Statistical Fit and Algorithmic Fairness in Risk Adjustment for Health Policy
share
While risk adjustment is pervasive in the health care system, relatively little attention has been devoted to studying the fairness of these formulas for individuals who may be harmed by them. In practice, risk adjustment algorithms are often built with respect to statistical fit, as measured by p-values or R2 statistics. The main goal of a health plan payment risk adjustment system is to convey incentives to health plans such that they provide health care services efficiently, a component of which is not to discriminate in access or care for persons or groups likely to be expensive. In an attempt to accomplish this, risk adjustment formulas include indicators for the presence of health conditions associated with higher costs. The salient issue is that incentives mainly operate at a group level, not an individual level; plans can discriminate at the group level in ways they cannot at the person level. Because health plans providing sparse care for certain illnesses is a key policy concern, group-level fit is arguably one of the most important metrics for formula evaluation. Giving primacy on the basis of individual fit when group fit may be the larger concern can lead to harmful decision making. We therefore discuss the role of p-values and statistical fit for this policy problem while considering the fairness of the risk adjustment algorithm for vulnerable groups. Enrollees with mental health and substance use disorders have been found to be subject to the adverse incentives noted above. We apply our ideas to this vulnerable group with a group-level net compensation metric of the incentives to health plans to underprovide services.
READ FULL TEXT