Doubly robust win statistics for hierarchical survival endpoints in individual- and cluster-randomized trials
Abstract
Prioritized composite endpoints compare trial participants according to a prespecified clinical hierarchy, and the win estimands are increasingly used to summarize treatment effects. When the component event times are right-censored, a pairwise comparison may remain unresolved, and treating unresolved pairs as ties yields a win statistic whose value depends on the censoring distribution. We define win, loss, and tie probabilities at a fixed restriction time for prioritized time-to-event outcomes in individual-randomized trials and extend them to cluster-randomized trials, where cluster-average and individual-average estimands assign equal weight to clusters and to participants, respectively. We propose doubly robust estimators that combine inverse probability of censoring weighting with joint outcome models to incorporate information recorded before censoring. Under covariate-dependent censoring, the resulting U-statistics are consistent when either the censoring models or joint outcome models are correctly specified for both treatment arms, but not necessarily both. We establish asymptotic normality and derive analytic variance estimators for computationally efficient inference. Simulation studies under both designs evaluate bias, precision, and confidence interval coverage under working model misspecification. Applications to two real-world studies demonstrate the use of the proposed methods under individual- and cluster-randomized trials.
Related
- Interpreting Hierarchical Composite Endpoints with Survival and Longitudinal Outcomes: Application to Amyotrophic Lateral Sclerosis Trials
- RECaST-Surv: A Calibrated Borrowing Method for Survival Endpoints in Unequal Randomized Trials
- Evaluating prognostic score adjustment and information borrowing in randomized trials with binary endpoints
- Detecting adverse high-order drug combinations from individual case safety reports using computational statistics on disproportionality measures
- Modeling Transition Dynamics and Network Structure in Cross-National Process Data: A Hierarchical Multi-State Survival Framework
BibTeX
@misc{fang2026doubly,
title = {{Doubly robust win statistics for hierarchical survival endpoints in individual- and cluster-randomized trials}},
author = {Xi Fang and Fan Li},
howpublished = {arXiv},
year = {2026},
month = nov,
doi = {10.48550/arxiv.2611.05853},
eprint = {2611.05853},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2611.05853},
}
RIS
TY - UNPB
TI - Doubly robust win statistics for hierarchical survival endpoints in individual- and cluster-randomized trials
AU - Xi Fang
AU - Fan Li
PY - 2026
DA - 2026/11/04
PB - arXiv
DO - 10.48550/arxiv.2611.05853
UR - https://arxiv.org/abs/2611.05853
AB - Prioritized composite endpoints compare trial participants according to a prespecified clinical hierarchy, and the win estimands are increasingly used to summarize treatment effects. When the component event times are right-censored, a pairwise comparison may remain unresolved, and treating unresolved pairs as ties yields a win statistic whose value depends on the censoring distribution. We define win, loss, and tie probabilities at a fixed restriction time for prioritized time-to-event outcomes in individual-randomized trials and extend them to cluster-randomized trials, where cluster-average and individual-average estimands assign equal weight to clusters and to participants, respectively. We propose doubly robust estimators that combine inverse probability of censoring weighting with joint outcome models to incorporate information recorded before censoring. Under covariate-dependent censoring, the resulting U-statistics are consistent when either the censoring models or joint outcome models are correctly specified for both treatment arms, but not necessarily both. We establish asymptotic normality and derive analytic variance estimators for computationally efficient inference. Simulation studies under both designs evaluate bias, precision, and confidence interval coverage under working model misspecification. Applications to two real-world studies demonstrate the use of the proposed methods under individual- and cluster-randomized trials.
ER -