Machine learning versus logistic regression for propensity score estimation: a trial emulation benchmarked against...

Area of Science:

• Epidemiology
• Biostatistics
• Health Informatics

Background:

• Machine learning (ML) is increasingly used for propensity score estimation to enhance causal inference.
• The effectiveness of data-driven ML approaches for confounder selection and adjustment in causal inference remains uncertain.

Purpose of the Study:

• To benchmark ML-based propensity score methods against traditional logistic regression for causal inference.
• To evaluate the impact of automated feature selection in ML propensity score models.

Main Methods:

• Emulated a secondary analysis of the PARADIGM-HF trial using observational data from U.S. veterans (2016-2020).
• Compared three propensity score approaches: logistic regression (pre-specified confounders), generalized boosted models (GBM) with pre-specified confounders, and GBM with expanded covariates and automated feature selection.

Main Results:

• Logistic regression yielded estimates closest to the trial results.
• GBM with pre-specified confounders showed no improvement over logistic regression.
• GBM with automated feature selection introduced substantial bias, increasing estimation error.

Conclusions:

• ML-based propensity score methods do not inherently improve causal estimation.
• Automated feature selection in ML models may introduce overadjustment bias.
• Careful confounder specification and causal reasoning are crucial for reliable causal inference, outweighing algorithmic complexity.