Machine learning for propensity score estimation: A systematic review and reporting guidelines.

Area of Science:

• Statistics
• Computer Science
• Social Sciences

Background:

• Machine learning (ML) methods are increasingly utilized for propensity score (PS) estimation in quasi-experimental research.
• This systematic review synthesizes 179 applications of ML for PS estimation across diverse fields over four decades.

Purpose of the Study:

• To systematically review the application of ML for PS estimation.
• To identify frequently used ML methods, software packages, and reporting practices.
• To highlight deficiencies in reporting critical analysis steps and propose guidelines.

Main Methods:

• Systematic literature review of 179 studies using ML for PS estimation.
• Analysis of ML methods (e.g., gradient boosting machine, random forest), software packages (e.g., R), and reporting of key analysis components.
• Examination of reporting completeness for covariate balance, sensitivity analysis, and hyperparameter configuration.

Main Results:

• Gradient Boosting Machine (GBM) and Random Forest were the most common ML methods for PS estimation.
• Significant underreporting was observed: 48.04% omitted covariate balance evaluation, and 13.97% misused p-values for balance assessment.
• Only 22.8% conducted sensitivity analyses, and 46.9% reported hyperparameters, indicating critical gaps in methodological transparency.

Conclusions:

• While ML methods like GBM are prevalent for PS estimation, current reporting practices are often inadequate.
• Underreporting of covariate balance, sensitivity analyses, and hyperparameter details hinders reproducibility and methodological rigor.
• Guidelines are proposed to enhance the transparent and accurate reporting of ML-based propensity score analyses in research.