Issues in win ratio estimation in presence of missing and censored data

Music City Center 

Since its introduction in 2012 by Pocock et al., the win ratio (WR) has become increasingly popular and has been used to analyze the results of multiple clinical trials. This approach is attractive because it combines information from multiple endpoints and takes into account their relative importance. However, it has been shown that win ratio estimate can be biased when the data are censored or missing. Unfortunately, in practical applications, these issues are often ignored and more ties often result from subjects with censoring or missing data than those without. Although some approaches have been proposed to deal with censoring for time-to-event endpoints, their implementation in practice is lagging and this is still an active area of research. In this presentation, we demonstrate that when all endpoints are time-to-event, the win ratio parameter depends on the length of follow up. Furthermore, in the presence of censoring, traditional WR estimator as described by Pocock et al. estimates a new parameter - a weighted average of win ratios at multiple lengths of follow up. Recognizing that the hypothesis testing in practice is often based on this new parameter, we show how to evaluate the sample size requirements in a trial design where the traditional WR estimator is used for primary analysis, under the assumption of administrative censoring due to uniform enrollment.