Improving propensity score estimators' robustness to model misspecification using super learner.

Area of Science:

• Epidemiology
• Biostatistics
• Machine Learning in Healthcare

Background:

• Propensity score (PS) estimation is crucial for causal inference, relying on accurate PS model specification.
• Traditional main-effects logistic regression for PS estimation may violate underlying model assumptions.
• Machine learning offers nonparametric alternatives for robust PS estimation.

Purpose of the Study:

• To evaluate the effectiveness of Super Learner (SL), a machine learning method, for propensity score estimation.
• To compare SL-based PS estimation against traditional methods under varying degrees of model misspecification.

Main Methods:

• A simulation study with 1,000 datasets (n=500) was conducted across four scenarios of model misspecification.
• Super Learner (SL) was employed for nonparametric PS estimation.
• Average treatment effects were estimated using PS matching and inverse probability of treatment weighting (IPTW).

Main Results:

• All methods showed adequate covariate balance, but SL demonstrated superior performance with highly unbalanced variables under model misspecification.
• SL-based estimators achieved the lowest bias, especially in cases of severe model misspecification.
• PS prediction accuracy, covariate balance, bias, standard error, coverage, and mean squared error were key evaluation metrics.

Conclusions:

• Utilizing Super Learner (SL) for propensity score (PS) estimation can significantly enhance covariate balance.
• SL-based methods effectively reduce bias in causal inference when the treatment assignment model is seriously misspecified.
• This approach offers a valuable alternative to traditional logistic regression for PS estimation in challenging scenarios.