A Sensitivity Analysis for Clinical Trials with Informatively Censored Survival Endpoints

Abstract

Analyses of clinical trials with time-to-event endpoints typically employ the assumption of non-informative censoring. While this assumption is usually appropriate for end-of-study (EOS) censoring, its applicability to lost-to-follow-up (LTFU) censoring is often suspect and may result in biased estimates of the treatment effect. To assess the robustness of estimates to departures from non-informative censoring, authors have proposed sensitivity analyses that assume a semiparametric model for the censoring mechanism, with the parameters representing associations between censoring and increased or decreased rates of survival. The parameters are varied over a plausible range resulting in a corresponding range of estimates for the treatment effect. We consider such an approach for two-arm trials in which the sensitivity parameters represent hazard ratios within a proportional hazards model with a time varying covariate comparing subjects who have been lost to follow-up to all other subjects. Using hypothesized hazard ratios for each arm separately, we multiply impute the unobserved data as it might have been observed in the absence of informative censoring. The treatment effect estimates computed using the imputed data are then summarized in a graphical display. Of particular interest in this research is the robustness of our approach to violations of the proportional hazards assumptions used when imputing the missing data. On the basis of extensive simulation studies, we find that the accuracy of the sensitivity analyses are relatively unaffected by departures from the semiparametric assumptions.