Utilizing Machine Learning for Valid Causal Inference

• In Event: Observational Causal Inference

Abstract

Despite its ubiquity, the linear regression suffers from several well-known flaws. First, specification choices can affect inference. Second, the regression is a correlative tool, generally not returning an average causal effect. Third, the method assumes no interference among observations. We introduce a method that addresses these shortcomings. First, the method combines a machine learning approach to control for background covariates and learn patterns of interference with a regression to estimate the coefficient on the treatment variable of theoretical interest. Second, the method models the treatment variable as well as the outcome, allowing recovery of a causal effect. The method applies whether the treatment variable is a binary, continuous, or a count variable. We give assumptions under which the method's estimate is consistent for the average causal effect, and we provide diagnostics to assess these conditions. We show that its standard errors are asymptotically valid and semiparametrically efficient. The incorporation of random effects allows for a much greater use than existing methods. A simulation study shows that the proposed method performs favorably relative to several recent machine learning methods, and application to real-world datasets illustrates the method's utility.

Author

• Marc Thomas Ratkovic, Princeton University