Improving propensity score weighting using machine learning.

Area of Science:

• Statistics
• Machine Learning
• Epidemiology

Background:

• Propensity scores are crucial for estimating treatment effects in observational studies.
• Logistic regression is a common method for propensity score estimation.
• Machine learning offers potential improvements over traditional methods.

Purpose of the Study:

• To compare the performance of various classification and regression tree (CART)-based propensity score models against logistic regression.
• To evaluate these methods under different scenarios of covariate associations.

Main Methods:

• Simulated data with varying sample sizes (n=500, 1000, 2000) and covariate complexities.
• Propensity score estimation using logistic regression, CART, pruned CART, bagged CART, random forests, and boosted CART.
• Performance evaluation based on covariate balance, standard error, bias, and confidence interval coverage.

Main Results:

• All methods performed adequately when only non-linearity or non-additivity was present.
• Logistic regression showed subpar performance when both moderate non-linearity and non-additivity were present.
• Ensemble methods, especially boosted CART, demonstrated superior bias reduction and more consistent confidence interval coverage.

Conclusions:

• Ensemble machine learning methods, particularly boosted CART, are effective alternatives to logistic regression for propensity score weighting.
• These methods are especially beneficial in complex scenarios with non-linear and non-additive covariate associations.
• Boosted CART shows promise for improving the reliability of causal inference from observational data.