Methods for Evaluating and Improving Plan Payment Risk Adjustment and Plan Defaults in Medicare Part D

Abstract

Chapter one expands on concepts from the statistics, computer science, and health economics literature to develop new fair regression methods for health plan payment risk adjustment formulas that build fairness considerations directly into the objective function. A suite of metrics is proposed as necessary to evaluate risk adjustment formulas more fully, including a novel measure of fair covariance. The data application uses the IBM MarketScan Research Databases to measure and reduce undercompensation for individuals with Mental Health and Substance Use Disorders in the Individual Health Insurance Marketplaces. These new fair regression methods can lead to massive improvements in group fairness (e.g., 98%) with only small reductions in overall fit (e.g., 4%).

Chapter two develops a machine learning method for "group importance" to identify unprofitable groups defined by multiple attributes in the risk adjustment formula. This approach is designed to evaluate the risk adjustment formulas used in the U.S. health insurance Marketplaces and Medicare. A number of previously unidentified groups with multiple chronic conditions are found to be undercompensated in the Marketplaces’ risk adjustment formula, while groups without chronic conditions tend to be overcompensated in the Marketplaces. The magnitude of undercompensation when defining groups with multiple attributes is larger than with single attributes. No complex groups were found to be consistently under- or overcompensated in the Medicare risk adjustment formula.

Chapter three evaluates whether smart defaults for plan assignment of Low-Income Subsidy recipients in Part D improves upon current random default procedures, which may lead to underutilization of needed prescription drugs. Assignment to plans that better match previous prescription drug use increases adjusted prescription drug spending by $115 on average compared to a random default. Furthermore, there is heterogeneity across beneficiaries in the spending differences from assignment to high fit plans. Smart defaults limited to beneficiaries that are estimated to benefit the most still outperforms random defaults.