17  HCCs & Risk Adjustment

Two patient populations can have different healthcare costs even when they receive equally effective care. One may contain more people with complex chronic conditions, more older adults, or more patients who need ongoing treatment. Comparing their spending without considering those differences can lead us to the wrong conclusion.

Risk adjustment accounts for specified patient characteristics when estimating expected cost or outcomes. In this chapter, we’ll focus on claims-based cost models, particularly hierarchical condition categories (HCCs). We’ll explain what a risk score measures, how diagnosis codes become model inputs, and why a calculated score is different from a payment or a complete description of someone’s health.

The central question is simple: given the information the model uses, what cost would we expect? The answer is useful only when we understand the population, time period, and outcome for which the model was developed.

17.1 Separate three uses of patient information

Claims can support several related but distinct tasks.

A description of illness burden might report the prevalence of diabetes, the number of chronic conditions per patient, or the distribution of functional limitations when those data are available. These describe selected aspects of a population.

A prediction model estimates an outcome, such as next year’s medical spending or the probability of a hospital admission. Its usefulness depends on how well it predicts that outcome in the intended population. A model that predicts dollars is not automatically a good model of avoidable admissions, clinical severity, or care management needs.

A payment methodology uses model outputs together with program rules to determine payments, benchmarks, or transfers. The model is one component of that methodology. The CMS introduction to risk adjustment explains its role in estimating expected spending for patients with different needs.

This distinction matters in daily work. A higher HCC score can contribute to a higher expected cost under a payment model. It does not establish that a patient needs a particular treatment, that their care was poor, or that every increase in their spending was justified.

17.2 Understand the model family

“HCC model” is not a complete specification. Different programs use different condition categories, coefficients, populations, and timing conventions.

Model or program	Main purpose	Important distinction
CMS-HCC	Risk adjustment for Medicare Part C medical spending in applicable populations	Generally prospective: diagnoses in a base period inform expected spending in a later period
CMS ESRD models	Risk adjustment for beneficiaries in specified end-stage renal disease states	Dialysis, transplant, and functioning-graft circumstances have distinct rules
RxHCC	Risk adjustment for Medicare Part D	Predicts the relevant prescription drug liability rather than Part C medical spending
HHS-HCC	Risk adjustment in applicable individual and small-group insurance markets	Concurrent models use information from the benefit year to estimate that year’s plan liability
PACE	Payment for the Program of All-Inclusive Care for the Elderly	A program with its own applicable models and payment rules, not a single interchangeable HCC model

CMS publishes the Medicare model resources through its risk adjustment page. The original HHS-HCC model paper explains its concurrent design and separate adult, child, and infant models. Annual specifications determine the model actually used for a benefit year.

HHS-HCC should not be treated as a universal model for all commercial insurance. Its payment context is the applicable individual and small-group markets. Similarly, an ACO can use risk adjustment in its financial benchmark without becoming an insurance plan or following every Medicare Advantage payment rule.

For HHS-operated risk adjustment, risk scores feed a transfer methodology that accounts for differences among participating plans. State-market transfers are designed to balance payments and charges within the relevant pool; the program also includes a high-cost risk pool. “Zero-sum” is therefore an incomplete explanation of an individual plan’s payment. See the CMS transfer reports and methodology resources.

17.3 Read a risk score correctly

A risk adjustment factor is a relative model output. A score above the model’s reference level indicates higher expected expenditure under that model. The reference population and scaling matter: a score of 1.0 does not necessarily equal the average of the health plan or patient panel in front of us.

Risk scores are not probabilities. A score of 1.2 does not mean a 120% chance of hospitalization. Nor does it guarantee an individual will incur 20% more cost than another patient. Models estimate patterns across populations; individual spending can vary widely around those expectations.

For an analytical comparison, retain the score’s full identity: model family, version, payment or benefit year, diagnosis period, model segment, and whether the value is raw, normalized, or adjusted for payment. A column named risk_score without that context is difficult to interpret and easy to misuse.

17.4 From diagnoses to HCCs

HCC stands for hierarchical condition category. Diagnosis codes are mapped into condition categories, and model rules determine which categories contribute to the score. The hierarchy prevents specified related categories from contributing independently when a higher category supersedes a lower one. It is not a rule to retain only the patient’s single most severe condition.

The basic process has several stages.

17.4.1 Establish the population and segment

For a CMS-HCC calculation, first establish enrollment and demographic information. Depending on the model, relevant distinctions include age, the model’s sex variable, disability and original entitlement information, Medicaid dual eligibility, and community or institutional status.

Use the exact definition of each segment. An inpatient hospital claim does not, by itself, establish long-term institutional status for risk adjustment. A newly insured member in a particular health plan is also not necessarily a new enrollee under the Medicare model’s definition.

New-enrollee scoring addresses situations without the required diagnostic history and can use demographic factors rather than the continuing-enrollee disease model. Separate ESRD rules apply where appropriate. The CMS model evaluation report explains the rationale for these population distinctions; use the applicable year’s software and specifications to implement them.

17.4.2 Identify eligible diagnosis evidence

Select diagnoses from the allowed data sources and observation period. A diagnosis can be a valid ICD-10-CM code and still be ineligible for a particular payment model because of its source, service circumstances, submission status, or timing.

For Medicare Advantage encounter submissions, the MAO-004 report communicates whether diagnoses meet the relevant risk-adjustment filtering rules. Passing that filter does not guarantee that a diagnosis maps to a payment HCC. The CSSC guide to Model Output Reports explains this distinction.

Resolve replacements and deletions before accumulating diagnoses. Ten copies of one accepted diagnosis should not generate ten instances of the same HCC. Preserve the supporting records so each modeled condition can be traced back to its source.

17.4.3 Map, apply hierarchies, and calculate factors

Use the diagnosis mapping and software corresponding to the selected model and run. The same diagnosis can map differently across model versions. The same HCC number can also mean different things in different models.

After mapping, apply the published hierarchies. Then calculate the applicable demographic factors, surviving condition factors, and any interactions or count factors specified by that model. An interaction represents an additional modeled effect of a combination of characteristics; it must come from the specification, not an analyst’s intuition.

CMS publishes annual code mappings and model software, including separate initial and midyear/final resources where applicable. For example, the 2026 model resource page provides these artifacts. Keep their versions together so the calculation does not accidentally combine a mapping from one release with coefficients from another.

17.5 Work through the logic

Consider a hypothetical continuing enrollee with three accepted diagnosis records. Two are repeated reports of the same condition; the third represents a related, more severe manifestation. Assume the selected model’s published mapping and hierarchy say that the more severe category supersedes the less severe category.

Stage	Result
Source evidence	Three accepted diagnosis records
Mapping and deduplication	Two distinct condition categories
Hierarchy	One contributing category; the other is superseded
Scoring	Applicable demographic factors plus the surviving category and any specified additional factors

This example deliberately omits HCC numbers and coefficients because those depend on the actual model. It illustrates three rules: repeated documentation does not multiply a condition’s factor; a hierarchy can remove a related category’s contribution; and unrelated conditions can still contribute separately.

If the patient also has a diagnosis that does not map to a payment HCC, that diagnosis remains meaningful clinical information. It simply does not add a condition factor in this model. Likewise, a patient without a contributing HCC can still have a nonzero demographic risk score.

17.6 Move from a model output to a payment score

A raw score is the output of the specified model factors. A normalized score adjusts that output using the applicable normalization methodology. A payment score may also reflect coding-pattern adjustments, model blends, or other program rules.

When two models are blended, calculate each model completely and apply its own required adjustments at the specified stage. Different models can have different normalization factors. Averaging raw scores and dividing by a single convenient factor will generally not reproduce a methodology that separately normalizes them.

A dated example helps show why “current model” is insufficient. For calendar year 2026, CMS completed the transition to the 2024 CMS-HCC model, commonly called V28, for the applicable non-PACE CMS-HCC population. This does not mean V28 is the model for every Medicare beneficiary or benefit: ESRD and Part D have separate models, and PACE has separate transition rules. The 2026 implementation memorandum specifies those distinctions, while the 2026 Rate Announcement supplies the payment-year policy.

A final payment still requires the rest of the program’s payment calculation. Do not label a raw HCC score as funded revenue, or multiply it by an arbitrary average premium and present the result as a CMS payment.

17.7 Reconcile with CMS reports

For Medicare Advantage work, several files provide different views of the process:

File	What it helps establish
Monthly Membership Detail Report, commonly called the MMR	Member-level payment information, risk factors, and adjustments for identified payment periods
Model Output Report (MOR)	The HCCs or RxHCCs used in a specified model run
MAO-004	Risk-adjustment eligibility information for diagnoses from encounter submissions

The Plan Communications User Guide documents the report layouts. Interpret effective periods and adjustment records rather than treating the latest file as one final row per person.

When a local score differs from a reported score, compare the inputs in order: population and segment, diagnosis window and cutoff, accepted evidence, mapped categories, hierarchy, factors, and payment adjustments. A difference in aggregate scores alone does not identify which stage is wrong.

17.8 Interpret change with care

A population’s average score can rise because its health needs changed. It can also rise because diagnosis capture improved, claims became more complete, the model changed, or different people entered the population. These explanations have different implications.

For trend analysis, separate changes in membership from changes among continuing members. Hold the model specification constant when the question is about changes in modeled burden, and report the actual payment specification separately when the question is about reimbursement. Align observation windows and claims maturity before comparing scores.

Coding intensity is especially relevant because a claims-based model sees documented conditions. More complete documentation can increase measured risk without an equivalent change in underlying health. Missing diagnoses can lower measured risk without indicating better health. Accurate documentation is the objective; an increasing score is not, by itself, evidence of improvement.

CMS’s Medicare Advantage Risk Adjustment Data Validation program checks whether submitted diagnoses are supported by medical records. A suspected condition or a condition carried forward from an earlier year should not be treated as qualifying current evidence without meeting the applicable requirements.

Risk adjustment makes comparisons more informative, but it cannot remove every difference between populations. Read the score alongside cost, utilization, clinical measures, enrollment history, and data completeness. The most useful analysis explains both what the model accounts for and what remains outside it.