A0590
Title: Causal Inference with Missing Covariates: A Unified Bayesian Perspective and Improved Methods
Authors: Arman Oganisian - Brown University (United States) [presenting]
Abstract: Comparative effectiveness studies often leverage large observational databases such as electronic health records or insurance claims and use causal inference methods to adjust for confounding of the treatment effect by a set of pre-treatment covariates. A key challenge, however, is dealing with covariate missingness. Many ad-hoc approaches are encountered in the wild when estimating causal effects in the presence of missing covariates. Some include missingness indicator columns as adjustment covariates in either a propensity score or outcome model; others are two-step approaches that first impute covariates, then apply the usual causal methods (standardization, weighting, etc.). Others combine imputations and indicators. Many papers have been written comparing statistical performance of such methods but provide little fundamental and causal insight. Here, we offer a unified Bayesian perspective. The Bayesian paradigm is unifying because uncertainty about all unknown quantities - including parameters, missing covariate values, and missing counterfactuals - is encoded via a single joint distribution. Taking this approach, we show how different classes of methods 1) imply distinct assumptions about the joint treatment-missingness mechanism, 2) can be recovered under distinct factorizations of the joint distribution, and 3) pose distinct modeling challenges. We also propose some Bayesian solutions to these challenges.
