Treatment Effect Performance of the X-Learner in the Presence of Confounding and Non-Linearity

This study critically evaluates a recent machine learning method called the X-Learner, that aims to estimate treatment effects by predicting counterfactual quantities. It uses information from the treated group to predict counterfactuals for the control group and vice versa. The problem is that studies have either only been applied to real world data without knowing the ground truth treatment effects, or have not been compared with the traditional regression methods for estimating treatment effects. This study therefore critically evaluates this method by simulating various scenarios that include observed confounding and non-linearity in the data. Although the regression X-Learner performs just as well as the traditional regression model, the other base learners performed worse. Additionally, when non-linearity was introduced into the data, the results of the X-Learner became inaccurate.