Collaborative targeted learning using regression shrinkage.

Area of Science:

• Causal inference
• Semiparametric statistics
• Machine learning in epidemiology

Background:

• Model selection and covariate set reduction are critical for efficient causal inference.
• Existing Collaborative Targeted Minimum Loss-based Estimation (CTMLE) adaptively limits propensity score model complexity.
• Cross-validation is used to assess errors based on outcome models.

Purpose of the Study:

• To generalize stepwise variable selection CTMLE using regression shrinkage.
• To introduce two new algorithms for stepwise penalization parameter selection in regression shrinkage.
• To evaluate the performance of these methods in reducing bias and mean squared error.

Main Methods:

• Generalization of CTMLE with propensity score regression shrinkage.
• Development of two novel algorithms involving stepwise penalization parameter selection.
• Simulation studies under misspecified outcome models.
• Application to electronic medical data for asthma therapy safety evaluation.

Main Results:

• The proposed CTMLE procedures with penalized regression shrinkage reduce mean squared error and bias.
• Penalizing individual covariates in the propensity score is effective, particularly with misspecified outcome models.
• The methods are demonstrated on real-world electronic medical data.

Conclusions:

• Regression shrinkage offers a powerful generalization of CTMLE for causal inference.
• The new algorithms improve estimation efficiency by adaptively controlling propensity score model complexity.
• These methods provide robust tools for causal effect estimation in complex observational studies.