Prospective Risk Adjustment

Prospective Risk Adjustment in Health Plan Payment

Under prospective risk adjustment, the payer works out each enrolled person's expected cost for the year ahead from information recorded before that year began, usually age, sex and the diagnoses coded in the previous year. A plan whose members were sicker last year is paid more this year, before any of this year's costs are known. Health economists treat it as the main tool for paying competing insurers under capitation without rewarding the enrolment of healthy people. Medicare Advantage in the United States has adjusted payments to private plans with the prospective CMS-HCC model since 2004, and the Dutch Central Insurance Fund once used it to allocate money among sickness funds. The defining choice is timing, and that choice trades predictive power against resistance to manipulation.

Base year and payment year

Every prospective model separates two windows, a base period in which information is collected and a later period whose spending is predicted and paid for. Ellis describes the usual framework as information from a base period, usually one year, used to predict spending in the subsequent period. Pope and colleagues, describing the CMS hierarchical condition category (CMS-HCC) model that Medicare introduced in 2004, calibrated it on diagnoses from 1999 and spending in 2000.

$$\widehat{C}_{i,t} = \sum_{k=1}^{K} \beta_k X_{ik,t-1}$$

where $\widehat{C}_{i,t}$ is predicted spending for person $i$ in payment year $t$, $X_{ik,t-1}$ is risk factor $k$ (an age and sex cell, a recorded condition or another adjuster) observed for that person in base year $t-1$, $\beta_k$ is the estimated cost weight for that factor and $K$ is the number of risk factors.

Model timing and payment timing are different things. Ellis notes that Medicare calculates payments prospectively each January from the previous calendar year's data and then makes retrospective adjustments for late-arriving data. Pope and colleagues explained that, under the 2004 design, provisional rates were set by 1 January on data lagged six months, with final rates by 30 June of the payment year. That reconciliation completes the base-year records and does not make the model concurrent.

Why payers prefer information from the year before

The case for prior-year information comes from the selection problem in insurance markets. Ellis, writing in the New Palgrave Dictionary of Economics, describes risk-adjusted payment as a way of reducing plans' incentives to seek favourable selection and of avoiding the market failures caused by adverse selection. A plan paid a flat amount per member profits from enrolling people who will cost less than that amount, which is the incentive behind cream skimming.

Ellis gives two reasons why most payment systems use the prospective form. The signals are less endogenous, because a plan cannot change last year's recorded conditions by treating people differently this year, and the risk factors can be measured a year before the spending they predict. The cost of this choice is lower explanatory power, since Ellis states that concurrent models always explain more of the variation in spending than prospective ones.

The alternative is concurrent risk adjustment, which uses diagnoses from the same period as the spending. Kautter and colleagues explain that the risk adjustment model under the Affordable Care Act for the individual and small group markets was built as a concurrent model. Prior-year diagnoses would not exist when the scheme started in 2014, and people move in and out of enrolment in those markets. Prospective adjustment suits schemes with stable membership.

Turning recorded conditions into a payment

A diagnosis-based prospective model has two layers of risk factors. In the CMS-HCC model, demographic cells and Medicaid status form the first layer, and diagnoses grouped into hierarchical condition categories form the second, so the most severe manifestation of a disease process counts and unrelated conditions add together.

A published vignette by Pope and colleagues shows how the increments combine. A hypothetical woman aged 79, eligible for Medicaid, with acute myocardial infarction, angina, chronic obstructive pulmonary disease, renal failure, chest pain and an ankle sprain receives USD 2,562 for her age and sex cell, USD 616 for Medicaid, USD 1,885 for the infarction, USD 1,936 for the lung disease and USD 2,908 for renal failure. Angina adds nothing because the infarction sits above it in the coronary hierarchy, and the chest pain and sprain are excluded. Her predicted cost is the sum, USD 9,907.

$$R_{i,t} = \frac{\widehat{C}_{i,t}}{\overline{C}_{t}}, \qquad P_{j,t} = B_t \times \frac{1}{N_j} \sum_{i \in j} R_{i,t}$$

where $R_{i,t}$ is the relative risk score of person $i$, $\overline{C}_{t}$ is mean predicted spending in the reference population, $P_{j,t}$ is the simplified payment per member to plan $j$, $B_t$ is the base rate per member for year $t$ and $N_j$ is the number of members of plan $j$.

Each year CMS divides scores by a normalisation factor that keeps the average fee-for-service score at 1.0, and applies a separate coding pattern adjustment.

Worked example with one plan and three risk groups

All numbers in this example are illustrative and the payment rule is simplified. A payer uses a prospective model with base-year cost weights of GBP 3,000 for people aged 65 to 74, GBP 4,500 for people aged 75 and over, GBP 2,000 for recorded diabetes and GBP 3,500 for recorded heart failure. Mean predicted spending in the reference population is GBP 5,000, which is also the base rate.

1. Predict spending for each group. Group A, aged 70 with no recorded conditions, is predicted at GBP 3,000. Group B, aged 78 with diabetes, is predicted at 4,500 + 2,000 = GBP 6,500. Group C, aged 72 with diabetes and heart failure, is predicted at 3,000 + 2,000 + 3,500 = GBP 8,500.

2. Convert to risk scores. Dividing by GBP 5,000 gives scores of 3,000 / 5,000 = 0.60 for A, 6,500 / 5,000 = 1.30 for B and 8,500 / 5,000 = 1.70 for C.

3. Average across the plan. The plan has 600 members in group A, 300 in B and 100 in C. Its mean score is (600 × 0.60 + 300 × 1.30 + 100 × 1.70) / 1,000 = (360 + 390 + 170) / 1,000 = 0.92.

4. Set the payment. The plan receives 0.92 × 5,000 = GBP 4,600 per member, or GBP 4,600,000 in total, against GBP 5,000,000 under a flat rate. The adjustment reflects a membership that was healthier than average in the base year.

5. Apply normalisation. Suppose recorded morbidity across the whole market has drifted so that the average score is 1.04 although underlying need is unchanged. Dividing by that factor gives 0.92 / 1.04 = 0.8846, a payment of about GBP 4,423 per member and about GBP 4,423,077 in total.

6. Follow a new diagnosis. If a group A member develops heart failure during the payment year, the plan still receives 0.60 × 5,000 = GBP 3,000 for that person this year, whatever the treatment costs. Next year the prediction becomes 3,000 + 3,500 = GBP 6,500, a score of 1.30, provided the diagnosis is recorded. A concurrent model would instead have recognised the condition in the year it arose.

7. Follow a coding change. If 50 group A members have diabetes that was never recorded, recording it adds 50 × 2,000 = GBP 100,000 to the plan's predicted spending before normalisation. Whether this is a correction or upcoding depends on whether the condition is real and documented.

Under this method the plan carries the cost of new illness within the year, while the payer carries the risk that recorded morbidity differs from true morbidity.

How well prospective models predict

Prospective models explain a modest share of individual spending. Pope and colleagues report an R-squared of 9.8% for the 70-category CMS-HCC model. Ellis cites work by Newhouse and colleagues suggesting that only about 20 to 25% of variation in total spending is explainable from time-invariant prospective information, and Ellis puts the upper bound nearer 30% in more recent data. He also reports that validated R-squared values are often 5 to 50% lower than in the development sample.

Group-level accuracy matters more for payment. Pope and colleagues measure it with predictive ratios, the mean predicted cost of a group divided by its mean actual cost. Their earlier full diagnostic model gave a ratio of 1.23 for the fifth with the lowest base-year spending and 0.69 for the top 1%, so the most expensive people remained underpredicted.

Incentives the method creates

Every risk factor in a payment formula becomes something a plan can influence. Ellis notes that policy makers have mainly worried about plans recording more diagnoses once payment depends on them. Some commenters on the CMS 2027 announcement argued that the statutory minimum coding adjustment understates coding differences between Medicare Advantage and fee-for-service.

Unbiased average payment is not the same as removing selection incentives. Ellis summarises work by Glazer and McGuire showing that, because risk signals are imperfect, plans can still distort service offerings to attract the healthy within each payment category, and that deliberately overpaying signals of high cost can offset this. Van Kleef and colleagues show for the Netherlands that constraining regression coefficients can reduce under- or overcompensation for chosen groups at the cost of increasing it for others.

Predicting spending is also not the same as predicting need. Obermeyer and colleagues found that a widely used care-management algorithm, which predicted health care costs, assigned Black patients the same scores as White patients who were less sick, because unequal access to care meant less was spent on Black patients. That algorithm guided care management rather than payment. Any prospective model calibrated on past spending still inherits past patterns of access.

Prospective, concurrent and retrospective adjustment

The terms overlap in the literature. Prospective adjustment uses base-period information for a later period, while concurrent adjustment uses information from the period whose spending is predicted. Ellis notes that concurrent adjustment is sometimes called retrospective. The retrospective risk adjustment page covers adjustments calculated after a coverage period ends. All three are forms of risk adjustment, and Ellis suggests keeping risk adjustment for spending and payment distinct from case-mix adjustment of outcomes for an episode of care. The diagnosis-based risk score page covers how coded conditions become a score.

Common misreadings of prospective scores

Most errors come from treating a payment score as something it was not built to be. Each leads to a wrong reading of plan payments.

1. Reading a score as one person's bill. With individual R-squared near 10%, actual spending varies widely around the score.
2. Reading a higher score as a sicker population. It may reflect fuller coding.
3. Treating good average fit as the end of selection. High-cost groups and groups without an explicit adjuster need their own predictive ratios.
4. Treating a cost-calibrated model as a measure of need. It reproduces whatever base-year spending reflected, including unequal access.

Sources

• Centers for Medicare and Medicaid Services. Announcement of Calendar Year (CY) 2027 Medicare Advantage (MA) Capitation Rates and Part C and Part D Payment Policies. 6 April 2026. https://www.cms.gov/files/document/2027-announcement.pdf
• Ellis RP. Risk adjustment in health care markets: concepts and applications. In: Lu M, Jonsson E, eds. Financing Health Care: New Ideas for a Changing Society. Wiley-VCH; 2008. p. 177-222. https://doi.org/10.1002/9783527611294.ch8 (author draft, 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. Palgrave Macmillan; 2011. Author manuscript, 19 October 2011. https://blogs.bu.edu/ellisrp/files/2012/01/Ellis_Palgrave-Dictionary-of-Economics_Risk-Adjustment_20111019.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 and Medicaid Research Review. 2014;4(3):mmrr2014-004-03-a03. https://doi.org/10.5600/mmrr.004.03.a03
• Obermeyer Z, Powers B, Vogeli C, Mullainathan S. Dissecting racial bias in an algorithm used to manage the health of populations. Science. 2019;366(6464):447-453. https://doi.org/10.1126/science.aax2342
• Pope GC, Kautter J, Ellis RP, Ash AS, Ayanian JZ, Iezzoni LI, Ingber MJ, Levy JM, Robst J. 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/
• van Kleef RC, McGuire TG, van Vliet RCJA, van de Ven WPPM. Improving risk equalization with constrained regression. European Journal of Health Economics. 2017;18(9):1137-1156. https://doi.org/10.1007/s10198-016-0859-1