Machine learning methods for propensity and disease risk score estimation in high-dimensional data: a plasmode...

Area of Science:

• Epidemiology
• Biostatistics
• Machine Learning

Background:

• Propensity score (PS) and disease risk score (DRS) estimation are crucial for causal inference in observational studies.
• Machine learning (ML) offers scalable alternatives for PS estimation, but its performance in DRS estimation is not well-understood.

Purpose of the Study:

• To compare the performance of ML methods against traditional logistic regression for PS and DRS estimation.
• To evaluate ML methods using real-world data and plasmode simulations.

Main Methods:

• A cohort study of 632,201 UK primary care patients comparing antihypertensive users and non-users.
• Plasmode simulations with synthetic data to assess bias and covariate balance.
• Comparison of four methods: logistic regression (reference), LASSO, Multi-layer Perceptron (MLP), and Extreme Gradient Boosting (XgBoost).

Main Results:

• ML methods, particularly XgBoost, generally outperformed the reference logistic regression for PS estimation in terms of covariate balance and bias.
• Disease risk score (DRS) estimation methods performed worse than PS estimation methods across all scenarios.
• XgBoost demonstrated the highest performance among the evaluated ML methods for PS estimation.

Conclusions:

• ML methods are reliable alternatives for propensity score (PS) estimation in observational research.
• ML-based DRS methods were less effective than PS methods, potentially due to outcome rarity.
• The findings support the use of advanced ML techniques for improving causal inference.