Machine learning outcome regression improves doubly robust estimation of average causal effects.

Area of Science:

• Statistics
• Causal Inference
• Machine Learning

Background:

• Doubly robust (DR) estimation offers unbiased average treatment effect (ATE) estimation, provided propensity score (PS) and outcome models are correctly specified.
• However, DR estimators can be more biased than standard weighting estimators when both PS and outcome models are misspecified.

Purpose of the Study:

• To evaluate machine learning (ML) methods for estimating conditional potential outcome means.
• To enhance the robustness of DR estimators against model misspecification by reducing bias and standard error.
• To compare the performance of various ML techniques within the DR framework.

Main Methods:

• Assessed ML methods including least squares, tree-based models, generalized additive models, and shrinkage methods for outcome prediction.
• Employed the Super Learner (SL), an ensemble method combining multiple learners, for estimating conditional means.
• Conducted simulations across diverse scenarios varying in PS/outcome model complexity and treatment prevalence.

Main Results:

• Shrinkage methods demonstrated robust performance, yielding low bias and mean squared error in DR estimates, especially with rich models including 2-way covariate interactions.
• The Super Learner (SL) achieved performance comparable to the best individual method in each simulated scenario.
• Both shrinkage and SL methods effectively enhanced the accuracy of DR estimators.

Conclusions:

• Machine learning methods, specifically shrinkage methods with interaction models or the Super Learner, are recommended for improving the accuracy of doubly robust estimators.
• These advanced methods offer enhanced robustness against model misspecification in causal effect estimation.