Propensity score and proximity matching using random forest.

Area of Science:

• Biostatistics
• Epidemiology
• Machine Learning in Health Research

Background:

• Matching methods are crucial for unbiased inference from observational data by balancing treatment and control groups.
• Traditional propensity score matching has limitations including model mis-specification and handling of complex data.
• Existing methods struggle with categorical variables, missing data, and nonlinear relationships.

Purpose of the Study:

• To introduce novel random forest-based matching methods for improved propensity score estimation and group balancing.
• To address limitations of existing propensity score matching techniques.
• To leverage the capabilities of random forests for more accurate and robust causal inference.

Main Methods:

• Utilizing random forest algorithms for nonparametric propensity score estimation.
• Employing the proximity matrix from random forests as a distance measure for matching.
• Applying proposed methods to National Health and Nutrition Examination Survey (NHANES) data.

Main Results:

• Random forest matching successfully produced well-balanced treatment and control groups.
• The methods demonstrated effectiveness in handling covariates with missing data.
• Achieved more accurate and less model-dependent propensity score estimates.

Conclusions:

• Random forest-based matching offers a powerful, flexible alternative to traditional propensity score methods.
• These novel approaches enhance the reliability of causal inference from observational health studies.
• The methods provide a robust framework for addressing common challenges in real-world data analysis.