Racial Bias in Risk Adjustment Algorithms and Implications for Racial Health Disparities: Evidence from Dual-Eligible Medicare/Medicaid Long-term Care Patients in New York

点击翻译按钮获取中文摘要

PROJECT SUMMARY/ABSTRACT A growing body of evidence demonstrates the presence of racial bias in data algorithms. In healthcare, racial bias could arise due to systematic biases in classification and coding, data availability, or data accuracies that differ across racial groups. For instance, algorithms that use data on healthcare costs—rather than illness—to predict need tend to allocate too few resources to Black patients who are underserved by our current system and generate lower spending than white patients with the same health conditions. This issue is increasingly relevant because most U.S. public health insurance programs operate capitated managed care systems, in which beneficiaries enroll in private insurance plans, and the government pays insurers a fixed monthly capitation payment per enrollee. These per-capita payments are typically calculated using risk-adjustment algorithms, in which patient costs are predicted with information on age, gender, and selected health conditions from data on past enrollees. However, race is often excluded from these algorithms, raising the possibility that risk-adjusted managed care could widen racial disparities in care and outcomes among patients. Yet there is little empirical evidence on the impacts of risk-adjusted managed care systems on racial differences in care and health outcomes, especially in high-cost settings, such as long-term care. This project will advance knowledge on these issues by studying the causal effects of risk-adjusted managed long-term care (MLTC) on racial disparities in care and outcomes among dual-eligible Medicare/Medicaid long-term care beneficiaries in New York, using 8 years of administrative data on Medicaid and Medicare enrollment, claims, and assessment records. The project will identify the effects of risk-adjusted MLTC on a range of care utilization and health outcomes, including inpatient, post-acute, nursing home, and at-home care, prescription drug use, and mortality, separately by patient race/ethnicity. Leveraging the county-by-county rollout of managed care mandates, the analysis will use difference-in-differences models to compare within-county changes in outcomes of patients in New York from before to after managed care was implemented. We will estimate separate models by race/ethnicity of the patient, testing for statistical differences in MLTC effects. The project will also identify subgroups who are most severely affected by racial bias in risk-adjustment algorithms, through sub-group analyses that compare effects by gender, age, presence of chronic conditions, and zip code level median income. The project will additionally examine the role of managed care plan features in driving racial disparities in health care utilization and health outcomes. Results will help policymakers, healthcare organizations, providers, and patients to understand the implications of bias in risk-adjustment algorithms on patient health, identify subgroups of patients who are most severely impacted, and learn about effective plan features that could curb or eliminate racial health disparities in managed care settings.