Can We Train Machine Learning Methods to Outperform the High-dimensional Propensity Score Algorithm?

Area of Science:

• Health Informatics
• Epidemiology
• Biostatistics

Background:

• Retrospective health care claims datasets often lack complete information on confounders.
• High-dimensional propensity score (HDPS) algorithms use patient data as proxies for unobserved confounders to reduce bias.
• Machine learning (ML) offers alternative methods for confounder selection in complex datasets.

Purpose of the Study:

• To compare the performance of the HDPS algorithm with popular ML methods for confounder selection.
• To evaluate the effectiveness of these methods in a real-world cohort study of post-myocardial infarction statin use.
• To assess a hybrid approach combining ML and HDPS.

Main Methods:

• Utilized a cohort study of post-myocardial infarction statin use (1998-2012).
• Compared HDPS algorithm with random forest, least absolute shrinkage and selection operator (LASSO), and elastic net for confounder selection.
• Employed a plasmode framework to mimic empirical data for bias-based analysis.

Main Results:

• ML methods performed comparably to the HDPS algorithm when applied with sound epidemiologic principles.
• A hybrid ML and HDPS approach demonstrated slightly improved performance in terms of mean squared error.
• Both approaches effectively addressed bias stemming from unobserved confounders in claims data.

Conclusions:

• Machine learning methods are viable alternatives for confounder selection in high-dimensional health care claims data.
• Hybrid ML-HDPS strategies may offer enhanced bias reduction in epidemiological studies.
• Careful application of analytical principles is crucial for accurate results from claims data analysis.