Diagnosis-Aware Cost Sharing: Welfare Gains from Changes in Cost-Sharing Provisions of Health Insurance Contracts

Background: The use of risk adjustment software by newly created health insurance exchanges makes understanding this adjustment in the 18-65 year old market crucial. To stabilize premiums against adverse selection, payment transfers will be made between insurers based on risk adjustment scores of enrollees. As evidenced from the Massachusetts health insurance exchange, the length of time individuals will be enrolled in a policy purchased from the exchange may vary due to dropping in and out of the private insurance market from employer sponsored coverage. Annualized scores will be used on the federal exchange for risk adjustment payments to insurers.

Objectives: We examine the effects of changing the duration of risk adjustment across the distribution of total health spending. Quantifying differences in the amount of adjustment given by selected time periods by estimating the optimal model for each time period allows us to show gains to alternate time periods at predicting spending and analyzing selection incentives. 

Data: We use data from the MarketScan health insurance claims database for 2008-2010. This database includes enrollment and claims data for employees and dependents enrolled in selected employer sponsored insurance plans.

Methods: We use a sample of 6,875,767  enrollees continuously enrolled in employer sponsored insurance plans for a three-year period. To calculate risk adjustment payments to insurers, we use the Hierarchical Condition Categories risk adjustment software for the federal exchanges developed by the Department of Health and Human Services. Rather than relying on the annualized coefficients, we use the binary condition indicators to estimate the relationship between the presence of each condition and spending within a given period. We estimate regression models of realized health spending using different time periods – one month, three months, six months, 12 months, 18 months, and 24 months – as well as concurrent versus prospective models. To compare models, we use the R² and predictive mean ratios; we compare a one year period calculated with each of the shorter time periods compared to the one year period calculated with one year model. For the 18 and 24 month periods, we divide the predicted spending by 1.5 and 2, respectively, to compare to the predicted annual spending. To compare models, we rely on R² values for the entire distribution and predictive mean ratios based on quintiles of realized annual health expenditures.

Results: Among concurrent models, R² values increase with increasing duration of risk adjustment information from monthly to annual time periods. However, R² values decrease from 12 to 18 months and again from 18 to 24 months. The models using quarterly HCC information perform substantially better than the models using monthly HCC information. Moving from quarterly to longer time periods (six and 12 months), improvements are modest up to one year; however, the highest and lowest quintiles of the distribution are better predicted. Concurrent models perform better than prospective models for the whole distribution of health expenditure as well as in each quintile of spending.