistic regression using different values for the prior probability of inclusion, and also a Beta distributed hyper prior, and examine the selected models using both Bayes Factors and the median model. We compare the fit of the selected models using the Area under the curve (AUC) of Receiver Operator Characteristic (ROC) functions.