An Assessment of the Treatment Effect in Treatment of Physician Choice Trials

In treatment of physician choice (TPC) trials, patients randomized to the control arm are treated with one of several drugs which are predetermined prior to the initiation of the trial. Statistical analysis in TPC trials typically hinge on a comparison of the experimental arm to the control arm, with the data from the control arm pooled across treatments. The robustness of the treatment effect for time-to-event endpoints estimated using the Cox proportional hazard model depends on the proportionality assumption. This assumption is violated if the magnitudes of the effects of the control treatments relative to the experimental treatment vary. The resulting effect is an averaged effect and is sensitive to censoring. We express the resulting hazard ratio from the Cox model where data from the control arm are pooled, as a weighted average of the head-to-head effects comparing controls treatments to the experimental treatments. These weights are function of the physicians preferences of the control treatments. Through simulation studies, we propose a method of identifying trial scenarios where the heterogeneity of the control treatments might lead to misleading results. We also propose an alternative estimator for the averaged effect which is unaffected by censoring.