A sensitivity analysis for causal parameters in structural proportional hazards models

Abstract

Deviations from assigned treatment occur often in clinical trials. In such a setting, the traditional intent-to-treat analysis does not measure biological efficacy but rather programmatic effectiveness. For all-or-nothing compliance situation, Loeys and Goetghebeur (2003) recently proposed a Structural Proportional Hazards method. It allows for causal estimation in the complier subpopulation provided the exclusion restriction holds: randomization per se has no effect unless exposure has changed. This assumption is typically made with structural models for noncompliance but questioned when the trial is not blinded. In this paper we extend the structural PH model to allow for an effect of randomization per se. This enables analyzing sensitivity of conclusions to deviations from the exclusion estriction. In a colo-rectal cancer trial we find the causal estimator of the effect of an arterial device implantation to be remarkably insensitive to such deviations.