Potential Pitfalls and Improvements to the ACA's Risk Adjustment Mechanism
This research examines two questions: 1) how effective is the Affordable Care Act’s (ACA’s) risk adjustment mechanism in limiting insurers’ exposure to selection by enrollees and curtailing incentives to actively select patients (e.g., through marketing, plan/network design, service) on the basis of risk or expected profits and 2) what improvements to the methodology or countermeasures would be feasible and effective in further mitigating selection pressures or their negative effects?
Starting with individual-level data including health care spending and health conditions from years 2005 to 2012 of the Medical Panel Expenditure Survey, we limit the sample to those with private insurance and adjust the sample weights to better reflect characteristics of individual insurance enrollees in 2014. We estimate a model of insured health spending analogous to the HHS-HHC risk adjustment model, which controls for age by gender categories, Hierarchical Condition Categories (HCCs), and selected interaction terms. Using potential predictors of health spending not included in the HHS-HCC model but feasibly observable or proxied by health insurers (such as income, employment status, occupation, and subjective health measures), and estimating insured cost models using market segmentation/predictive modeling methods (chi-squared automated interaction detection [CHAID] and boosted regression), we examine how much “residual predictability” the HHS-HHC model leaves on the table that insurers could feasibly exploit to select (or attract) more profitable enrollees. Beyond standard statistical measures, we quantify the economic value of this residual predictability.
We then conduct simulations over a range of scenarios in a hypothetical market in which two insurers compete in a situation where one insurer has more ability or willingness to engage in selection of enrollees who are more profitable net of premiums and risk adjustment transfers. We consider selection based on diagnostic factors not captured by HCCs, selection on expected profits predicted by easily observed socio-demographic measures, and selection on expected profits predicted by a broader set of measures that may be feasibly proxied (such as self-reported health status).
We find evidence of residual predictability (in a hold-out sample) of sufficient magnitude that if it were strategically exploited at even 10 percent effectiveness, simulations show it could produce negative margins for the other insurer. We explore potential improvements to the current system including: adding additional diagnostic detail not picked up on the current HCCs, adding income and employment status as risk factors, and a permanent reinsurance mechanism.