Alternative evaluation metrics for risk adjustment methods.

Area of Science:

• Health Economics
• Biostatistics
• Health Services Research

Background:

• Risk adjustment is crucial for equitable healthcare payments, aiming to match payments with expected expenditures.
• Current methods often lead to residual risks at the individual level, particularly for high-cost patients, incentivizing health plans to avoid them.
• This risk selection poses a challenge to the fairness and efficiency of healthcare markets.

Purpose of the Study:

• To comprehensively compare the predictive accuracy of 19 different risk adjustment estimators.
• To evaluate performance across group-level, tail distribution, and individual-level predictions.
• To identify an optimal risk adjustment strategy considering trade-offs between different accuracy metrics.

Main Methods:

• A comparative analysis of 19 estimators, including parametric regression, machine learning, and distributional models.
• Utilized the 2013-2014 MarketScan database for empirical evaluation.
• Assessed prediction accuracy at group, tail, and individual levels.

Main Results:

• No single estimator demonstrated superior performance across all evaluated metrics.
• Machine learning and distribution-based estimators generally outperformed parametric regression at the group level.
• Parametric regression estimators showed higher accuracy for tail distributions and individual-level predictions, especially for high-expenditure individuals.

Conclusions:

• A trade-off exists between group-level payment accuracy and minimizing individual-level residual risk.
• Optimal risk adjustment method selection requires considering statistical performance alongside the simulation of health plan risk selection behaviors.
• Statistical metrics alone are insufficient for determining the best risk adjustment approach.