Metalearners for estimating heterogeneous treatment effects using machine learning

Area of Science:

• Statistics
• Machine Learning
• Econometrics

Background:

• Estimating heterogeneous treatment effects is crucial for personalized interventions.
• Standard machine learning algorithms are not designed to directly estimate conditional average treatment effects (CATE).

Purpose of the Study:

• To introduce meta-algorithms, specifically the X-learner, for estimating CATE.
• To demonstrate the efficiency and applicability of the X-learner in various settings.

Main Methods:

• Developed meta-algorithms that leverage existing supervised learning methods (e.g., random forests, BART, neural networks).
• Introduced the X-learner, a meta-algorithm designed for efficiency, especially with imbalanced treatment groups.
• Proposed X-learner versions utilizing random forests and Bayesian additive regression trees as base learners.

Main Results:

• The X-learner shows provable efficiency, particularly with unequal treatment group sizes.
• Simulations indicate favorable performance of the X-learner compared to other meta-learners.
• The X-learner successfully targeted treatment regimes and elucidated mechanisms in political science field experiments.

Conclusions:

• The X-learner provides a flexible and effective framework for estimating heterogeneous treatment effects.
• This approach enhances the utility of machine learning in causal inference.
• The developed methods and software package facilitate practical application in research.