Bayesian Nonparametric Estimation of Regression Models in Event History Analysis

• Activity Number: 553
• Type: Invited
• Date/Time: Wednesday, August 4, 2010 : 2:00 PM to 3:50 PM
• Sponsor: Section on Nonparametric Statistics
• Author(s): Pierpaolo De Blasi*+
• Companies: University of Turin/Collegio Carlo Alberto
• Keywords: Bayesian nonparametrics ; beta process ; competing risks ; hazard regression ; random measures

Abstract

A new semiparametric specification of the competing risks model is studied. Cause-specific hazard (CSH) is modelled as the product of the conditional probability of a failure type and the overall hazard rate. Regression analysis on CSHs is introduced via a logistic relative risk factor in the specification of the conditional probability, while the overall hazard is left unspecified. The model can be adapted to describe the case of independence between failure times and failure types. Frequentist estimators are based on partial likelihood methods and are shown to possess good large sample properties. Bayesian inference involve a beta process prior for the overall cumulative hazard and a Jeffreys-type prior for the finite-dimensional parameters. Data on prognostic factor on cancer is considered for illustration.