Description:

Title:  A Double Robust Estimator For Mann Whitney Wilcoxon Rank Sum Test When Applied for Causal Inference in Observational Studies

Speaker:  Ruohui (Matt) Chen, University of California, San Diego

Abstract:  The Mann-Whitney-Wilcoxon Rank Sum Test (MWWRST) is widely used to compare two treatment groups in randomized control trials when data distributions are highly skewed, especially in the presence of outliers. As the MWWRST generally yields invalid causal inference when applied to observational study data, Wu et al. (2014) introduced an approach to address confounding effects by incorporating the inverse probability weighting technique into this rank-based statistic. More importantly, their approach addressed limitations of an earlier attempt by Rosenbaum (2002) based on a randomization inference technique by positing a constant treatment effect between two potential outcomes across all subjects. This assumption not only completely ignores subject level differences, but also is unverifiable in real studies. Since any assumption on potential

outcomes cannot be empirically checked, methods require such assumptions have very limited applications.

In this paper, we address an important limitation in Wu et al. (2014) by extending their approach to a doubly robust setting to provide more robust causal inference by integrating functional response models with the inverse probability weighting and mean score imputation. Additionally, we define causal effects that are consistent with the rank-based MWWRST. We demonstrate performances of the approach through both simulated and real study data.