Laura Liu (Pitt): Bayesian Causal Inference and Orthogonality in High Dimensions
Abstract: This talk is based on two papers, “Bayesian Double Machine Learning for Causal Inference” and “Bayesian Causal Inference and Orthogonality in High Dimensions,” where we develop Bayesian methods for causal inference with a low-dimensional target and high-dimensional nuisance. Standard priors on the nuisance can unintentionally imply dogmatic beliefs about the target and therefore induce bias. In the first paper, we propose Bayesian Double Machine Learning for partial linear models, which recovers the causal effect from the reduced-form covariance matrix. In the second paper, we consider a general framework for causal inference, which incorporates Neyman orthogonality as local neutrality of the nuisance prior and constructs orthogonal working likelihoods that keep the restrictions identifying the target while transforming the nuisance. Examples include average treatment effects, missing data, heterogeneous treatment effects, and many instruments.
Economics
Email