The win ratio/Finkelstein-Schoenfeld approach has been gaining momentum as a promising alternative to the time-to-first event approach and other tests for randomized trials with composite time-to-event endpoints. We aimed to compare the statistical performance of a wide range of statistical tests to design trials with multiple time-to-event outcomes. Tests based on generalized pairwise comparison (GPC) or the restricted mean time in favor (RMT-IF) of treatment test leverage an ordering of components by severity. Conversely, the O'Brien and Wei-Lachin tests combine component-wise treatment effects across all components. Tests such as the log-rank and MaxCombo tests on the time-to-first event, along with a test based on the area under the cumulative event count curve, do not differentiate among components. We performed in-depth simulation studies for a composite endpoint with three components. We generated component-wise treatment effects of different magnitudes or directions under both proportional and non-proportional hazards settings. We then evaluated empirical power and size for each test. We defined and compared the percentages of information used by each test. Our findings indicate that no single test was uniformly most powerful. Across different simulation scenarios, GPC and RMT-IF tests showed lower power when treatment effect was greater on less important events or when there was a delayed effect on the most important event. When the overall treatment effect was beneficial, the RMT-IF and Wei-Lachin tests demonstrated relatively robust power, while the RMT-IF test showed a mildly elevated type I error in small samples. Both tests utilize all observed events and allow for component-wise decomposition of treatment effects. We illustrated all the tests on data from a cardiovascular trial.
Lin et al. (Fri,) studied this question.