Machine Learning Algorithms to Estimate Propensity Scores in Health Policy Evaluation: A Scoping Review.

Area of Science:

• Health policy analysis
• Causal inference methodologies
• Machine learning applications

Background:

• Quasi-experimental designs are prevalent for health policy impact evaluation.
• Non-randomized treatments introduce bias, often mitigated by propensity score (PS) methods.
• Machine learning (ML) offers advanced techniques for PS estimation.

Purpose of the Study:

• To conduct a scoping review on the utilization of ML algorithms for PS estimation.
• To identify ML models used for PS estimation and assess their accuracy.
• To characterize studies employing ML for causal inference in health policy.

Main Methods:

• Scoping review methodology adhering to PRISMA-ScR guidelines.
• Comprehensive literature search across multiple scientific and gray literature databases.
• Focus on identifying ML models, their performance, and study characteristics for PS estimation.

Main Results:

• Seven studies were included from 3018 references.
• Tree-based ML models were predominantly used for PS estimation.
• Performance metrics for ML models were frequently not reported or discussed.

Conclusions:

• ML algorithms are increasingly explored for PS estimation in health policy evaluation.
• Under-reporting of model development and evaluation hinders comprehensive understanding.
• Further research should emphasize transparent reporting of ML model performance in causal inference studies.