Using machine-learning algorithms to improve imputation in the medical expenditure panel survey

Area of Science:

• Health economics
• Data science
• Survey methodology

Background:

• The Medical Expenditure Panel Survey (MEPS) imputes healthcare expenditures using predictive mean matching (PMM) with linear regression.
• Accurate imputation is crucial for understanding healthcare utilization and costs.

Purpose of the Study:

• To evaluate the feasibility and effectiveness of applying machine learning (ML) methods to enhance imputation in the MEPS.
• To compare ML algorithms against traditional linear regression within the PMM framework.

Main Methods:

• Replaced the linear regression model in PMM with five ML alternatives: Gradient Boosting, Random Forests, Extreme Random Forests, Deep Neural Networks, and a Stacked Ensemble.
• Introduced an alternative matching scheme using a vector of predicted expenditures by sources of payment, rather than a single total expenditure.

Main Results:

• ML algorithms outperformed Ordinary Least Squares (OLS) in both prediction and matching imputation.
• The Stacked Ensemble approach yielded the best results, improving expenditure prediction R² by 108% and final imputation R² by 227%.
• Matching on a prediction vector enhanced the alignment of payment sources between imputed and observed data.

Conclusions:

• ML algorithms and the novel matching scheme substantially improve the quality of MEPS expenditure imputation.
• These advanced imputation techniques hold potential value for other national surveys utilizing PMM or similar imputation methods.