B0716
Title: A general framework for treatment effect estimation in semi-supervised and high dimensional settings
Authors: Abhishek Chakrabortty - Texas A\&M University (United States) [presenting]
Abstract: Semi-supervised (SS) settings are of growing relevance in modern studies. However, their full scope and benefits for causal inference problems are not yet well explored. Using the average treatment effect (ATE) as a prototype case, a general understanding of causal inference is provided in SS settings, where one has labelled (or supervised) data on a treatment, a response, and a set of (possibly high dimensional) covariates, and a much larger unlabeled (or unsupervised) data without the response. It is generally of interest to investigate how the additional unlabeled data available in the SS setting can be exploited to improve (efficiency and/or robustness) upon a fully supervised approach. A family of SS ATE estimators are developed with a flexible construction and gives a full characterization of their properties, revealing several key benefits of SS settings. In particular, they are ensured to be (1) more robust and (2) more efficient (and optimal, too, in some cases) than their supervised counterparts. Moreover, beyond the standard double robustness that can be achieved by supervised methods, root-n consistency and asymptotic normality of the SS estimators are also established whenever the propensity score model is correctly specified, without requiring any specific forms for both the nuisance models. Such an improvement in robustness arises from the use of the massive unlabeled data and thus is generally unachievable in a purely supervised setting.
