Stacked Survival Models for Weighted Methods with Censoring

• Activity Number: 604
• Type: Contributed
• Date/Time: Thursday, August 7, 2014 : 8:30 AM to 10:20 AM
• Sponsor: Biometrics Section
• Author(s): Kyle D. Rudser*+ and Andrew Wey and John Connett
• Companies: University of Minnesota and University of Minnesota and University of Minnesota
• Keywords: Survival prediction ; Ensemble ; Conditional distribution ; Brier score

Abstract

A variety of methods rely on estimation of nuisance quantities that one would like to allow to depend on covariates for additional flexibility and robustness. Estimation of these quantities, though not of direct scientific interest, can have a substantial impact on the performance of estimating parameters that are of interest. Nonparametric estimators can be preferred to parametric and semi-parametric estimators due to relaxed assumptions that enable robust estimation. However, even when misspecified, parametric and semi-parametric estimators can possess better operating characteristics in small samples due to smaller variance. Estimating conditional distribution functions through stacked survival models is explored for the censored data setting. By minimizing prediction error, stacking estimates optimally weighted combinations of survival models that span parametric, semi-parametric, and non-parametric models. As such, the low variance of approximately correct parametric models can be exploited while maintaining the robustness of nonparametric models. A simulation study demonstrates stacked survival models consistently outperform alternatives across a wide variety of situations.