Impact of informative censoring on estimation and testing in randomized trials with delayed treatment effects.

Time-to-event endpoints like progression-free survival in oncology randomized trials sometimes demonstrate differential censoring patterns between study arms which can be indicative of informative censoring, depending on censoring reasons. Informative censoring can bias treatment effect estimates but few simulation studies characterized the magnitude of its impact, particularly in the context of therapies with delayed treatment effects. We used copula methods to model dependent censoring data and assessed the impact of informative censoring. To improve the understanding of copula models in this context, we proposed a new measure of the strength of informative censoring, the probability of events being informatively censored. We further proposed a visual tool for examining the underlying correlation pattern between censoring and event time. We conducted simulation studies to assess the impact of informative censoring on estimation bias for hazard ratios, as well as on empirical power of unweighted, weighted log-rank tests, and the MaxCombo test. We implemented data generation algorithm for copula survival models with piece-wise exponential marginals to introduce various censoring patterns under scenarios with delayed treatment effect. We found large overestimation of hazard ratio of the experimental arm versus the control arm and loss in power when there was a positive correlation between event time and censoring time in the control arm and a negative correlation in the experimental arm. When correlations in both arms were of the same direction and degree, we observed minimal impact on hazard ratio estimate and statistical powers.