On Double Robustness in Double Machine Learning

Double Machine Learning (DML) is widely used for causal estimation from observational data and is often assumed to be doubly robust. While this holds for the Z-estimator proposed by Chernozhukov et al., many practical implementations rely on the Robinson estimator, which crucially depends on correct treatment model specification. This misunderstanding has important implications, as many practitioners incorrectly assume robustness to misspecification. We provide analyses clarifying when double robustness holds for popular DML estimators. Based on these insights, we develop a maximum likelihood estimator that achieves double robustness, providing a likelihood-based alternative to the Z-estimator.