Algorithms to Improve Fairness in Medicare Risk Adjustment

Payment system design creates incentives that impact healthcare spending, access, and outcomes. With Medicare Advantage accounting for more than half of Medicare spending, changes to its risk adjustment algorithm have the potential for broad consequences.

To develop risk adjustment algorithms that can achieve fair spending targets, and compare their performance to a baseline that emulates the least squares regression approach used by the Centers for Medicare and Medicaid Services.

Retrospective analysis of Traditional Medicare enrollment and claims data between January 2017 and December 2020. Diagnoses in claims were mapped to Hierarchical Condition Categories (HCCs). Algorithms used demographic indicators and HCCs from one calendar year to predict Medicare spending in the subsequent year.

Data from Medicare beneficiaries with documented residence in the United States or Puerto Rico.

A random 20% sample of beneficiaries enrolled in Traditional Medicare. Included beneficiaries were aged 65 years and older, and did not have Medicaid dual eligibility. Race/ethnicity was assigned using the Research Triangle Institute enhanced indicator.

Prospective healthcare spending by Medicare. Overall performance was measured by payment system fit and mean absolute error. Net compensation was used to assess group-level fairness.

The main analysis included 4,398,035 Medicare beneficiaries with a mean age of 75.2 years and mean annual Medicare spending of $8,345. Out-of-sample payment system fit for the baseline regression was 12.7%. Constrained regression and post-processing both achieved fair spending targets, while maintaining payment system fit values of 12.6% and 12.7%, respectively. Whereas post-processing only increased mean payments for beneficiaries in minoritized racial/ethnic groups, constrained regression increased mean payments for beneficiaries in minoritized racial/ethnic groups and beneficiaries in other groups residing in counties with greater exposure to socioeconomic factors that can adversely affect health outcomes.

Constrained regression and post-processing can incorporate fairness objectives in the Medicare risk adjustment algorithm with minimal reduction in overall fit.