MODEL MISSPECIFICATION IN PROPORTIONAL HAZARDS REGRESSION

The proportional hazards model is frequently used to evaluate the effect of treatment on failure time events in randomised clinical trials. Concomitant variables are usually available and may be considered for use in the primary analyses under the assumption that incorporating them may reduce bias or improve efficiency. In this paper we consider two approaches to including covariate information: regression modelling and stratification. We focus on the setting where covariate effects are nonproportional and we compare the bias, efficiency and coverage properties of these approaches. These results indicate that our intuition based on linear model analysis of covariance is misleading. Covariate adjustment in proportional hazards models has little effect on the variance but may significantly improve the accuracy of the treatment effect estimator.