A General Framework for Treatment Effect Estimation in Semi-Supervised and High Dimensional Settings

  • 2022-01-03 04:12:44
  • Abhishek Chakrabortty, Guorong Dai, Eric Tchetgen Tchetgen
  • 22

Abstract

In this article, we aim to provide a general and complete understanding ofsemi-supervised (SS) causal inference for treatment effects. Specifically, weconsider two such estimands: (a) the average treatment effect and (b) thequantile treatment effect, as prototype cases, in an SS setting, characterizedby two available data sets: (i) a labeled data set of size $n$, providingobservations for a response and a set of high dimensional covariates, as wellas a binary treatment indicator; and (ii) an unlabeled data set of size $N$,much larger than $n$, but without the response observed. Using these two datasets, we develop a family of SS estimators which are ensured to be: (1) morerobust and (2) more efficient than their supervised counterparts based on thelabeled data set only. Beyond the 'standard' double robustness results (interms of consistency) that can be achieved by supervised methods as well, wefurther establish root-n consistency and asymptotic normality of our SSestimators whenever the propensity score in the model is correctly specified,without requiring specific forms of the nuisance functions involved. Such animprovement of robustness arises from the use of the massive unlabeled data, soit is generally not attainable in a purely supervised setting. In addition, ourestimators are shown to be semi-parametrically efficient as long as all thenuisance functions are correctly specified. Moreover, as an illustration of thenuisance estimators, we consider inverse-probability-weighting type kernelsmoothing estimators involving unknown covariate transformation mechanisms, andestablish in high dimensional scenarios novel results on their uniformconvergence rates, which should be of independent interest. Numerical resultson both simulated and real data validate the advantage of our methods overtheir supervised counterparts with respect to both robustness and efficiency.

 

Quick Read (beta)

loading the full paper ...