B1377
Title: Assessing covariate balance in matched observational studies with high-dimensional categorical variables
Authors: Massimiliano Russo - Harvard Medical School (United States) [presenting]
Abstract: Inferring causal effects in matched non-randomized studies requires that the exposure groups are approximately balanced. This implies that the two groups share a similar joint distribution of many confounding factors. There is vast literature on how to achieve such balance, but less attention has been devoted to assessing if the balance has been reached at different scales, including high-order interactions of the confounding factors. Focusing on the common case of multivariate categorical data, we describe a global test for balance that assesses if the joint probability mass function differs between the exposure groups, estimating a Bayes Factor via either an efficient Gibbs sampler or a variational approximation. We discuss explicit control of type-I error via a permutation scheme. If the imbalance is detected, local tests with explicit false discovery rate control are performed to detect which terms or interaction terms are responsible for the imbalance. We compare our methods with popular competitors, and discuss improved performance in simulations and real cohort studies.
