PS36: Covariate Adjustment in Practice: Robust Implementation and Extension to Win Statistics

For designs and analyses of randomized clinical trials (RCTs), the importance of covariate adjustment has been increasingly recognized, driven by recent advancements such as the FDA's 2023 guidance document, EMA's support for Prognostic Covariate Adjustment (PROCOVA), and methods such as g-computation and digital twins. In principle, adjusting for baseline covariates, particularly those that are prognostic for the outcome, can enhance precision in treatment effect estimation. Therefore, model-assisted approaches (e.g., g-computation) are in general recommended to improve efficiency and robustness against model misspecifications. However, there is a lack of robust implementation of various covariate adjustment methods that is ready to be used for GxP purposes. The first presentation of this session will feature the RobinCar2 R package (ROBust estimation and INference for Covariate Adjustment in Randomized clinical trials), which has been developed under the ASA BIOP covariate adjustment scientific working group. Moreover, in clinical trials, evaluations often rely on multiple endpoints to comprehensively capture various aspects of treatment effects. In some therapeutic areas, designing a clinical trial with a feasible sample size and adequate statistical power using traditional clinical endpoints becomes challenging, particularly when event rates are low (e.g., mortality rates in diabetes studies). To address this, the win ratio, together with the win odds and net benefit, has been increasingly popular to handle multiple outcomes in a hierarchical order. Therefore, when lower-priority outcomes (e.g., progression in oncology) occur earlier, they do not "mask" more important outcomes (e.g., death). The win ratio and the win odds have been applied in clinical drug development, particularly in design and analysis of Phase III clinical trials, with the win ratio also supporting regulatory approval processes. However, the win statistics (win ratio, the win odds, and net benefit) are non-collapsible, which means that the marginal (overall) treatment effect differs from the conditional (subgroup-specific) treatment effect, even if the treatment effect is identical across subgroups. This non-collapsibility issue can lead to differences in interpretation and create challenges in aligning statistical analyses with regulatory requirements and clinical objectives. Due to the nature of win statistics as nonparametric measures of treatment effect, the second presentation of this session will introduce a novel nonparametric analytical solution for win statistics.