Insurance by BogdanovaThe Encyclopedia of Insurance and Risk Science

RESEARCH DOSSIER

Health Risk Adjustment Models and Calibration

Abstract. Health risk adjustment transforms measured differences in expected insured health expenditures into payment differentials among health plans. It is a statistical procedure embedded in a regulated market, not simply an exercise in predicting a person's medical spending. This monograph investigates the construction of diagnosis- and utilization-sensitive risk scores, estimation under skewed and zero-inflated expenditure distributions, calibration across populations and years, risk-equalization transfer design, coding incentives, residual selection, and model governance. It combines formal analysis with primary statutory and administrative records from the United States, the Netherlands, Germany, and Switzerland and with published economic and actuarial research. The central finding is that accurate average payment calibration, equitable compensation for predictable risk, and resistance to endogenous coding and plan selection are separate empirical properties. Improvements in a conventional fit statistic do not establish that transfers eliminate profitable selection; a highly predictive variable can also intensify incentives to alter recorded diagnoses or utilization. An assessment of a risk adjustment regime must therefore connect its estimand, eligible population, time horizon, validation sample, risk-score normalization, transfer formula, coding rules, and residual financial exposure. Contemporary implementations demonstrate persistent institutional choices concerning prospective versus concurrent information, common benefit baskets, high-cost risk sharing, and regulatory auditing. The analysis supplies derivations, source-specific comparisons, a reproducible validation protocol, and bounded substantive conclusions without introducing a new theory or asserting unperformed statistical estimation.

Keywords: Health risk adjustment; risk equalization; prospective payment; hierarchical condition categories; morbidity adjustment; health insurance; calibration; predictive ratio; risk selection; diagnosis coding; Medicare Advantage; ACA risk adjustment; risk sharing; out-of-sample validation; residual risk; international comparative insurance research.

Introduction

Health insurers differ in their average expected claims costs even when their administrative efficiency, contractual service obligations, and provider purchasing practices are identical. A market that permits people to choose among insurers can reward insurers for attracting members whose expected medical spending is low relative to the premiums and regulated transfers generated by those members. Risk adjustment modifies financial flows in response to observable characteristics correlated with the insured population's expected expenditure, so that legitimate competition over service, quality, and cost management becomes less entangled with the composition of the risk pool.

This scientific object is rooted in selection theory. Einav and Finkelstein's insurance-selection synthesis explains why adverse and advantageous selection cannot be inferred from a single cross-sectional correlation between observed cost and insurance take-up. Geruso and Layton's health-market policy review establishes that selection can affect both whether insurance is purchased and which plan is chosen. Consequently, a well-specified health risk adjustment model must be evaluated together with market structure and contract choice.

Three questions organize this monograph. First, how are the statistically relevant insured population, diagnosis and spending variables, score weights, and forecast horizons specified? Second, what do calibration and external validation reveal about systematic under- and overcompensation, and what can changes in diagnostic recording conceal? Third, how do prospective payments and risk equalization leave residual risk, and how do the remaining incentives differ across institutions? The study distinguishes exact mathematical implications, propositions conditional on explicitly stated assumptions, causal estimates reported by other researchers, and administrative decisions enacted for particular benefit years.

The international comparison is deliberately selective. United States ACA risk adjustment, Medicare Advantage, Dutch statutory basic insurance, German statutory sickness funds, and Swiss mandatory health insurance operate with different benefit and premium arrangements; they illustrate institutional mechanisms rather than a common global estimator. The comparative units are the payment rule, insured person, insurer or health plan, observation year, and regulatory version. No patient-level proprietary data were acquired or analyzed here. Numerical examples below are illustrative calculations, identified as such.

The economic and institutional object of risk adjustment

Risk adjustment is a mapping from eligible enrollee attributes and recorded health information into financial compensation for expected covered expenditures. For insured person i, define Yᵢ,t as covered expenditure or plan liability over evaluation period t under the applicable benefit design; Xᵢ,t−1 as information available before that period; and Rᵢ,t = f(Xᵢ,t−1; θ) as the predicted risk score under parameters θ. The expenditure target can be allowed claims, insurer liability after member cost sharing, or another administratively defined amount. These targets imply different interpretations and cannot be interchanged.

The essential financial question is which portion of expected variation should be redistributed through a common risk pool. Variation caused by unavoidable morbidity, age, disability, or related clinical factors can create selection incentives. Variation caused by discretionary intensity, contracted prices, provider supply, plan design, or inefficient management can carry legitimate incentive effects. In any given scheme the empirical separation is imperfect. A risk variable that predicts expenditure because providers can induce it is statistically useful and potentially problematic for incentive design at the same time.

A decomposition clarifies the object. Write Yᵢ = μ(Xᵢ) + εᵢ, where μ(Xᵢ) is conditional expected covered expenditure under a specified data-generating and contracting environment and E[εᵢ | Xᵢ] = 0 in that environment. The risk score attempts to approximate μ, often after dividing by a reference mean. If insurers sort on additional information Z unobserved by the model, then E[εᵢ | plan j] can differ from zero even though residuals average to zero conditional on model covariates in the estimation population. This is the analytical origin of residual selection under imperfect adjustment.

Historically, Newhouse, Manning, Keeler, and Sloss used the RAND-related expenditure analysis to compare health status and prior use as capitation adjusters. Their specific results refer to the experimental sample and to an explained share of predictable variance; they are not contemporary estimates of any national risk adjustment system. The study shows why previous utilization can improve prediction while changing the financial incentives of capitated payment.

A basic distinction is between risk classification and the allocation mechanism. Model prediction assigns a score or expected cost. Transfer rules determine actual revenue and incorporate premiums, actuarial value, geographic factors, enrollment, budget constraints, and retrospective settlement. Insurer net exposure is the revenue remaining after all these steps minus actual benefit expenditures. Judging an insurer's incentive from a risk score without the associated transfer rule is therefore incomplete.

Historical evolution and source hierarchy

Early adjusted capitation relied heavily on demographic attributes and administrative eligibility categories. The later use of diagnosis groups and pharmaceutical histories increased the available information on morbidity, allowing plans and regulators to differentiate insured persons with otherwise similar demographics. The 1989 capitation study is an early empirical illustration of this progression. The Dutch article by van de Ven, van Vliet, and Lamers documented the use of diagnostic and pharmacy cost groups in Dutch premium subsidies in 2004.

Regulatory systems subsequently attached richer classification structures, hierarchy rules, age interactions, duration terms, and separate risk-pool components. These changes reflect program-specific evidence and political choices. A model revision is not automatically a demonstration of improved causal fairness; it may address coding practices, benefit changes, predictability, or fiscal distribution.

The documentary hierarchy in this study begins with binding statutory texts and effective administrative rules, followed by official methodology and payment-year records, then peer-reviewed analyses and official research evaluations. In the United States, the Patient Protection and Affordable Care Act, section 1343, establishes the underlying statutory risk-adjustment mechanism for relevant health-plan markets. Later federal payment notices specify year-specific model and transfer details. A rule announced in May 2026 about benefit year 2027 must be described as a 2027-model rule, not as observed 2027 transfers.

In Germany, the Federal Office for Social Security records the evolution of morbidity-based equalization and separately publishes classification and data procedures. Dutch annual instructions distinguish ex ante allocations from subsequent settlement. Swiss authorities specify their own risk characteristics and adjustment agency. Comparability is achieved through explicit variables and institutional mechanisms, rather than by assuming identical legal instruments.

The methodological distinction between explanation and prediction is also fundamental. Shmueli's explanation-versus-prediction analysis explains that in-sample explanatory fit is not equivalent to future predictive performance. Health risk adjustment makes this distinction operational: historical coding and clinical practice generate the training associations, while future plan payments react to the implemented prediction rule.

Risk-score architecture and statistical estimation

Defining the target and observation window

A defensible estimation begins with a clearly defined target. Let Lᵢ,t be plan liability for covered services during benefit year t, adjusted to common benefits where the program requires such a basis. A prospective estimator uses variables measured before year t, whereas a concurrent estimator uses some events recorded within the payment year. A prospective model more directly limits payment dependence on current utilization, but may predict less accurately for newly emerging illness; a concurrent model can reflect current morbidity and simultaneously reward intensity of recording.

Enrollment duration creates another choice. If a member is enrolled for m months and annual spending is modeled as if exposure were twelve months, naïvely annualizing short exposures can distort estimates for people with clustered care episodes. A regulator may instead specify duration groups, exposure-weighted outcomes, or partial-year factors. Death, newborn status, switching plans, and coverage transitions require explicit data and timing rules.

Diagnosis-based scores require a reproducible classification pipeline: eligible sources of diagnosis, standardized coding version, mapping from codes to condition categories, hierarchy exclusions, interactions, and eligibility restrictions. Demographic factors may enter separately, and pharmaceutical cost groups can serve as indicators of treatment-associated morbidity. A diagnosis code is an observation of a clinical and documentation process, not an invariant biological trait.

A general scoring representation

A simplified score is Rᵢ = α + ∑ₖ βₖ Dᵢₖ + ∑ₗ γₗ Aᵢₗ + ∑ₘ δₘ Iᵢₘ, where D are mutually reconciled condition-category indicators, A are demographic or exposure terms, and I are specified interactions. This is a pedagogical representation of common additive risk-classification approaches, not a transcription of a particular regulator's full operational algorithm. Programs can use infant severity cells, distinct segments, pharmaceutical factors, and policy adjustments that require different structures.

A normalization such as rᵢ = predicted liabilityᵢ / mean predicted liability in the reference population yields a relative score near one on average when the reference and calculation definitions align. If the average enrollee has score one, the interpretation of a score of 1.4 is 40 percent higher predicted liability relative to that benchmark, not a 40 percent probability of illness. Risk scores are dimensionless multipliers under this normalization; financial transfers are monetary flows and depend on further rules.

Regression on highly skewed, heavy-tailed expenditures can be unstable if extreme expenses dominate the objective. Ordinary least squares is simple and may align with an average-dollar prediction target. Alternative generalized linear models, two-part occurrence-and-severity models, winsorization, or high-cost pooling alter how the upper tail contributes to estimation. Such choices should be assessed using expected payment losses, individual and subgroup predictive ratios, and out-of-period performance. Altering the target before estimation can change what the model is designed to compensate.

Variable selection and hierarchy

A risk category should be tested for clinical coherence, predictive value, robustness to code transitions, and manipulation incentives. Hierarchical categories prevent simultaneously paying multiple severity weights for mutually nested conditions when the design specifies this. Interactions address nonadditivity but also increase variance and may respond sharply to small coding differences. Sparsely represented diseases make subgroup estimates noisy even when they are clinically important. Pooling or partial regularization can stabilize estimates at the cost of a structural assumption about shared effects.

The estimation population is a substantive choice. Training on fee-for-service beneficiaries while paying managed care plans, or training on one set of benefit designs while compensating another, creates possible covariate and outcome shifts. The original U.S. CMS-HCC model evaluation is an institutional example of assessing models across identified beneficiary subsets. Modern transfer models must also respect whether premiums and costs are defined on the same geographical and benefit basis.

Estimation uncertainty

For a model vector θ̂, the point predictions f(X; θ̂) inherit sampling variation, specification uncertainty, coding changes, and macroeconomic trend error. The last two are frequently more important for deployment than a small standard error calculated as if administrative records were independent and stationary. Clustered uncertainties at plan, provider, and region levels should be considered; serial dependence across years requires a longitudinal validation design. Report confidence intervals for group predicted-to-actual ratios where underlying observations are available, and separate parameter uncertainty from policy scenario uncertainty.

Calibration, discrimination, and validation

Calibration asks whether predicted expenditure is aligned with actual expenditure for a specified population and period. Define for group g the predictive ratio PR_g = (∑ᵢ∈g Ŷᵢ)/(∑ᵢ∈g Yᵢ), where exposures and prices are harmonized and the denominator is positive. PR above one means aggregate overprediction and below one means underprediction. This ratio is directly relevant to potential uncompensated groups. A ratio near one for the whole market can coexist with large errors for high-cost patients, people with disabilities, mental-health diagnoses, or members with a brief enrollment duration.

An alternative observed-to-expected ratio O/E = (∑Y)/(∑Ŷ) reverses the direction. Reporting which convention is used prevents erroneous sign interpretation. Group calibration is meaningful only after defining group membership independently of outcome leakage and accounting for any reweighting. For example, suppose a hypothetical group of 1,000 enrollees has $4.8 million predicted and $6 million observed covered expenditure. PR = 0.80: the model underpredicts that group by 20 percent of observed expenditure. If that group is concentrated in one insurer and no compensating transfer term exists, the model creates a potential revenue disadvantage; the example itself does not establish actual market selection.

Discrimination describes the ability to distinguish higher from lower expected cost. A model may rank people accurately but consistently underpredict a subgroup. Conversely, a model with modest individual R² may yield well-calibrated annual average payments to large insurer groups. Thus calibration intercept, calibration slope, overall R² or squared error, high-cost-tail metrics, and group predictive ratios measure different properties. For a continuous nonnegative spending target, a classification AUC describes only a defined binary outcome and cannot substitute for expenditure calibration.

Out-of-sample validation means that the test observations were not used to choose variables, tune complexity, or estimate coefficients. A random person-level split can still leak temporal, family, provider, and institutional dependencies. The stronger validation for next-year payment is chronological: train on earlier benefit years, freeze the specification, and score a later year using the information permitted at payment time. Compare actual and predicted plan liability at both the enrollee and plan level. A second transport test should apply the frozen rule to different insurer mixes or territories when lawful data permit.

Recalibration may update an intercept to correct an average level, a slope to correct dispersion, or all category coefficients to reflect changed costs and coding. These operations have distinct economic effects. A uniform multiplicative correction may improve aggregate dollars while leaving residual group losses unchanged in relative terms. Re-estimating category weights can improve group alignment but also increase the marginal payment associated with a recorded diagnosis. There is no model-independent theorem that more frequent recalibration increases welfare.

For U.S. benefit year 2027, the finalized federal model recalibration uses separately estimated coefficients on 2021, 2022, and 2023 EDGE enrollee-level data and then averages them; the document also preserves the basic state transfer formula. This is an administrative calibration specification. It is not a publicly observed prospective test of 2027 outcomes, because payment-year outcomes occur later. The rule explicitly discusses historical pandemic-related composition concerns, showing that input-year choice is itself a model-risk decision.

A complete validation record should document data cutoffs; insurer eligibility; exclusions; coding versions; the covered benefit basket; geographic cost adjustment; target price level; exposure adjustment; uncertainty intervals; model version; and any retrospective reconciliation. Without this ledger, the term "calibrated" has insufficient empirical content.

From individual predictions to insurer transfers

Let plan j enroll n_j members with risk scores rᵢ. Its exposure-weighted mean score is R̄_j = (∑ᵢ wᵢ rᵢ)/(∑ᵢ wᵢ), with wᵢ representing eligible person-time or other rule-specific weights. The difference between this score and a reference average can feed a transfer formula, but transfers need additional terms to avoid compensating plans twice for factors already recognized in permitted premiums or benefits.

For a simplified budget-neutral illustration, assume each plan has the same regulated benefit and rating terms, risk scores are measured consistently, and there are N total covered exposures. Let B be a common monetary base per exposure, and R̄ = (∑_j n_j R̄_j)/N. Then define T_j = B n_j(R̄_j − R̄). Summing over plans gives ∑_j T_j = 0 exactly. High-scoring plans receive, and low-scoring plans contribute. This identity explains one transfer mechanism; it is expressly not the full HHS state transfer formula.

Consider two hypothetical plans: A enrolls 10,000 exposure-equivalent members with mean score 1.20 and B enrolls 10,000 with score 0.80. The common mean equals 1.00. At B = $6,000, T_A = $12 million and T_B = −$12 million. A model change that raises A's score to 1.22 while leaving B at 0.80 raises the market mean to 1.01, so T_A = $12.6 million and T_B = −$12.6 million. This demonstrates that a coding change can shift real revenue even when there is no newly observed health expenditure. It does not imply that either payment is correct without independently examining underlying health risk and applicable rule details.

In an actual regulator's formula, plan actuarial value, induced demand, geographic cost, eligible premiums, rating factors, and risk-pool segmentation can change the transfer. In a federal premium-stabilization system, state and market-pool boundaries also matter. The ACA legal provision establishes relative low- versus high-actuarial-risk charges and payments; CMS rules determine the operational methodology. The premium-stabilization program inventory links benefit-year transfer and validation documents, making the distinction between a model-year forecast and an observed transfer report explicit.

Residual risk arises because actual cost Y_j deviates from expected funded cost, because expected cost conditional on unmodeled traits can vary by plan, or because the transfer rule covers only part of modeled morbidity differences. Random claims volatility calls for capital or risk sharing; persistent predictable losses call for a different diagnostic response. High-cost pooling or risk corridors can protect insurers against tail exposure, but may weaken incentives to manage utilization depending on the marginal sharing fraction.

United States: ACA individual and small-group markets

The ACA permanent risk adjustment program addresses enrollment selection in eligible individual, small-group, and merged market structures. Section 1343 establishes the statutory payment and charge principle. The operational approach uses HHS hierarchical condition category models and state-market transfer formulas rather than a single flat per-enrollee capitation score. The 2026 Payment Notice final rule specifies recalibration using 2020–2022 EDGE data, a phased change in treatment of hepatitis C drug costs, and a PrEP-associated factor for the adult and child models.

The 2027 final Payment Notice updates the calibration window to 2021–2023 data. It distinguishes adult, child, and infant model structures and identifies factors such as age-sex categories, HCCs, prescription drug categories and interactions as applicable to each model. The adult and child models include additional affiliated cost factors. The 2027 rule also maintains the high-cost risk pool threshold of $1 million and a 60 percent coinsurance rate above that threshold. These are specific regulatory parameters for the specified benefit year and are not universal actuarial recommendations.

The EDGE system connects insurer-submitted enrollee-level data with a federally prescribed methodology. Data quality is therefore a necessary condition for both statistical prediction and legitimacy of transfers. An error in eligibility, diagnosis sourcing, benefits, or enrollment days can propagate through a risk score, through a plan-level average, and into a charge or payment. CMS maintains a risk adjustment guidance inventory with yearly algorithm instructions and technical files; a replication should pin one version rather than combine coefficients and eligible-code definitions from different years.

Regulatory auditing matters here because a classification algorithm rewards recorded evidence of morbidity. A sufficiently high score does not prove that all underlying diagnoses are valid or that the associated spending is unavoidable. Data validation rules operate as a separate control. The statutory, algorithmic, and audit layers should be independently documented in any evaluation.

A particular risk in interpreting annual transfer reports is to mistake an aggregate charge for an insurer penalty or an aggregate payment for profit. Transfers are redistributive functions of plan composition, premium-related terms, and applicable adjustments. Assessing insurer financial adequacy requires studying premiums, claims, administrative costs, reserves, and retrospective settlements in addition to transfer totals. The CMS transfer documentation supplies program-specific reporting, not a causal comparison of competing plans' care quality.

United States: Medicare Advantage as a distinct payment system

Medicare Advantage is analytically related to, but institutionally different from, ACA individual-market risk adjustment. Federal Medicare payments to private MA organizations use the CMS-HCC framework with Medicare-specific rate setting, benchmarks, eligibility segments, coding adjustment, and normalization. Directly transplanting ACA model coefficients or transfer interpretations into MA creates category errors.

In its 2026 MA rate announcement, CMS completed the three-year phase-in of the 2024 CMS-HCC risk adjustment model for non-PACE organizations, using that updated model for 100 percent of their 2026 risk-score calculation. The announcement also discusses fee-for-service risk-score normalization and a distinct Part D risk adjustment update. Those choices alter the calibration target and institutional baseline, and must be read as payment-year implementation rules.

A major empirical concern is diagnostic intensity. Geruso and Layton's Medicare upcoding study found private-plan diagnosis-based scores 6 to 16 percent above counterfactual fee-for-service scores in their analyzed comparisons, with stronger coding responses in vertically integrated organizations. Their estimates are study-specific, using the authors' methods and period, and are evidence about recording incentives rather than a universal fraud rate.

The HHS Office of Inspector General's 2024 HRA evaluation found that diagnoses appearing only in health risk assessments and linked chart reviews, rather than other 2022 service records, were associated with an estimated $7.5 billion in 2023 MA risk-adjusted payments. The report raises alternative concerns about unsupported diagnoses and possible unmet medical needs. The finding does not establish that every isolated diagnosis was false, and the estimated payments should not be reclassified wholesale as proven fraud.

This example shows why documentation integrity and health status are analytically distinct. A score can be mathematically accurate given the submitted codes and financially misleading if coding propensity varies by organization. Conversely, improved detection of real previously unrecorded illness can raise scores for a clinically justified reason. Audit designs must therefore examine chart support, subsequent care, provider behavior, encounter sources, and the counterfactual cost trajectory rather than inferring manipulation solely from a score increase.

The earlier CMS-HCC evaluation program illustrates regulator-directed model assessment for beneficiary subgroups. Future work needs to distinguish error in risk prediction, error in diagnoses, and the difference between fee-for-service and managed-care observation processes.

Netherlands: ex ante equalization and retrospective settlement

Dutch statutory basic health insurance combines acceptance obligations and regulated basic benefits with risk equalization to compensate insurers for unequal expected enrollee costs. The Zorginstituut explanation describes use of age, sex, region, medicine, hospital information, and related risk classes. Government-defined normative amounts attached to risk categories determine equalization contributions, making classification and annual coefficients central to insurer revenue.

The 2026 Dutch program materials identify the annual ministerial regulation, institute policy rules, the ex ante model advice, and updated reference files. The institutional sequence is significant: a forecast allocation is required before the coverage year for price setting, but actual enrollment and claims information subsequently becomes available. The institute describes a multiyear settlement cycle with preliminary contributions and eventual final determination.

This arrangement distinguishes ex ante payment allocation from ex post correction. A rule incorporating finalized enrollee composition after the year does not thereby become a concurrent patient-cost model in the estimation sense; the dimensions are different. Temporal information used to predict morbidity and timing of account reconciliation should be identified separately. An insurer can receive an ex ante forecast, an adjustment reflecting actual eligible lives, and a later settlement under the same legal framework.

van de Ven, van Vliet, and Lamers's Dutch health-adjusted subsidy study documents complementary diagnostic and pharmacy cost groups in 2004. It is historical evidence of the role of clinical and pharmaceutical adjustment, not direct validation of 2026 coefficients. The current primary institutional sources must take precedence for contemporary features.

The Dutch comparison also highlights the significance of guaranteed issue and uniform basic coverage. When insurers must accept applicants and cannot perfectly price individual morbidity, inadequate equalization produces incentives for indirect selection through plan features, marketing, service responsiveness, or provider arrangements. A model's predictive ratio for chronic-condition groups is therefore an economic warning indicator even if the insurer never rejects an applicant.

Germany: morbidity-based risk structure adjustment

The German statutory health insurance system applies risk structure adjustment among sickness funds through the health fund. The Federal Office for Social Security's official system overview describes a solidarity-based environment in which funds cannot freely choose whom to insure and insured populations differ in their demographic and morbidity composition.

The Office's explanation of Morbi-RSA states that morbidity is recorded through billing data, diagnoses, and prescribed medications. It characterizes the classification model as prospective and describes annually fixed specifications. That description supplies an essential distinction from a purely retrospective reimbursement of current individual costs: forecastable morbidity affects allocations while sickness funds retain financial responsibility for spending decisions.

Reforms effective from the 2021 equalization year widened the classification to a full disease spectrum, introduced regional components, and established a high-cost risk pool. The Federal Office's reform account describes 80 percent reimbursement of eligible expenses above an annually adjusted threshold within that pool. The 80 percent pertains to the German regime's specified risk-pool mechanism, not to every medical expense and not to the U.S. program's 60 percent excess-cost coinsurance.

The Office also publishes data audit requirements for morbidity reporting and enrollee time, and its risk-group data repository distinguishes information by settlement year and decision stage. These records demonstrate that risk adjustment is a sustained administrative measurement system. They do not supply a straightforward internationally comparable estimate of insurer-level selection without harmonized outcomes and adjustment denominators.

The analytical implication is that additional disease categories can shrink some predictable losses, while coding, classification instability, regional prices, and expenses above risk-pool thresholds still shape incentives. The magnitude of the effect remains an empirical question for a defined year and coverage population. The presence of sophisticated morbidity adjustment is insufficient by itself to establish the absence of selection.

Switzerland: pharmaceutical cost groups and insurer equalization

Swiss mandatory health insurance applies risk equalization among insurers with different enrolled risk distributions. The Swiss Federal Office of Public Health identifies mandatory-insurance equalization variables including age, sex, prior hospitalization or nursing-home residence, and pharmaceutical cost groups derived from the previous year's dispensing. The Joint Institution under the Swiss Health Insurance Act administers the equalization, and the pharmaceutical category list is updated through the applicable administrative process.

The choice of prior drug dispensing illustrates the trade-off in measurable proxies. Medicines can reveal chronic disease when diagnosis data are incomplete, but dispensing can vary with prescribing practice, access to pharmacies, adherence, benefits, and provider incentives. Because these are behaviorally affected, an intervention to increase access to beneficial treatment may also change the recorded risk adjustment category. This need not be improper; it requires causal care in interpreting increased score weights.

The Swiss authority identifies inclusion of pharmaceutical cost groups from 2020 and references an evaluation of their effects. This is administrative documentation of a model modification; no independent re-estimation of its policy impact is undertaken in this dossier. The distinction between a regulator's program description and independent identification of behavioral effects is maintained throughout.

Swiss experience also reinforces the significance of the unit of equalization. Mandatory standardized benefits with insurer premiums subject to legally defined constraints make the redistribution problem materially different from markets where plan benefit generosity varies substantially. The mathematics of residual selection remains applicable, but the observed mechanisms and available policy responses differ.

Gaming, coding intensity, and endogenous measurement

In a static prediction model, increasing the number of correctly observed morbidity indicators can improve information. Under risk-adjusted reimbursement, the financial reward to coding changes the data-generation process. Let a provider or plan choose coding effort c with cost K(c), and let recorded score be r(c,h), depending on underlying health h and documentation effort. If expected plan revenue includes B·r(c,h), the marginal incentive is B·∂r/∂c − K′(c), abstracting from any additional compliance or audit costs. A change in coefficient B or in category weights can alter coding intensity without any change in underlying health.

This expression is a behavioral implication of a hypothetical incentive structure; it does not assert a universal utility function for all insurers. It explains why holding model fit constant while adding a payment-sensitive variable can alter the future measurement process. The empirical upcoding evidence supports studying this channel in Medicare Advantage rather than treating it as purely theoretical.

Upcoding should be used precisely. One possibility is unsupported addition of diagnoses; another is more complete documentation of true illness; a third is shifts in the locus or intensity of legitimate encounters. Administrative score changes alone cannot distinguish these cases. The OIG health-assessment findings identify a high-stakes area for verification, while preserving uncertainty about whether unsupported recorded diagnoses reflect inaccurate documentation or missing follow-up care.

Potential defenses include independent chart validation, source restrictions, clinical hierarchy, coding-pattern monitoring, cross-plan benchmarking, retrospective audits, and adjustments for abnormal coding intensity. Each defense has side effects. Overly strict diagnosis eligibility may undercompensate people with poor access to care or conditions documented outside eligible billing settings. A coding adjustment based on broad averages may penalize organizations that document genuine unmet need. These trade-offs demand stratified evaluation and appropriate due process.

Endogeneity also affects utilization-based variables. If historical pharmacy dispensing drives future revenue, plans may have financial incentives to facilitate access to medications, but the measured association may change after policy adoption. If prior hospitalization drives weights, efforts that successfully prevent future admissions may reduce subsequent risk scores even though care quality improves. Model designers must consider whether the classification system rewards the signal, the treatment, or the health need it represents.

Data leakage is a separate methodological risk: using information from the prediction target period in what is labeled a prospective model overstates deployable accuracy. Freezing feature availability at the time of payment calculation is a basic reproducibility requirement.

Residual risk, selection, welfare, and equity

A competitive insurer evaluates the expected net margin on a marginal member as expected permitted revenue plus expected transfers less expected covered spending and administration. Define Mᵢ = Pᵢ + Tᵢ − E[Yᵢ | Iᵢ] − Aᵢ, with Iᵢ denoting all information relevant to the insurer and Aᵢ attributable administrative cost. If E[Mᵢ | subgroup g] varies systematically across groups, selection incentives can remain. A whole-market calibration ratio of one does not imply equal margins by group, especially when premiums, benefits, and information differ.

Residual risk is multi-component. Idiosyncratic claims risk can be diversified across a large portfolio. Covariate-linked residual cost can become concentrated when plans attract different clinical or demographic groups. Tail risk may be correlated across members during epidemics or region-wide shocks. Model parameter uncertainty and exogenous price inflation also generate systematic variation. Insurance capital and risk sharing address different parts of this residual structure from ex ante predictive adjustment.

A richer model may improve average compensation and still worsen some forms of selection. Geruso, Layton, McCormack, and Shepard's two-margin analysis demonstrates theoretically and with Massachusetts data that transfer policies can change both plan generosity choices and the extensive margin of obtaining coverage. A financial adjustment that shifts relative premiums can affect uninsured entry, so its welfare effects cannot be read from within-market risk-score accuracy alone.

Fairness is not exhausted by equal average prediction errors. An underpredicted subgroup can be disadvantaged through product design even when explicit discrimination is prohibited. A plan rewarded for covering low-risk people may alter networks, specialist access, formularies, complaint resolution, and marketing in ways that shift participation. Risk adjustment should therefore be evaluated against access, enrollment and outcome indicators as well as accounting transfers. Such indicators require separate causal designs before being used to attribute effects to one formula change.

There are also tensions between individual prediction and group equity. A model optimized for total squared error gives substantial influence to high-dollar individuals, while a model optimized for small group ratios may redistribute more across identifiable groups. The correct objective must follow the legal and social purpose of the program. Purely technical loss minimization cannot determine the desired distribution of financial responsibility.

The broad selection literature cautions against assuming that all higher-cost enrollment represents the same informational process. Some populations may exhibit advantageous selection or different willingness to pay. Applying a diagnosis-based compensator in a market with different enrollment elasticities does not mechanically reproduce the results of a regulated compulsory insurance system.

Comparative interpretation and identification limits

The reviewed systems share one economic problem: compensating plans for unequal expected eligible expenditures under rules that limit direct risk pricing. Their architectures differ enough that a common numeric league table based on reported R² would be misleading. The U.S. ACA program uses state-market risk adjustment transfers and a separately defined high-cost pool. U.S. Medicare Advantage uses risk-adjusted federal plan payment and its own normalization and coding processes. The Netherlands makes annual ex ante allocations with multiyear reconciliation. Germany operates morbidity-sensitive sickness-fund allocations and its high-cost risk pool. Switzerland adds pharmaceutical history to its regulated equalization characteristics.

These distinctions can be expressed across five comparable dimensions: the target expenditure; the timing and origin of predictors; the mechanism translating risk into plan revenue; the role of retrospective risk sharing; and the permissible scope of selection and premium differentiation. Each system needs its own effective-year legal record. Claims of comparative superiority would require harmonized and representative cost data, the ability to replicate statutory formulas, and a design that isolates model improvements from changes in benefits, coding, premiums, and enrollees. That evidence is absent from this literature synthesis; therefore no jurisdiction is ranked as definitively best.

The original analytic synthesis yields four implications. First, calibration must be decomposed into level, subgroup distribution, and financial transfer rather than represented by one statistic. Second, coding is an endogenous input in reimbursement regimes, so a frozen-model test on retrospective data may not characterize the equilibrium after implementation. Third, budget neutrality can hold algebraically while distributional errors persist across groups. Fourth, an insurer's residual financial risk must be separated into predictable selection exposure and unpredictable claims volatility, because different policy instruments address them.

These implications follow from the definitions, transfer identity, and cited comparative evidence. They are methodological conclusions within established insurance economics and applied statistics; no new doctrine, proprietary index, or unverified priority claim is proposed.

Limits of available evidence

The official pages document legal design and year-specific administrative choices. They do not independently establish clinical accuracy or program welfare. Peer-reviewed studies identify behavior in their own markets and periods, not globally transportable causal parameters. The historical RAND sample and the early Dutch comparison are valuable for mechanism and chronology but cannot validate current 2026 or 2027 predictions. Publicly linked regulatory sources do not provide an accessible unified patient-level multi-country dataset. Hence, this dossier makes no claims of new coefficient estimates, patient-level replication, numerical cross-country accuracy rankings, or formal meta-analysis.

Some empirical material comes through journal article records and abstracts rather than fully independently reproduced datasets; that access level is documented in the evidence ledger. Administrative policy details are as of the research cut-off of November 14, 2026, and any later change requires a separately documented revision rather than retroactive alteration of the evidence.

Reproducible research and technical audit specification

An empirical replication should begin by freezing one jurisdiction, one payment year, one program version, and its canonical governing documents. Preserve a manifest of source URLs and their capture dates, a legal crosswalk mapping the target population to eligibility rules, and a versioned diagnosis-to-category mapping. Describe member eligibility, enrollment start/end times, death, births, plan switches, deductibles, cost-sharing, plan actuarial value, claims runout, pharmacy claims, provider prices, and excluded benefits. The dataset must be governed by lawful access and privacy protection.

Split historical observations chronologically. Estimate model parameters using only permitted historical information and validate on a later cohort. Preserve a strict prediction timestamp for each variable, including any retrospective corrections that the actual payment algorithm would allow. Compute individual error, total dollars, group predictive ratios, plan-level errors, top-tail error, and within-market transfers. Bootstrap or otherwise quantify sampling uncertainty with clustering appropriate to plan and repeated-member correlation.

Run at least four specified stress analyses: a coding-sensitivity scenario where category capture probabilities differ by insurer; a secular-trend scenario where price and utilization growth differ; a selection scenario where high-need groups sort disproportionately into some plans; and an extreme-cost scenario before and after any high-cost pooling rule. These are diagnostic scenarios, not observed findings, until parameters are empirically estimated. To evaluate payment fairness, calculate post-transfer expected margins for relevant groups under the actual legal formula, not merely changes in the predicted score.

Implement checks for zero and negative expenditure anomalies, impossible enrollment duration, age-boundary inconsistencies, incompatible diagnosis versions, duplicate records, inconsistent care dates, high-cost truncation errors, and plan-year identifier changes. Verify that a small perturbation to one category changes scores only through authorized hierarchy and interaction terms. Confirm that an annual transfer implementation satisfies its own budget constraint when the statute requires budget neutrality, allowing for explicit program adjustments.

The bibliographic audit has the same rigor: each external contextual link must point to a source represented in the evidence ledger and in the complete References; the linked phrase must state only what the source substantiates. A method paper can support a methodological distinction but cannot prove the regulatory details of a later program. An official regulator statement can prove the announced parameter, but not that the parameter optimizes welfare. The date and jurisdiction belong to the claim.

All numerical illustrations in this monograph are algebraic demonstrations using stated hypothetical inputs. No live enrollee, provider, claims, or risk-score dataset was processed. Reproduction of actual predictions remains an independently scoped empirical project requiring data access and statutory model components.

Conclusion

Health risk adjustment is the institutional conversion of recorded morbidity and other eligible risk characteristics into financial compensation for expected insured expenditures. Its success depends on a chain of linked decisions: the cost target, information window, diagnosis mapping, statistical estimation, calibration population, plan-level aggregation, market-specific transfer mechanism, and rules governing retrospective compensation and data verification.

The formal analysis establishes that aggregate calibration can coexist with subgroup underpayment; an exact budget-neutral transfer identity can coexist with residual insurer selection incentives; and statistical predictive improvement can coexist with greater sensitivity to endogenous coding. These results are compatible with the economic evidence on health insurance selection and the empirical record of diagnosis-based payment behavior. They also explain why data validation is a constitutive element of risk adjustment governance rather than a peripheral administrative check.

The comparative primary-source review documents distinct implementations in the United States, Netherlands, Germany, and Switzerland. Their differences in financing, timing, predictors, risk sharing, and premiums delimit the interpretation of any model performance result. Effective-year legal and technical documents govern each specific system. An international assessment should test standardized and subgroup calibration, expected post-transfer margins, actual access and selection outcomes, and coding responses under clearly defined denominators.

The scholarly contribution is an integrated, source-specific reconstruction of established mechanisms and their interactions, including the distinction between prediction quality and transfer adequacy. The study produces analytical conclusions without constructing a new theoretical system or claiming evidence that has not been measured. Future research can test these conclusions with secure longitudinal data and frozen algorithm versions; the scientific record in this monograph is complete within the documentary and analytical scope stated above.

References

  • Centers for Medicare & Medicaid Services (CMS). 2011. Evaluation of the CMS-HCC Risk Adjustment Model, final report, prepared by Gregory C. Pope, John Kautter, Melvin J. Ingber, Sara Freeman, Rishi Sekar, and Gordon Newhart, RTI International. CMS evaluation record.

  • Centers for Medicare & Medicaid Services (CMS). 2025. HHS Notice of Benefit and Payment Parameters for 2026 Final Rule, fact sheet. 2026 payment notice.

  • Centers for Medicare & Medicaid Services (CMS). 2025. 2026 Medicare Advantage and Part D Rate Announcement, April 7. 2026 MA announcement.

  • Centers for Medicare & Medicaid Services (CMS). 2026. Premium Stabilization Programs: Risk Adjustment and Data Validation, program documentation and benefit-year reports. Premium-stabilization inventory.

  • Centers for Medicare & Medicaid Services (CMS). 2026. Marketplace Regulations and Guidance, program documentation and risk adjustment software instructions. Regulatory guidance inventory.

  • Einav, Liran, and Amy Finkelstein. 2011. “Selection in Insurance Markets: Theory and Empirics in Pictures.” Journal of Economic Perspectives 25 (1): 115–138. Journal record.

  • Federal Office for Social Security (Bundesamt für Soziale Sicherung). 2026. “Risikostrukturausgleich: Überblick.” System overview.

  • Federal Office for Social Security (Bundesamt für Soziale Sicherung). 2026. “RSA – einfach erklärt.” Explanatory statement.

  • Federal Office for Social Security (Bundesamt für Soziale Sicherung). 2026. “Weiterentwicklung” (risk adjustment reforms). Reform documentation.

  • Federal Office for Social Security (Bundesamt für Soziale Sicherung). 2026. “Prüfungen” (risk adjustment data audits). Audit documentation.

  • Federal Office for Social Security (Bundesamt für Soziale Sicherung). 2026. “Datenzusammenstellungen und Auswertungen” (risk-group distributions and yearly files). Statistical archive.

  • Geruso, Michael, and Timothy J. Layton. 2017. “Selection in Health Insurance Markets and Its Policy Remedies.” Journal of Economic Perspectives 31 (4): 23–50. Journal record.

  • Geruso, Michael, and Timothy J. Layton. 2020. “Upcoding: Evidence from Medicare on Squishy Risk Adjustment.” Journal of Political Economy 128 (3): 984–1026. PubMed bibliographic record.

  • Geruso, Michael, Timothy J. Layton, Grace McCormack, and Mark Shepard. 2023. “The Two-Margin Problem in Insurance Markets.” Review of Economics and Statistics 105 (2): 237–257. PubMed bibliographic record.

  • Newhouse, Joseph P., Willard G. Manning, Emmett B. Keeler, and Elizabeth M. Sloss. 1989. “Adjusting Capitation Rates Using Objective Health Measures and Prior Utilization.” Health Care Financing Review 10 (3): 41–54. Full-text journal article.

  • Shmueli, Galit. 2010. “To Explain or to Predict?” Statistical Science 25 (3): 289–310. Publisher DOI.

  • Swiss Federal Office of Public Health (Bundesamt für Gesundheit). 2026. “Krankenversicherung: Risikoausgleich.” Government program documentation. Swiss equalization policy.

  • United States Congress. 2010. Patient Protection and Affordable Care Act, Pub. L. 111-148, § 1343, codified at 42 U.S.C. § 18063. Enacted statutory text.

  • United States Department of Health and Human Services (HHS). 2026. “Patient Protection and Affordable Care Act, HHS Notice of Benefit and Payment Parameters for 2027; and Basic Health Program.” Federal Register 91 (May 20), 29526 ff., FR Doc. 2026-10050. 2027 final rule.

  • United States Department of Health and Human Services, Office of Inspector General (HHS-OIG). 2024. “Medicare Advantage: Questionable Use of Health Risk Assessments Continues To Drive Up Payments to Plans by Billions.” Report OEI-03-23-00380, October. Official evaluation.

  • van de Ven, Wynand P. M. M., René C. J. A. van Vliet, and Leida M. Lamers. 2004. “Health-Adjusted Premium Subsidies in the Netherlands.” Health Affairs 23 (3): 45–55. PubMed bibliographic record.

  • Zorginstituut Nederland. 2025–2026. “Zvw 2026” (risk-equalization regulations, annual ex ante model, and methodological attachments). 2026 program file.

  • Zorginstituut Nederland. 2026. “Wat is risicoverevening?” (fund distribution, ex ante contribution, and settlement sequence). Risk-equalization explanation.