Generalized pairwise comparisons using pseudo-observations for time-to-event censored data in a randomized controlled trial setting

Generalized pairwise comparison (GPC) methods are extensions of the Mann-Whitney approach that allow comparisons of outcomes through prioritized ranking, and they have been widely applied in randomized controlled trials (RCTs). Importantly, GPC methods can be adapted to handle censored time-to-event data. GPC methods are based on assigning scores to pairs of subjects where all pairs of treatment and control subjects are evaluated: the outcome of each treatment subject is compared with each control subject. The GPC test statistic can be expressed as a treatment effect for the therapeutic intervention by measures such as the net benefit, win odds, win ratio (WR), or probability index. As the focus for this study, the WR has an alternative interpretation as the inverse of the hazard ratio under proportional hazards. However, its estimate could be biased in the presence of substantial censoring. Censoring increases the number of indeterminate treatment and control pairs, where the win or loss is undetermined due to the censored observation(s) and a definitive score cannot be assigned. We propose a novel method leveraging pseudo-observations to address the issue of uninformative pairs resulting from censoring for a time-to-event outcome. We compare the performance of our method with existing GPC methods in simulations under various censoring scenarios. For equal drop-out and administrative censoring, our method provides results that are comparable to existing GPC methods. However, for unequal drop-out, which is common in clinical trials, the performance of our approach relative to existing methods depends on the censoring proportion and distribution. The proposed approach reduced bias and root mean squared error relative to Gehan and Latta under several censoring conditions, but these improvements did not extend to gains in statistical power. Lastly, we illustrate this new GPC approach using two reconstructed RCT datasets.