On kernel machine learning for propensity score estimation under complex confounding structures.

Area of Science:

• Biostatistics
• Health Informatics
• Machine Learning

Background:

• Post-marketing data, often from electronic health records (EHRs), are valuable for clinical research but present analytical challenges due to complex confounding.
• Massive datasets in EHRs introduce computational difficulties for traditional statistical methods.

Purpose of the Study:

• To propose a statistically robust and computationally efficient propensity score (PS) method for comparative effectiveness analysis using observational data.
• To address complex confounding structures, including nonlinear and nonadditive interactions, within large-scale EHR datasets.

Main Methods:

• A kernel-based machine learning approach for flexible and robust propensity score modeling.
• Utilizing estimated propensity scores in a second-stage analysis to determine average treatment effects.
• Implementing a split-and-merge algorithm to manage big data computational demands and provide variance estimation.

Main Results:

• The proposed propensity score method demonstrated statistical robustness and computational efficiency.
• The kernel-based machine learning approach effectively handled complex confounding structures in observational data.
• The split-and-merge algorithm facilitated scalable analysis of large EHR datasets.

Conclusions:

• The developed propensity score method offers a practical and effective solution for comparative effectiveness research using real-world evidence.
• The approach outperforms competing methods in handling complex confounding and big data challenges.
• This methodology enhances the utility of EHR data for informing clinical practice and policy.