Association for Public Policy Analysis & Management

Risk adjustment is used in insurance markets to adjust payments for individuals’ health conditions. Risk adjustment models typically take the form of a regression model with coefficients for demographics and documented health conditions. However, current risk adjustment models are consistently inaccurate at tails of the risk distribution. In particular, they consistently underestimate spending for high risk, multi-morbid individuals while overestimating spending for low-risk, healthy individuals. This inaccuracy creates substantial selection incentives.

First, I propose an explanation for this consistent misprediction. Current models assume that health conditions have a constant or minimally stepped marginal effect on health care costs. However, for many health conditions, costs may increase substantially in the presence of other health conditions. If so, current and proposed risk adjustment formulas will necessarily mispredict costs in the manner observed.

Next, I evaluate and find evidence for this claim in Truven Marketscan commercial claims data. For the 10 most frequently occurring health conditions, the marginal effect of a health condition on costs increases monotonically in the number of other health conditions. This empirical pattern suggests that risk adjustment methods need to account more comprehensively for interactions between health conditions.

Accounting for interactions leads to a high-dimensional statistical problem, which necessitates the use of tools from machine learning. I evaluate several machine learning based tools for risk adjustment that allow for heterogenous marginal effects. Model performance is evaluated throughout the distribution of predicted risk. It is also benchmarked against models similar to widely used risk adjustment models.