Comparison of methods to address confounding in estimation of the win ratio in an observational setting

Music City Center 

The win ratio (WR) is a method for comparing composite endpoints. Originally used in randomized clinical trials, there is interest in applying this approach to observational studies with multiple clinically relevant endpoints. The unmatched WR approach compares all participants in one treatment arm to all participants in another arm and declares a 'winner' according to a hierarchy of potential outcomes for each pairwise comparison. The WR is then estimated as the ratio of 'winners' in the treatment arm to 'winners' in the control arm, excluding any hierarchical 'ties'. While this provides unbiased estimates in settings where there is balance in groups on measured and unmeasured confounders, in observational settings, confounding related to non-random assignment of treatment must be considered. Methods to address potential covariate imbalance include matching, stratification, and inverse probably of treatment weighting, however, each has potential limitations such as limiting sample size and representativeness. We explore use of the WR in an observational setting using a retrospective study comparing composite endpoints in Hospital-acquired/Ventilator-Associated Pneumonia by prescribed antibiotic treatment. We examine imbalance in measured confounding; characterize how each method affects the WR estimate; and discuss implications for their use in observational studies.