Stanford Health Policy Professors Marissa Reitsma, PhD, and Sherri Rose, PhD, joined Harvard Medical School Health Economics Professor Thomas G. McGuire, PhD, to develop risk adjustment algorithms that can promote fairer spending for minoritized racial and ethnic groups. They analyzed a random sample of Medicare fee-for-service beneficiaries to assess existing levels of net compensation, which is the difference between predicted and observed spending, by racial and ethnic group. They proposed a basic measure of health-care spending disparity that informed potential fair spending targets and then developed and evaluated two algorithms to achieve fair spending targets.

Here is an excerpt from their policy brief for the Stanford Institute for Human-Centered AI and Stanford Health Policy:

Medicare spending accounts for 14% of the federal budget and Medicare Advantage accounts for more than half of Medicare spending. Medicare Advantage plans receive risk-adjusted payments for each beneficiary they enroll. Currently, risk adjustment is based on a least squares regression, which generates spending predictions from observed data (i.e., beneficiaries’ demographic characteristics and clinical conditions). Prior research has examined approaches to improve the risk adjustment algorithm by increasing the accuracy of spending predictions, mitigating opportunities for upcoding (i.e., where providers document more severe conditions to increase payments), and reducing the potential for favorable selection (i.e., where insurers aim to attract profitable beneficiaries and avoid enrolling unprofitable beneficiaries).

However, few prior studies have examined fair regression methods, which optimize for both overall and group-level performance. An important finding from previous studies is that adding marginalized group indicators as predictors in the risk adjustment algorithm can reinforce data-embedded inequities in spending between populations, necessitating alternative approaches. Past literature has not specifically examined algorithms to achieve fairness goals across multiple minoritized racial and ethnic groups in Medicare.

This research gap matters for several reasons. First, a greater percentage of Black, Hispanic, and Asian/Pacific Islander beneficiaries, compared to non-Hispanic white beneficiaries, are enrolled in Medicare Advantage plans. Second, the population aged 65 years and older, which is the largest Medicare-eligible population, is projected to become more racially and ethnically diverse in the coming decades. Third, although Medicare eligibility reduces racial and ethnic disparities in insurance coverage, disparities persist in healthcare access, utilization, spending, and outcomes. Historical fee-for-service spending data, which are used to estimate risk scores and determine payments, embed many of these long-standing disparities.

The researchers introduced two algorithms to improve fairness in Medicare Advantage plan payments: constrained regression and post-processing. They evaluated their impacts using enrollment and claims data from more than 4.3 million beneficiaries.

Go To Full Policy Brief