Resampling Plans for Frailty Models

Abstract

Obtaining the standard error of the estimated heterogeneity in shared frailty models is in general difficult. Klein and Moeschberger (1997) show that the use of the observed information matrix is often not feasible because of its high dimension. Therneau and Grambsch (2000) use a nonparametric bootstrap algorithm to obtain standard errors for the estimated parameters in a shared frailty model. For parametric shared frailty models we define two model-based resampling schemes and use them to obtain standard errors. Based on a simulation study, we show that model-based resampling compares favorable to nonparametric resampling and that for all resampling schemes robustness is an issue of concern.