Posterior Predictive Model Choice for Survival Data

• Activity Number: 252
• Type: Contributed
• Date/Time: Tuesday, August 13, 2002 : 2:00 PM to 3:50 PM
• Sponsor: Section on Bayesian Stat. Sciences*
• Author(s): Fei Wang*+ and Alan Gelfand+
• Affiliation(s): Boston University and University of Connecticut
• Keywords: proportional hazard models ; censoring ; deviance loss ; Monte Carlo integration; ; accelerated life models ; grouped data models

Abstract

Currently, increasingly complex models for survival data are being employed, typically including random effects and nonparametric aspects. With such high, possibly infinite, dimensional models, how does one avoid overfitting? How can one encourage parsimony? Customary penalized likelihood approaches are not applicable. We offer a predictive utility-based approach, which extend the work of Gelfand and Ghosh (1998). It carries the same benefits as the criterion in rewards fidelity to the observed data, as well as precise prediction, without requiring specification of model dimension. The utility (or loss) is motivated by familiar deviance notions. The resulting criterion emerges as a sum of two terms, one measuring goodness-of-fit, the other penalizing for overfitting. The performance of the criterion is examined for a bone marrow transplant dataset.