Analyses of Dependent Right-Censored Data.

Abstract

Right-censored data arise when the time until the occurrence of a terminal event is not observed for all members of a study population. A class of models for the analysis of right-censored data with dependent censoring has been proposed (Lagakos, 1978). Simulations based on this cone-class of models are used to examine the robustness of the maximum likelihood estimator for an exponential distribution of termination times to departures from the assumption of independent censoring. Other estimators for an exponential distribution of termination times derived from this cone-class of models are presented. The robustness of these estimators to misspecifications of the model within the cone-class is examined. In addition, these estimators for an exponential distribution of termination times are compared under conditions of independent censoring. A methodology for generating dependent right-censored data which does not depend on the assumption of a cone-class model is developed. This procedure is flexible in the types of dependent censoring which may be modelled. This methodology for generating simulated right-censored data provides investigators with a powerful tool for studying the robustness of estimators for the distribution of termination times to departures from assumptions concerning the nature of the censoring dependence. Simulations are performed using this method to generate data in which estimators for the termination time distribution are computed under conditions of positive dependent, negative dependent, and independent censoring. Results of these studies indicate that the cone-class models provide investigators with a useful procedure for estimating an exponential distribution of termination times when there is positive dependent censoring. Cone-class models more general than those in the statistical literature are described and results from simulations reported.