B1819
Title: Approximate balancing weighting for treatment effects: Justifications, choices, and fundamental limitations
Authors: Chad Hazlett - UCLA (United States) [presenting]
Abstract: Weighting approaches used in applied causal inference settings may rest on the estimation of the propensity score, on directly seeking to balance covariates, or on a combination of these goals. Accordingly, these approaches are provably unbiased under claims regarding the specification of the propensity score, the specification of the (conditional expectation of) outcome, or under a weaker combined assumption. We review these varying motivations for weighting approaches, emphasizing the (less widely-recognized) justification that requires only assumptions on the outcome model. An analysis of the bias in the estimated treatment effect illustrates the specification dependencies inherent in weighting approaches, aiding investigators in answering the practical question of what functions of the covariates should be balanced. This offers one motivation for a family of kernel-based weighting estimators proposed by authors on this panel. Finally, we turn to a central tradeoff at the heart of these methods: weighting approaches allow us to make weak specification assumptions, but attempt to obtain balance on features that may be irrelevant. Indeed, weights that produce worse balance, or even offsetting imbalances, can have better finite-sample performance. We discuss how this limits the practical performance of weighting-only approaches, and alternative or hybrid options.
