Project Summary
Description: Abstract accessible to people in the field or closely related fields. 30 lines max.
Risk adjustment is widely-used in American healthcare policy - nearly two-thirds of Medicare and
Medicaid dollars are distributed in risk adjusted payments. Risk adjustment is used to adjust outcomes
predictions for patient characteristics. Current risk adjustment models typically use a linear regression and
consistently underestimate bad outcomes for multi-morbid patients. These inaccuracies greatly impair policy-
makers ability to measure quality, modernize payment systems, and incentivize innovation in patient care.
Medicare managed care is one of the largest sources of risk adjusted payments in health policy. It
represents 32% of Medicare spending, 1% of US GDP, and the source of health insurance coverage for 20
million people. Medicare's current risk adjustment methods substantially underestimate costs for high risk,
multimorbid patients. They also overestimate costs for low risk, healthy patients.
Preliminary analysis has identified a critical weakness in traditional risk adjustment models that likely
contributes to these misestimates. Traditional models (including the widely-used CMS-HCC model and ARHQ
Elixhauser index) largely assume that health conditions have a constant effect on costs, with a handful of
exceptions for limited interaction terms. However, in preliminary analysis, all common health conditions cost
substantially more in the presence of comorbidities. Consequently, the costs of health conditions are
underestimated in multimorbid patients and overestimated in healthy patients.
In Medicare Managed Care, inaccurate risk adjustment reduces the accessibility of care for multimorbid
patients and increases the costs of covering healthy patients. Insurers are systematically underpaid to care for
multimorbid patients. As a result, insurers try to avoid enrolling multimorbid patients through a variety of
strategies that ultimately reduce these patients' access to care and harm their health. In addition, Medicare
overpays companies to cover relatively healthy patients, creating incentives to selectively enroll healthy
patients and unnecessarily increasing government expenditures.
My preliminary analysis suggests a promising and actionable direction to improve the accuracy of risk
adjustment – substantially relaxing the assumption of constant effects to create “non-constant effect” models.
At least one new Medicare model has moved in this direction; the proposed project will extend this work by
developing and carefully evaluating a much broader range of non-constant effect models.
The central hypothesis of the proposed research is that non-constant effect models will substantially
improve cost predictions. The specific aims of this project are: 1) Develop non-constant effect models that
incorporate the heterogenous effects of health conditions. 2) Determine if non-constant effect models
meaningfully improve upon the statistical performance of current models for reasonable losses of model
parsimony. 3) Evaluate whether non-constant effect models reduce incentives for selective patient enrollment.