Model Selection in Competing Risk Settings

• Activity Number: 630
• Type: Topic Contributed
• Date/Time: Thursday, August 7, 2014 : 10:30 AM to 12:20 PM
• Sponsor: Section on Statistics in Epidemiology
• Author(s): Ruth Maria Pfeiffer*+ and Stephanie Kovalchik
• Companies: National Cancer Institute and RAND Corporation
• Keywords: absolute risk ; model building ; competing risk ; cause specific hazard ; influence function

Abstract

Absolute risk (``crude risk'' or ``cumulative incidence''), the probability of experiencing an event of type m in the presence of M-1 competing causes, is a useful quantity for making clinical decisions and for establishing policies for disease prevention. Absolute risk can be estimated by properly combining estimates of cause-specific hazard functions for each cause in an integral expression. However, standard methods for model diagnosis and covariate selection of the cause-specific hazard model have limited applicability in the absolute risk setting because they do not consider the other components of the model. We thus introduce novel measures based on the absolute risk influence components to assist with the modeling of the competing events components, which is challenging, owing to their second-order effects on the absolute risk estimate. When there are multiple competing risks, there may be uncertainty about how they should be classified, that is, how narrowly events should be defined. The first measure, based on relative information a competing cause contributes to the absolute risk model relative to all competing causes, helps decide whether a cause is sufficiently