Estimating the effect of treatment in a proportional hazards model in the presence of non-compliance and contamination

Methods for adjusting for non-compliance and contamination, which respect the randomization, are extended from binary outcomes to time-to-event analyses by using a proportional hazards model. A simple non-iterative method is developed when there are no covariates, which is a generalization of the Mantel-Haenszel estimator. More generally, a 'partial likelihood' is developed which accommodates covariates under the assumption that they are independent of compliance. A key feature is that the proportion of contaminators and non-compliers in the risk set is updated at each failure time. When covariates are not independent of compliance, a full likelihood is developed and explored, but this leads to a complex estimator. Estimating equations and information matrices are derived for these estimators and they are evaluated by simulation studies.