Double robust methods for flexible covariate adjustment in causal inference have proliferated in recent years. Despite their apparent advantages in overcoming model misspecification, these methods are rarely used by social scientists. It is also unclear whether these methods actually outperform more traditional methods in finite samples. This paper has two aims: It is a guide to some of the latest methods in double robust, flexible covariate adjustment using machine learning, and it tests them against more traditional statistical methods and flexible "single robust" methods using simulated, cross-sectional data where the treatment effect is known. Double robust methods covered include Augmented Inverse Probability Weighting (AIPW), Targeted Maximum Likelihood Estimation (TMLE), and Double/Debiased Machine Learning (DML). Results suggest that these methods do outperform traditional methods in a wide range of simulations, but only when paired with flexible machine learning estimators. Notably, G-computation with the same flexible estimators obtains almost identical results, and standard regression methods have only slightly higher bias. In the scenarios considered in this paper, flexibility and robustness to heterogeneous treatment effects appear to be more important properties than double robustness.
Nathan Isaac Hoffmann (Tue,) studied this question.