Machine Learning vs Logistic Regression in Causal Inference
Publication Title: Machine learning versus logistic regression for propensity score estimation: a trial emulation benchmarked against the PARADIGM-HF randomized trial
Summary
- Question
- This study investigated whether machine learning (ML) algorithms improve the accuracy of propensity score estimation—a statistical technique used to adjust for confounding variables (factors that may distort the relationship between treatments and outcomes) in observational studies. Specifically, the researchers compared ML-based approaches with traditional logistic regression to evaluate treatment outcomes in heart failure patients.
- Why it Matters
- Understanding the effectiveness of ML in propensity score estimation is important for advancing how we interpret real-world health data. Propensity scores help mimic randomized controlled trials by balancing differences between treatment groups. If ML can enhance this process, it could lead to more reliable conclusions about treatments, especially in complex medical scenarios such as heart failure, which affects millions globally. This has implications for researchers, clinicians, and policymakers aiming to improve evidence-based decision-making.
- Methods
- The researchers used data from U.S. veterans with heart failure who received either sacubitril/valsartan or standard ACEI/ARB therapy between 2016 and 2020. They emulated a previous clinical trial to compare three approaches: (1) logistic regression with pre-specified confounders, (2) generalized boosted models (GBM, an ML algorithm) using the same confounders, and (3) GBM with additional data-driven predictors. The study assessed how well each method aligned with the clinical trial results.
- Key Findings
- Logistic regression produced estimates closest to the clinical trial outcomes, indicating a lower risk of mortality for sacubitril/valsartan users compared to ACEI/ARB users. GBM with the same confounders showed no improvement in accuracy. When GBM included additional data-driven predictors, it introduced bias, producing estimates that deviated significantly from the trial results. Despite its ability to predict treatment assignment accurately, ML did not reduce bias in causal inference.
- Implications
- The findings suggest that ML does not inherently improve causal estimates in observational studies and may even introduce errors when combined with automated variable selection. This highlights the importance of careful study design and subject-matter expertise over reliance on algorithmic complexity. For researchers and clinicians, the study reinforces that traditional methods, when applied thoughtfully, can still provide reliable results in health research.
- Next Steps
- The authors recommend further research to explore the conditions under which ML might enhance propensity score estimation. Future studies should also investigate alternative ML algorithms and methods to mitigate overadjustment bias while maintaining interpretability in causal inference models.
- Funding Information
- This research was supported by grants from the National Heart, Lung, and Blood Institute (K23HL141644) and the National Institute on Drug Abuse (K99DA057487). The content is solely the responsibility of the authors and does not necessarily represent the official views of the National Institutes of Health.
Full Citation
Wang K, Rosman L, Lu H. Machine learning versus logistic regression for propensity score estimation: a trial emulation benchmarked against the PARADIGM-HF randomized trial. European Journal Of Epidemiology 2026, 41: 317-328. PMID: 41524886, DOI: 10.1007/s10654-025-01341-7.
This AI-assisted summary has been reviewed and approved by at least one of the study's authors to ensure it accurately reflects the research.
Authors
Kaicheng Wang, MD, MPH
First AuthorHaidong Lu, PhD
Last AuthorAssistant Professor of Medicine (General Medicine) and Epidemiology (Chronic Diseases)
Other Authors
Research Themes
Concepts
- Pre-specified confounders;
- Propensity score approach;
- Logistic regression;
- Causal inference;
- Propensity score methods;
- Overadjustment bias;
- Residual confounding;
- Confounder selection;
- Secondary analysis;
- Trial emulation;
- U.S. veterans;
- Scoring approach;
- Confounding;
- Logistic regression approach;
- Propensity score estimation;
- Propensity score;
- Randomized trials;
- Score estimation;
- Scoring method;
- Machine learning;
- Trials;
- Covariates;
- Regression