Using Hybrid MCMC for Logistic Regression Model Selection

• Activity Number: 353
• Type: Contributed
• Date/Time: Tuesday, August 3, 2010 : 10:30 AM to 12:20 PM
• Sponsor: Section on Statistical Computing
• Author(s): Robert Feyerharm*+
• Companies: Independent Consultant
• Keywords: Markov chain Monte Carlo ; logistic regression ; Akaike Information Criterion

Abstract

Logistic regression models are used in many fields to study binary or proportional outcome variables. Selecting the optimal logistic regression model from among a large number of candidate variables poses a challenging computational problem. In this paper I propose using the hybrid MCMC method to search a finite parameter space of candidate explanatory variables with the goal of maximizing a target density constructed from a normalized AIC function. Variables are eliminated from the final model where the MCMC procedure converges to a ßk=0 solution. To ensure differentiability and proper convergence properties, the AIC function is smoothed in the vicinity of degenerate points (ßk=0). Simulation data is generated to assess the convergence and efficiency of the MCMC procedure.