B1917
Title: The weighting representation of Bayesian causal effect estimators
Authors: Jared Murray - University of Texas at Austin (United States) [presenting]
Avi Feller - University of California at Berkeley (United States)
Abstract: Bayesian nonparametric (BNP) models are powerful tools for many causal inference tasks but are often opaque and difficult to assess in practice. A novel approach is presented to understanding BNP models by showing that the treatment effect estimates from popular methods (such as Bayesian additive regression trees (BART), Bayesian causal forests (BCF), and more general Gaussian processes models) have a representation as weighting estimators. This representation is used to introduce a range of new model checks and diagnostics tailored to causal inference, to facilitate comparisons of competing models and methods, to help guide model and prior specification, and to shed new light on the unreasonable effectiveness of some Bayes estimators even under significant model misspecification.
