Matching-Based Nonparametric Estimation of Group Average Treatment Effects.

Area of Science:

• Causal inference
• Statistical modeling
• Personalized medicine

Background:

• Heterogeneous treatment effects are vital for tailored strategies.
• Existing methods for group average treatment effects (GATEs) face instability and bias issues, particularly with extreme propensity scores.
• Focusing on key covariates for treatment decisions is common practice.

Purpose of the Study:

• To propose novel nonparametric methods for estimating group average treatment effects (GATEs).
• To address limitations of existing weighting-based and regression-based GATE estimation methods.
• To enhance the robustness and reduce bias in estimating heterogeneous treatment effects.

Main Methods:

• Developed a matching-based method for imputing potential outcomes, followed by nonparametric regression.
• Introduced a bias-corrected matching estimator incorporating outcome regression (OR) models to mitigate bias.
• Demonstrated consistency, double robustness, and asymptotic normality of the bias-corrected estimator.

Main Results:

• The proposed matching-based method avoids instability from extreme propensity scores.
• The bias-corrected matching estimator shows enhanced robustness and reduced bias, especially in high-dimensional covariate settings.
• Extensive simulations and a real-world application confirm the advantages of the novel methods.

Conclusions:

• The novel matching-based and bias-corrected matching methods provide more stable and accurate estimation of GATEs.
• The bias-corrected matching estimator offers desirable statistical properties, including consistency and double robustness.
• The open-source R package MatchGATE facilitates the application of these advanced causal inference techniques.