Risk Adjustment

Risk Adjustment in Health Plan Payment and Provider Comparison

Risk adjustment is the set of statistical methods a payer or analyst uses to estimate how much care, spending or adverse outcome a population would be expected to generate, given who is in it, and then to pay or judge an insurer, provider or area against that expectation. In health systems its main use is to scale capitated payments, so that an insurer or commissioner covering older and sicker people receives more per head than one covering a younger, healthier group. The same logic underlies risk-adjusted quality measures, where a hospital's observed deaths are compared with the number expected for its patient mix. Because the adjusted figure moves money and reputations, the choice of risk factors changes behaviour as well as estimates.

Two questions behind every model

Ellis, in a review chapter on risk adjustment in health care markets, defines it broadly as the use of patient-level information to explain variation in spending, resource use and health outcomes over a fixed period such as a year. The American Academy of Actuaries describes it more narrowly as an actuarial tool used to calibrate payments to health plans or other stakeholders based on the relative health of the at-risk populations.

Ellis separates two questions that the literature often asks. A prediction model asks what resource use a given person is expected to have, while a payment model asks what the best capitation formula for that person is, which need not be the best predictor once incentives are considered. His Palgrave dictionary entry adds programme evaluation and outcome comparison, where the same idea is often called case-mix adjustment or severity adjustment.

Why capitated insurance markets depend on it

When premiums cannot vary with health, a plan paid the same amount for every enrolee profits on the healthy and loses on the sick. Ellis describes a sponsor that collects contributions that need not equal expected cost and pays plans on a different basis, generally subsidising high-risk and low-income consumers. Without adjustment, that payment rewards cream skimming, and the sorting of consumers across plans feeds adverse selection.

In the simplest case Ellis sets out, a perfectly informative signal marks each person as high or low risk and the sponsor pays the average cost of each signal. Profit on every type is then zero, so a plan has no reason to prefer one enrolee to another. His Palgrave entry lists Medicare and Medicaid managed care in the United States and the competitive insurance markets of Belgium, Germany, Israel, the Netherlands and Switzerland as users of risk-adjusted payment.

Building the prediction and the risk score

Most payment models regress annual spending per person on demographic cells and condition indicators, then express the prediction relative to the population mean. Each coefficient is the expected extra spending linked to one risk factor, other factors held fixed.

$$\hat{C}_i = \beta_0 + \sum_{k=1}^{K} \beta_k X_{ik}$$

where $\hat{C}_i$ is predicted annual spending for person $i$, $X_{ik}$ is the $k$th risk factor for that person (an age-sex cell, a condition category or an interaction), $K$ is the number of risk factors, and $\beta_0$ to $\beta_K$ are coefficients estimated from a calibration sample.

$$R_i = \frac{\hat{C}_i}{\bar{C}}$$

where $R_i$ is the person's risk score and $\bar{C}$ is mean predicted spending in the reference population, so 1.0 marks an average person. Payment is the score multiplied by a base rate.

Diagnoses from claims, grouped into a diagnosis-based risk score such as the hierarchical condition category system, are the most widespread addition, with pharmacy information also common.

Choosing risk factors and the prediction window

The timing choice decides whether last year's information predicts this year's spending or this year's information explains it. Ellis states that most payment systems use prospective risk adjustment, because of endogeneity concerns and because risk factors can be measured a year ahead, while concurrent risk adjustment always explains more and is also used for quality measurement. Settling payments after the period ends, retrospective risk adjustment in the settlement sense, is a separate choice.

The factor list is a policy decision as much as a statistical one. Pope and colleagues excluded vague, discretionary or not medically significant diagnoses when building the CMS-HCC model for Medicare, and dropped categories defined by procedures or equipment so that payments would follow the medical problems present rather than the services offered. Kautter and colleagues set criteria, sometimes traded off against each other, that categories in the HHS-HCC model for the Affordable Care Act markets should not be especially subject to discretionary coding and should not primarily represent poor quality or avoidable complications of care.

Measuring how well a model performs

Individual R-squared is the usual summary, and its ceiling is low. Ellis cites an estimate by Newhouse and colleagues that only about 20 to 25% of the variation in total spending is explainable with time-invariant prospective information. Kautter and colleagues report R-squared of 35 to 36% for the concurrent adult HHS-HCC models, against generally less than 2% for demographic-only models and typically 10 to 15% for prospective diagnosis-based models such as the CMS-HCC.

Group-level measures matter more for payment, because plans are paid for populations. Ellis describes predictive ratios of predicted to actual spending, which in an ideal model are close to one for every group of interest. He warns that validated R-squared in a reserved sample is often 5 to 50% lower than in the development sample.

Fit, power and coding incentives

Geruso and McGuire judge payment systems on fit, power and balance. Fit is how closely payments track costs, power is how much of a marginal saving the plan keeps, and balance is how far power differs across types of service. They found that risk adjustment weakens cost control in practice. In their simulation of Marketplace payment rules, a simple reinsurance system gave similar fit, more power and better balance than the planned scheme based on concurrent risk adjustment alone.

Coding is the second channel. Ellis describes convincing evidence that coded severity rose after Medicare introduced diagnosis-related group payment, and notes that fewer than 75% of people with a serious chronic diagnosis coded in one year have it coded again the next.

Medicare, the ACA and NHS England allocations

MedPAC's payment basics state that a Medicare Advantage plan's monthly base payment is adjusted by each enrolee's CMS-HCC risk score, calibrated on traditional Medicare data. Because these plans tend to document diagnoses more thoroughly, the Secretary must reduce their risk scores by at least 5.9%. In the Affordable Care Act individual and small group markets, the CMS report for the 2025 benefit year states that the programme transfers funds from plans with lower-than-average risk enrolees to plans with higher-than-average risk enrolees.

England applies the idea to area budgets. NHS England's technical guide to allocation formulae for 2026/27 (PRN02116) describes a weighted capitation or fair shares formula that estimates the relative need per head of each Integrated Care Board population, under the principle of equal opportunity of access for equal need. Since the 2014/15 allocations its general and acute component has used a person-based model first developed by the Nuffield Trust, in which previous diagnoses and age were the most important variables. Supply variables are set to the national average when weighted populations are calculated, so areas are not penalised for lower use caused by less capacity.

Observed and expected outcomes in provider comparison

For quality comparison, a provider's expected count of an outcome is the sum of its patients' predicted risks. NHS England defines its Summary Hospital-level Mortality Indicator (SHMI) as the ratio between the actual number of patients who die following hospitalisation at a trust and the number expected on the basis of average England figures, given the characteristics of the patients treated there. This is indirect standardisation, the logic of the standardised mortality ratio. In an illustrative case, 60 observed deaths against 75 expected gives a ratio of 0.80. The ratio cannot separate quality from unmeasured risk or chance, and the SHMI bands each trust as higher than expected, as expected or lower than expected.

Worked example with two plans and one coding change

All numbers are illustrative. Two plans of equal size share a market in which healthy enrolees cost GBP 1,000 a year and enrolees with a chronic condition cost GBP 6,000. Plan A enrols 90% healthy people and Plan B 70%, so 80% of the market is healthy.

1. Flat payment. Mean cost is 0.8 × 1,000 + 0.2 × 6,000 = 2,000 GBP. Plan A's expected cost per member is 0.9 × 1,000 + 0.1 × 6,000 = 1,500 GBP, and Plan B's is 0.7 × 1,000 + 0.3 × 6,000 = 2,500 GBP. A flat GBP 2,000 gives Plan A a profit of GBP 500 per member and Plan B a loss of GBP 500.

2. Risk-adjusted payment. Risk scores are 1,000 / 2,000 = 0.5 for a healthy enrolee and 6,000 / 2,000 = 3.0 for a chronic one. Plan A's average score is 0.9 × 0.5 + 0.1 × 3.0 = 0.75, so it receives GBP 1,500 per member. Plan B's is 0.7 × 0.5 + 0.3 × 3.0 = 1.25, so it receives GBP 2,500. Both plans break even.

3. Coding under a fixed budget. Plan B records a chronic diagnosis for a further 5% of members who are in fact healthy. Its score rises to 0.65 × 0.5 + 0.35 × 3.0 = 1.375 while costs stay at GBP 2,500. The market average score becomes (0.75 + 1.375) / 2 = 1.0625, so holding spending at GBP 2,000 per member cuts the base rate to 2,000 / 1.0625 = 1,882.35 GBP. Plan A receives 0.75 × 1,882.35 = 1,411.76 GBP and Plan B 1.375 × 1,882.35 = 2,588.24 GBP.

4. Interpretation. Plan B gains GBP 88.24 per member from coding alone and Plan A loses the same amount. Without a fixed budget, as in Medicare Advantage, the extra coding would raise total payment instead, which is the gap the Medicare Advantage coding adjustment narrows.

Use in economic evaluation and health services research

Risk scores often serve as covariates when observational studies compare costs between treatments or care models using claims data. A score built from diagnoses recorded after treatment starts can absorb part of the treatment's own effect, a concern that parallels the HHS-HCC criterion on avoidable complications. Causal comparisons need confounding control with baseline covariates fixed before the intervention, the purpose of methods such as the propensity score.

Boundaries with related ideas

Risk pooling spreads financial risk across members of a fund, while risk sharing pays part of the costs a plan actually incurs. Ellis notes that fees, risk sharing and other non-capitated payments are often combined with risk-adjusted capitation, and he suggests economists keep risk adjustment and case-mix adjustment apart for clarity.

Common misreadings

Several misreadings recur. Each confuses a statistical summary with the economic question behind it.

1. Treating a high R-squared as proof of a good payment system. Geruso and McGuire show that better fit can cost power.
2. Reading a rise in average risk scores as a sicker population. Coding practice can move scores.
3. Reading an observed to expected ratio below one as proof of better quality. Unmeasured risk and chance can produce the same result.
4. Assuming every predictive variable belongs in the model. Variables reflecting supply or discretionary coding reward behaviour a payer wants to discourage.

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
• Kautter J, Pope GC, Ingber M, et al. 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/mmrr2014-004-03-a03
• Medicare Payment Advisory Commission. Medicare Advantage program payment system. Payment Basics, revised October 2026. https://www.medpac.gov/wp-content/uploads/2026/10/MedPAC_Payment_Basics_26_MA_FINAL_SEC.pdf
• NHS England. Summary Hospital-level Mortality Indicator (SHMI), Deaths associated with hospitalisation, England, June 2025 to May 2026. Published 8 October 2026. https://www.gov.uk/government/statistics/summary-hospital-level-mortality-indicator-shmi-deaths-associated-with-hospitalisation-england-june-2025-may-2026
• NHS England. Technical guide to allocation formulae and convergence for 2026/27 allocations. Publication reference PRN02116, December 2025. https://www.england.nhs.uk/wp-content/uploads/2025/12/allocations-2627-technical-guide-to-formulae-v2.pdf
• Pope GC, Kautter J, Ellis RP, et al. Risk adjustment of Medicare capitation payments using the CMS-HCC model. Health Care Financing Review. 2004;25(4):119-141. https://pmc.ncbi.nlm.nih.gov/articles/PMC4194896/