Machine Learning for Causal Inference

Machine learning (ML) methods are increasingly used for estimating causal effects, particularly in high-dimensional settings with large-scale data. This article reviews key ideas underlying the use of ML for causal inference, and the fundamental challenges posed by confounding and missing counterfactual outcomes. The article introduces common ML methods for causal inference, including ML-based estimators for average treatment effects (e.g., regression adjustment, inverse-propensity of treatment weighting (IPTW), the Robinson estimator, augmented inverse probability of treatment weighting estimator (AIPTW), and targeted maximum-likelihood estimator (TMLE)) as well as meta-learners for heterogeneous treatment effects (e.g., S-, T-, IPTW-, DR-, and R-learner). The article discusses these concepts within the broader frameworks of semiparametric efficiency and orthogonal learning theory, and discusses intricate connections and various insights for constructing and analyzing modern ML methods for causal inference. Various practical recommendations are provided to improve both the understanding and the reliable use of ML for causal inference.

Recommended citation:

@incollection{frauen2026machine,
  title={Machine Learning for Causal Inference},
  author={Frauen, Dennis and Melnychuk, Valentyn and van der Laan, Lars and Feuerriegel, Stefan},
  booktitle={Wiley StatsRef: Statistics Reference Online},
  pages={1--17},
  year={2026},
  publisher={Wiley},
  doi={10.1002/9781118445112.stat08670}
}