Conditional survival given covariates and marginal survival

Abstract

Assuming some regression model, it is common to study the conditional distribution of survival given covariates. Here, we consider the impact of further conditioning, specifically conditioning on a marginal survival function, known or estimated. We investigate to what purposes any such information can be used in a proportional or non-proportional hazards regression analysis of time on the covariates. It does not lead to any improvement in efficiency when the form of the assumed proportional hazards model is valid. However, when the proportional hazards model is not valid, the usual partial likelihood estimator is not consistent and depends heavily on the unknown censoring mechanism. In this case we show that the conditional estimate that we propose is consistent for a parameter that has a strong interpretation independent of censoring. Simulations and examples are provided.