Concurrent Risk Adjustment

Concurrent Risk Adjustment in Health Plan Payment

Concurrent risk adjustment explains each enrollee's spending in a year with the diagnoses and demographic details recorded during that same year, then uses the prediction to move money between health plans or to put cost comparisons on a common footing. It matters most in health insurance markets where insurers are paid by capitation or community-rated premiums, because it decides how much a plan that enrols sick people is compensated for them. The best-known current use is the HHS-HCC model behind Affordable Care Act (ACA) risk adjustment in the United States individual and small group markets, on and off the Marketplaces. The choice of a concurrent rather than a prospective design trades better statistical fit against weaker incentives to control costs and more room for coding behaviour.

Same-period information and the prediction equation

Every risk adjustment model has to fix which period supplies the risk factors and which period supplies the spending to be explained. Ellis describes the more common design, prospective risk adjustment, as using information from a base period, usually one year, to predict spending in the following period, while a concurrent model uses the diagnoses or other information from a year to predict spending or utilisation in that same year. Kautter and colleagues draw the same contrast for the HHS-HCC model.

$$\hat{C}_{it} = \beta_0 + \sum_{k=1}^{K} \beta_k X_{ikt}$$

where $\hat{C}_{it}$ is predicted spending for person $i$ in period $t$, $X_{ikt}$ is the $k$th risk factor (an age-sex cell, a condition category or an interaction) recorded for that person in period $t$, $K$ is the number of risk factors and $\beta_0$ to $\beta_K$ are coefficients estimated by regressing actual spending on the risk factors in a calibration sample. A prospective model has the same form with $X_{ik,t-1}$ in place of $X_{ikt}$.

The prediction is usually turned into a relative risk score.

$$R_{it} = \frac{\hat{C}_{it}}{\bar{C}_t}$$

where $R_{it}$ is the risk score and $\bar{C}_t$ is mean predicted spending in the reference population for period $t$, so that a score of 1.0 marks an average enrollee. The HHS-HCC risk score is defined this way, with 1.0 equal to average plan liability spending in the calibration sample.

Many people are enrolled for only part of a year. The HHS-HCC models therefore divided plan liability spending by the fraction of months each person was enrolled and were estimated by weighted least squares, weighting each person by that same fraction, an approach Ellis traces to Ash and colleagues.

Concurrent, prospective and retrospective as names

The naming in this area is inconsistent, and the word retrospective carries two different meanings. Ellis, in his 2008 review chapter and his New Palgrave Dictionary entry, calls the same-period approach concurrent and notes that it is sometimes called retrospective. In that usage retrospective refers to the data window of the model.

The second meaning concerns when payments are settled. Ellis noted that the prospective CMS-HCC model in Medicare Advantage still used retrospective adjustments to reflect late-arriving data. The American Academy of Actuaries makes the converse point, that a concurrent model reflects new diagnoses arising during the year, so retrospective payment adjustments are typically required. A concurrent model therefore typically needs settlement after the year, but settlement after the year does not imply a concurrent model. Retrospective risk adjustment in the settlement sense means payment adjustments calculated after the coverage period ends, so a clear description of any scheme gives the diagnosis window, the spending window and the settlement date separately.

Where concurrent models are used

Most payment systems use prospective models. In his dictionary entry Ellis gives two reasons, concern about the endogeneity of same-period signals and the practical advantage of measuring risk factors a year before the spending they predict. Geruso and McGuire list Medicare Advantage, some state Medicaid programmes and public health systems across Europe as prospective users and describe concurrent risk adjustment as nearly unique to the ACA Marketplaces.

Kautter and colleagues give the HHS reasons. Prior-year diagnoses would not be available for the first year of implementation in 2014, and people move in and out of individual and small group coverage, so prior-year data would be missing for many enrollees even later. The original HHS-HCC models were calibrated on 2010 MarketScan commercial claims, with fifteen separate models for three age groups (adult, child and infant) and five levels of plan cost sharing. For the 2025 benefit year CMS recalibrated them on 2019 to 2021 EDGE data.

The resulting transfers are calculated after the benefit year closes. The CMS summary report for the 2025 benefit year states that the methodology transfers funds from plans with lower-than-average risk enrollees to plans with higher-than-average risk enrollees. The stated aim is to reduce issuers' incentives to avoid high-risk enrollees and to lessen the influence of risk selection on premiums. Risk selection here covers cream skimming by insurers as well as adverse selection by consumers.

Outside payment, Ellis notes that a concurrent framework is also used for quality measurement and to normalise performance measures. In his worked comparison of five types of United States health plan, spending by HMO and capitated point-of-service plans diverged by about 20% from the others by age 40, but after adjustment with a concurrent DCG/HCC model the five lines differed by less than 5%. He read most of the raw difference as selection, an example of case-mix adjustment of claims data in health services research.

Explanatory power and what R-squared shows

Concurrent models explain much more of the variation in individual spending than prospective ones, because the spending being explained comes from the period that generated the diagnoses. Kautter and colleagues report R-squared values of 35% to 36% for the adult HHS-HCC models, 30% to 31% for the child models and 29% for the infant models. They compare this with generally less than 2% for demographic-only models and 10% to 15% for typical prospective diagnosis-based models such as the CMS-HCC.

The Society of Actuaries study by Hileman and Steele tested 19 concurrent and 23 prospective models from ten sources on a sample of one million people from 2012 and 2013 MarketScan data. Without censoring of high costs, R-squared for concurrent diagnosis-only models ranged from 24.2% to 52.7%, against 9.1% to 20.7% for prospective diagnosis-only models. The HHS-HCC concurrent model scored 41.3%. Capping individual costs at USD 250,000 raised every R-squared value, and the authors stress how strongly a few extreme values drive the statistic.

The gap narrows when individuals are pooled. For simulated groups of 10,000, mean absolute error for the concurrent diagnosis-only models ranged from 2.1% to 2.7%, against 2.8% to 3.1% for their prospective counterparts. Plan-level transfers depend on group averages, so for randomly formed groups the individual R-squared gap overstates the practical difference, although real plans are not random samples of the market.

Incentives for efficiency, coding and selection

A concurrent payment rises when a diagnosis is recorded during the year, and diagnoses are recorded when care is delivered. Geruso and McGuire measured this with the idea of power, the share of a cost saving that a plan keeps. They estimated that concurrent risk adjustment more than doubles fit compared with prospective risk adjustment in the Marketplace setting but reduces power considerably. In their published estimates, removing 10% of inpatient events cut payment by about USD 0.38 for each dollar of cost removed, a power of about 0.62.

The American Academy of Actuaries describes concurrent models as possibly less resistant to gaming, because current diagnoses and treatments affect payment. The HHS-HCC condition list responds to this. Kautter and colleagues report criteria excluding categories especially subject to discretionary coding, and grouping that lets related conditions count only once. They state that the models were built to balance predictive power against sensitivity to discretionary coding rather than purely to maximise R-squared.

On risk protection and selection the balance tips the other way. A concurrent model pays for acute events such as a fracture in the year they occur, and the Academy notes that such events predict current costs without necessarily predicting later ones. A prospective model leaves those costs uncompensated in the year they occur, so plans bear more of the risk of unpredictable spending. Hileman and Steele found prospective models underpredicting costs for people with a condition recorded in the prediction year, partly because some had no record of it the year before.

Worked example with three enrollees and two plans

All numbers in this example are illustrative and simplified. A concurrent model fitted to a hypothetical calibration sample gives an adult age-sex coefficient of GBP 1,200, GBP 2,400 for diabetes, GBP 6,000 for heart failure and GBP 3,000 for a hip fracture. Mean predicted spending in the market is GBP 3,000.

1. Predict spending for each enrollee. Enrollee A has no recorded condition, so predicted spending is GBP 1,200. Enrollee B is diagnosed with diabetes during the year, giving 1,200 + 2,400 = 3,600. Enrollee C has heart failure and a hip fracture in the same year, giving 1,200 + 6,000 + 3,000 = 10,200.

2. Convert to risk scores. Dividing by the market mean gives 1,200 / 3,000 = 0.40 for A, 3,600 / 3,000 = 1.20 for B and 10,200 / 3,000 = 3.40 for C. Under a prospective model, neither the new diabetes diagnosis nor the fracture would raise this year's payment.

3. Compute a simplified plan transfer. Two plans have equal enrolment. Plan X has an average risk score of 1.30 and Plan Y of 0.70, so the market average is (1.30 + 0.70) / 2 = 1.00. Each plan's transfer per member is its score less the market average, multiplied by an average premium of GBP 3,000. Plan X receives 0.30 × 3,000 = GBP 900 per member, and Plan Y pays the same GBP 900 per member, so the transfers sum to zero. The HHS formula adds actuarial value, geographic cost and cost-sharing terms and a 14% administrative cost reduction to the statewide average premium, so this version shows the logic only.

4. Show the power problem. Suppose Plan X's care management prevents the hip fracture that would have cost GBP 3,500. Under the concurrent model the fracture is never coded, so predicted spending for that person falls by GBP 3,000 and the plan keeps only GBP 500 of the saving, a power of 500 / 3,500 = 0.14 for that event. Under a prospective model this year's payment does not depend on this year's fracture, so the plan would keep the whole GBP 3,500 in the current year, and the fracture could at most change next year's score.

5. Interpretation. The concurrent design pays Plan X more for the costly conditions its enrollees actually develop, which protects plans that attract sick people. The same feature removes most of the reward for preventing a costly event. In real models the effect varies by condition, as Geruso and McGuire found across clinical areas.

Use in health services research and economic evaluation

Concurrent risk scores appear in studies comparing plan or provider spending, because they summarise contemporaneous morbidity well. In a cost comparison of treatments or care models, a concurrent score also captures conditions and complications arising after the intervention starts, so adjusting for it can remove part of the effect under study. The HHS-HCC models exclude categories that primarily represent complications or poor quality of care, such as pressure ulcers, for a related reason. An economic evaluation that needs baseline risk from routine data would take its risk factors from a period before the intervention, which is the prospective arrangement. A risk score from either design describes predicted spending and is no measure of a treatment effect.

Misreadings and limits

Four misreadings recur. Most read a timing label or a fit statistic out of context.

1. Treating a higher R-squared as a better payment system. Fit is one aim among several, and Geruso and McGuire show that the fit gain comes with lower power.
2. Comparing R-squared across studies without checking the data. Values move with censoring and population, as the Society of Actuaries results show.
3. Reading a risk-score rise as a sicker population. The CMS report for 2025 notes that coding practices, enrolment shifts and model changes all move scores.
4. Reading a score as a diagnosis. A diagnosis-based risk score predicts spending under one model's rules and coefficients.

Sources

• American Academy of Actuaries. Risk Assessment and Risk Adjustment. Issue brief, May 2010. https://www.actuary.org/wp-content/uploads/2017/11/Risk_Adjustment_Issue_Brief_Final_5-26-10.pdf
• Centers for Medicare & Medicaid Services, Center for Consumer Information and Insurance Oversight. Summary Report on Individual and Small Group Market Risk Adjustment Transfers for the 2025 Benefit Year. Released 30 June 2026. https://www.cms.gov/cciio/programs-and-initiatives/premium-stabilization-programs/downloads/ra-report-by2025.pdf
• Ellis RP. Risk adjustment in health care markets: concepts and applications. In: Lu M, Jonsson E, eds. Paying for Health Care: New Ideas for a Changing Society. Wiley-VCH; 2008. Author manuscript dated February 2007. https://people.bu.edu/ellisrp/EllisPapers/2007_Ellis_Riskadjustment25.pdf
• Ellis RP. Risk adjustment. In: The New Palgrave Dictionary of Economics, online edition. 2011. Author manuscript dated 19 October 2011. https://blogs.bu.edu/ellisrp/files/2012/01/Ellis_Palgrave-Dictionary-of-Economics_Risk-Adjustment_20111019.pdf
• Geruso M, McGuire TG. Tradeoffs in the design of health plan payment systems: fit, power and balance. Journal of Health Economics. 2016;47:1-19. https://doi.org/10.1016/j.jhealeco.2016.01.007
• Hileman G, Steele S. Accuracy of Claims-Based Risk Scoring Models. Society of Actuaries; October 2016. https://www.soa.org/globalassets/assets/files/research/research-2016-accuracy-claims-based-risk-scoring-models.pdf
• Kautter J, Pope GC, Ingber M, Freeman S, Patterson L, Cohen M, Keenan P. The HHS-HCC risk adjustment model for individual and small group markets under the Affordable Care Act. Medicare & Medicaid Research Review. 2014;4(3):mmrr2014-004-03-a03. https://doi.org/10.5600/mmrr.004.03.a03