Jared Huling, of the Department of Statistics at The Ohio State University, will present:
“Energy Balancing of Covariate Distributions for Estimation of Causal Effects”
Abstract: Bias in the estimation of causal effects is a function of distributional imbalance of covariates between treatment groups. Weighting strategies such as inverse propensity score methods attempt to mitigate bias by either modeling the treatment assignment mechanism or balancing specified covariate moments. In practice, these approaches can be quite sensitive to modeling decisions. This talk instead introduces a novel weighting method based on energy distance and is explicitly designed to balance weighted covariate distributions, thus targeting the source of bias. The method has several advantages compared with existing weighting techniques. First, our weighting strategy provides a model free approach for causal comparisons and can be flexibly utilized in a wide variety of downstream causal analyses, such as the estimation of average treatment effects, individualized treatment rules, and more. Second, our approach is based on a genuine measure of distributional balance, providing a means of precisely assessing the covariate balance induced by a given set of weights. Finally, the method is computationally feasible and provides strong theoretical guarantees under weak conditions. The effectiveness of our approach is demonstrated in the analysis of two real-world applications, the first a study in the safety of right heart catheterization and the second a study of the effectiveness of a transitional care intervention at a large Midwestern academic medical center.
A social tea will be held at 11:00 a.m. in A434 Mayo. All are Welcome.