Affordable Care Act Risk Adjustment: Overview, Context, and Challenges

Different Plan Actuarial Value Levels versus a Standard Benefit Level

The ACA establishes four tiers of plan actuarial value, or “metal levels,” plus a catastrophic plan in a separate risk pool. The metal levels are platinum, gold, silver, and bronze, which correspond, respectively, to plans that are expected to pay 90, 80, 70, and 60 percent of the medical expenditures of a standard population. Although all plans must cover “essential health benefits,⁶ ” the metal levels are defined by the cost sharing (deductibles, copayments, and coinsurance) the enrollee pays (not by the medical services the plan covers). The varying metal levels are in contrast to Medicare Advantage and Medicare Part D, where plans provide, at minimum, a standard coverage level.

The presence in the market of plans with different actuarial values poses two key considerations for the risk adjustment methodology. The first consideration is how to preserve premium differences that reflect differences in generosity of plan coverage. The risk transfers should counteract the effects of risk selection, but should not adjust away actuarial value differences among plans. A related issue that the ACA’s metal level actuarial value tiers create is differences in induced demand across plans, meaning that enrollees in plans with higher actuarial values are expected to use more services because of lower cost sharing. The policy goal is for the premiums that plans charge enrollees to reflect the different actuarial values of plans (and associated induced demand), but not to reflect the health status of enrollees. Other things equal, an individual should pay a higher premium for a platinum plan than a bronze plan to reflect the reduced cost sharing the individual pays when enrolled in the platinum plan. But, other things equal, an individual should not pay more to enroll in a platinum plan because it has sicker enrollees on average than the bronze plan. The second consideration raised by the presence of plans with different actuarial values is how to develop risk scores that appropriately reflect a given enrollee’s actuarial risk to a plan in light of the fact that plans pay a different portion of an enrollee’s total expenditures, depending on the plan metal level.

Allowed Rating Factors versus Uniform Premiums

The ACA allows individual and small group plans to rate premiums on four factors: age, tobacco use, family size, and geographic rating area. The age variation in premiums is constrained to 3:1 for 21 year olds and older, and the variation based on tobacco use is constrained to 1.5:1. In contrast, Medicare Advantage and Part D plans are required to charge all enrollees uniform premiums.⁷ In the presence of age rating variation, if a plan obtains higher revenues by charging its older enrollees more, it should not also be fully compensated for age variations through risk transfers. Age predicts medical expenditures and is typically included in risk adjustment models. How should the allowed premium rating for age be netted out of risk transfers?

Geographic rating area is the fourth source of allowed rating variation. ACA individual and small group markets are established within states. But states may elect to define multiple intra-state rating areas across which plans can vary premiums.⁸ Given that risk pools are defined within states, how should risk transfers differ when the “base” level of premiums and costs differs across rating areas? More generally, how can a methodology be established that is flexible enough to potentially be applied to all 50 states, with their different cost levels?

Employer-Sponsored versus Medicaid Data to Calibrate a Risk Adjustment Model.

Projections of the characteristics of the long-run (2019) ACA individual market population (both inside and outside the Marketplaces) have been made in comparison to the characteristics of employer-sponsored insurance enrollees and Medicaid enrollees (Trish, Damico, Claxton, Levitt, & Garfield, 2011). Although many projected characteristics of the ACA individual market enrollees lay between the characteristics of enrollees in employer-sponsored insurance and Medicaid enrollees, on average they tend to be closer to enrollees in employer-sponsored insurance. For this reason, we focused on claims data from employer-sponsored insurance to calibrate the HHS-HCC risk adjustment model. The specific employer-sponsored insurance claims dataset we chose is discussed in the companion article on the empirical risk adjustment model.

Prospective versus Concurrent Model

Risk adjustment models can only utilize available information to predict expenditures. Most risk adjustment models used for payment are “prospective,” meaning they use prior year information to predict current year medical expenditures. For example, the Medicare Advantage and Medicare Part D models are prospective. Prospective models tend to be favored because they emphasize the impact of ongoing chronic conditions on costs (as opposed to random current year costs that can be pooled as “insurance risk”).

However, for the first year of the ACA-established individual and small group markets in 2014, no previous year information on health status exists. A prospective model is, therefore, infeasible for the first year of the ACA state markets, and given the time required to accumulate and analyze data and pre-announce the model, it is realistically infeasible for at least the first several years of the Marketplaces. Even after the first few years of operation of the Marketplaces, assembling the data for a prospective risk adjustment model would be very challenging. For example, there are likely to be substantial flows of individual/small group participants among insurance statuses, including to/from Medicaid, to/from large-employer-based insurance, and even to/from uninsured status.

For these reasons, the 2014 HHS-HCC risk adjustment model is “concurrent,” meaning current year information is used to predict current year costs. Concurrent models tend to emphasize the prediction of costs associated with current year acute health events. However, a considerable amount of the costs of chronic conditions are associated with acute exacerbations. Concurrent models can also capture the very high costs of conditions, such as organ transplants, metastatic cancer, and low-birthweight babies, that reduce or eliminate the disincentive for plans to contract with providers who treat these conditions. In developing the concurrent model, we attempted to focus on conditions associated with systematic selection risk of enrollees or providers and to de-emphasize conditions such as injuries that are probably not a focus of plan selection behavior. We also adopted approaches intended to lessen the influence of differences in diagnostic coding patterns on risk scores, as described in more detail in the second companion paper. Further, because concurrent risk adjustment explains more of the variation in current (acute) costs, it reduces unsystematic risk, which may benefit small health plans that do not have enough enrollees to diversify away unsystematic risk. Finally, we include partial year enrollees in the sample to calibrate the risk adjustment model because, with a concurrent risk adjustment model, enrollees’ diagnoses will match their utilization for any period of enrollment. All enrollees (with at least one month of enrollment), including newborns and decedents—some of whom are typically among the highest-cost enrollees—are reflected in risk adjustment.

Revised Clinical Classification and Subpopulation Models

The HHS risk adjustment approach predicts expenditures using only enrollees’ age, sex, and diagnoses. Diagnosis is a key clinical factor that drives medical treatment decisions and costs, and is widely used in risk adjustment models (Lodh, Raleigh, Uccello, & Winkelman, 2010). Conceptually, diagnosis is distinct from treatment or utilization, and basing risk adjustment on diagnosis is neutral with respect to treatment modality and utilization. The heart of the empirical risk adjustment model is the clinical classification system that organizes the thousands of International Classification of Diseases (ICD) diagnosis codes into a coherent system of diagnostic categories.

The starting point for the HHS risk adjustment diagnostic clinical classification was the Centers for Medicare & Medicaid Services’ Hierarchical Condition Categories (CMS-HCC) clinical classification (Pope et al., 2004). The CMS-HCCs had to be adapted for three main reasons, which are elaborated on in the second companion paper: 1) prediction year—the CMS-HCC risk adjustment model is prospective rather than concurrent; 2) population—the CMS-HCCs were developed using data from the aged (age ≥ 65) and disabled (age < 65) Medicare populations, as compared to the private individual and small group, primarily under age 65, population; and 3) type of spending—the CMS-HCCs are configured to predict medical spending excluding outpatient prescription drug spending as compared to medical and prescription drug spending. We call the revised clinical classification that is the basis of HHS risk adjustment the HHS-HCC clinical classification.

Separate Adult, Child, and Infant Models

In addition to revising the Medicare CMS-HCC clinical classification to be applicable to the individual and small group markets that are largely under the age of 65, we considered subpopulation differences within the ACA risk adjustment population. Clinical reasoning and empirical investigation led us to conclude that separate adult (age 21+), child (age 2–20), and infant (age 0–1) models are desirable for the risk adjustment population.

Plan Liability Versus Total Expenditures

To account for differences in plan actuarial risk across actuarial value levels, we considered plan liability and total expenditure risk scores. A person’s total expenditure risk score predicts total medical expenditures. In contrast, a plan liability risk score predicts the medical expenditures that a plan is actually liable for, given its actuarial value and cost sharing structure. It incorporates the predicted effect of both health status and plan cost sharing on expected plan liability. An individual has a different plan liability risk score depending on what metal tier of plan he/she enrolls in.

The plan liability risk score cannot be obtained by simply multiplying a person’s total expenditure risk score by his/her plan’s actuarial value because the amount plans pay is not constant as expenditures increase (i.e., it is non-linear, primarily because of the presence of deductibles). We instead estimate separate plan liability models on the same population to determine each enrollee’s plan liability risk score.

Induced Demand Due to Cost Sharing Reductions

We also considered how to address the potential higher utilization among individuals who are enrolled in cost sharing reduction plans. A direct adjustment in the risk adjustment model for induced demand due to cost sharing reductions was not possible due to a lack of the required data in the risk adjustment model calibration sample. As an alternative, a multiplicative adjustment to the risk score was developed. We chose to account for induced demand associated with a more generous actuarial value of cost sharing reduction plans in the risk adjustment model, because premiums for cost sharing reduction plans are required to be the same for all actuarial value levels of cost sharing reduction plans (in contrast to differing metal levels where premiums can vary). For the Medicare Advantage program, induced demand due to lower cost sharing for Medicare-Medicaid dual eligible beneficiaries is adjusted for directly in the risk adjustment model by including a risk factor for dual eligible status. Similarly, for the Part D program, induced demand due to lower cost sharing for low-income beneficiaries is adjusted for directly in the risk adjustment model by calibrating separate models for low-income beneficiaries.