Bias in the IPCW estimator for censored pairwise comparisons: the importance of explicit estimands

Oregon Convention Center 

Net benefit and win ratio are gaining interest in oncology and cardiovascular research for assessing multifaceted clinical outcomes. They compare multiple outcomes hierarchically by using all possible pairs from treatment and control groups. For time-to-event outcomes suffering from censoring, the inverse probability of censoring weighting (IPCW) is available. However, Dong's (2021) original IPCW formulation does not distinguish (a) "ties" in uncensored prioritized outcomes and (b) "uninformative" comparison due to censoring when considering lower-priority outcomes. The resulting censoring-dependent bias has been overlooked because of the lack of clear estimands in pairwise comparisons. In this talk, we introduce explicitly defined estimands for net benefit/win ratio with censored outcomes in terms of the "probabilities that would be observed if we had removed censoring". We show the bias by the series of simulations by varying dependencies in censoring, outcome correlations, and treatment effects for separate outcomes. We propose the modified IPCW estimator that reduces the bias but sacrifices efficiency by excluding uninformative censored pairs.

generalized pairwise comparisons

inverse probability of censoring weighting

net benefit

prioritized outcomes 

Biopharmaceutical Section