Modeling pseudo-observations with covariate dependent censoring: robustness of the method against misspecified censoring models

Abstract

The so called pseudo-observations in survival analysis were introduced by recent studies that reviewed this method when estimating different parameters using regressions models (Andersen and Perme, Stat. Meth. Med. Res., 2010) with the condition that the censoring distribution is independent from covariates. If censoring depends on covariates, the method based on pseudo-observations requires modeling of the censoring distribution, which leads to the construction of alternative estimators based on censoring probability weighting. This master thesis will present the proposal of Andersen and Perme and -- by means of Monte Carlo simulation -- will also study its robustness if the model for the censoring distribution is misspecified. Two alternative estimators will be explained and used for the study of robustness of the method: the Cumulative Incidence Function and the Restricted Mean Lifetime.