Averages lie, distributions tell the truth: a better way to judge risk adjustment factor coding

A lower average risk adjustment factor (RAF) score in one clinic does not automatically mean that clinic is undercoding. RAF is a measure of the expected cost and clinical risk of a specific patient population — not a direct report card on coding quality. Judging documentation and coding performance on the absolute average RAF of the underlying population is one of the most common analytic mistakes in accountable care organization (ACO) and value-based care oversight, and it can lead groups to reward the wrong clinics and pressure the right ones.

The central error is straightforward: comparing average RAF scores across clinics without first asking whether the clinics care for comparable patients. A practice with a younger, healthier, community-dwelling Medicare population should generally have a lower average RAF than a practice whose panel includes more older beneficiaries, dual-eligible patients, originally disabled beneficiaries, patients who are institutionalized or patients with a genuinely higher burden of chronic illness. Centers for Medicare & Medicaid Services-Hierarchical Condition Categories (CMS-HCC) risk adjustment models intentionally include both demographic characteristics and documented diagnoses because both predict expected cost, according to the CMS’ own risk adjustment overview.

RAF has 2 very different components

At the patient level, RAF is not a synonym for the number of diagnoses coded. It is the sum of demographic factors, diagnosis-based HCC factors, and applicable interactions. Demographic factors include age, sex, Medicaid or dual status, original disability entitlement and community-versus-institutional status. Diagnosis factors arise when diagnoses, coded using “International Classification of Diseases, Tenth Revision, Clinical Modification,” map to HCCs and receive their associated weights, subject to hierarchies and interactions.

That means a patient can have a low RAF for two very different reasons. The patient may be truly younger and healthier, with little or no HCC-relevant disease. Or the patient may have clinically meaningful chronic conditions that were not assessed, documented, coded or submitted in a way that produces HCC risk. Those are fundamentally different situations, and an average RAF alone cannot tell you which one is occurring — or which clinic is coding better.

Why the distribution matters more than the mean

The more useful analysis is not the clinic’s mean RAF. It is the distribution of patient-level RAF scores across the panel. When you plot RAF for every attributed patient, many Medicare populations do not form a single smooth bell curve. They often show two or more concentrations, or peaks. There is typically a lower peak made up largely of patients whose RAF is driven primarily by demographics — age and sex — with few or no HCC-producing diagnoses. There is usually a higher, broader distribution of patients with one or more chronic conditions contributing to diagnosis-based risk. And there is often a long right tail of patients who are medically complex, frail, disabled or institutionalized.

The size, position and separation of those peaks tell you far more than the average does. A large cluster at low RAF driven almost entirely by demographic risk may reflect a genuinely healthier, younger Medicare panel — or it may reflect a diagnosis-capture gap. A broad middle peak reflects the typical mix of patients with common chronic disease burden. A long right tail reflects multimorbidity, frailty, or institutional status, and it demands that severity documentation be specific and compliant.

The presence of a low-RAF peak is not inherently a coding problem. A clinic may simply have more newly Medicare-eligible patients, relatively healthy retirees or patients with few chronic conditions. A practice that attracts or retains healthier patients should not be analytically penalized for it. The right question about that low-RAF peak is whether those patients are actually low risk or whether they have RAF-relevant conditions that are not being captured.

Why clinic-to-clinic averages can invert the truth

Consider two clinics in the same ACO. Clinic A has an average RAF of 0.85 and Clinic B has an average RAF of 1.20. The surface-level conclusion is that Clinic A is undercoding and Clinic B is coding well. The more plausible interpretation is very different. Clinic A may have a younger, healthier panel and may actually be documenting its lower-risk population completely and accurately. Clinic B may have an older, more dual-eligible, or clinically sicker panel and may still be missing common HCCs in a meaningful subset of its patients.

A high average RAF can result from a legitimately higher-risk mix, but it can also coexist with incomplete coding. A very high-risk panel can still miss diabetes with complications, chronic kidney disease (CKD) stage specificity, heart failure type, chronic obstructive pulmonary disease (COPD), vascular disease, major depression severity, morbid obesity, substance use disorders, or other clinically relevant risk-adjusting diagnoses that are actually being evaluated and managed. Conversely, a lower-RAF clinic may have excellent, disciplined documentation. Its physicians may accurately document every active, evaluated and managed condition — but the patients simply do not have as many conditions that map to HCCs.

Average RAF tells you the level of measured risk. It does not tell you whether documentation is complete. In an ACO where shared savings, benchmarks, and internal performance dashboards depend on risk scores, mistaking one for the other punishes accurate coders and rewards incomplete ones.

Compare like with like

Coding performance should be evaluated only after separating patient mix from documentation quality. A credible RAF review compares clinics using clinically and demographically similar cohorts — not raw average RAF. Useful stratifications include age bands, sex, dual eligibility, original disability entitlement, community versus institutional status, new-to-Medicare versus long-standing Medicare, number of chronic medications, prior-year HCCs and prior-year RAF, hospitalization and specialty utilization, and disease-specific cohorts such as diabetes, heart failure, COPD, CKD, dementia, serious mental illness and cancer.

Comparing the RAF distribution of 70- to 74-year-old, community-dwelling, non-dual beneficiaries across clinics is far more informative than comparing overall averages. If, within that matched cohort, one clinic has a substantially larger share of patients with no diagnosis-based risk at all, that is a signal worth investigating. It still does not prove undercoding. Rather, it identifies a subgroup worth understanding.

What to look for in the low-RAF, demographics-only peak

A sound coding assessment starts with patients, not clinic averages. For members in the demographics-only cluster, look for clinical evidence that a legitimate HCC-relevant diagnosis may need to be reassessed and, if appropriate, documented at the next face-to-face encounter. Signals worth reviewing include chronic medications suggesting an ongoing condition — insulin or multiple diabetes agents, inhalers with maintenance therapy, guideline-directed heart failure therapy, anticoagulants, antiepileptics, immunosuppressants, disease-modifying specialty medications, dialysis-related medications or advanced psychiatric therapies. Problem lists and historical claims may show chronic disease that does not appear in current-year encounter documentation. Specialist notes, discharge summaries, imaging and laboratory data may suggest active chronic conditions, such as an estimated glomerular filtration rate consistent with CKD stage, a hemoglobin A1c with complication documentation, or an ejection fraction and New York Heart Association Functional Classification for heart failure. Utilization patterns may be inconsistent with a patient who is genuinely disease-free. Prior-year HCCs may have dropped off without clinical evidence of resolution. Diagnoses may appear in narrative documentation but not be clearly assessed, linked to a plan, or coded to the appropriate level of specificity.

This is an exercise in accurate, compliant documentation — not a hunt for codes. A diagnosis should be reported only when supported by the medical record and relevant to that encounter’s assessment, care, treatment, monitoring or management. CMS risk adjustment depends on documented diagnoses submitted through qualifying encounters. Conditions that are not documented and coded appropriately will not contribute to the risk score, and conditions that are not clinically supported should not be reported.

A better way to look at RAF

A clinic-level RAF review should go well beyond an average. It should look at both the mean and the median, since a small number of patients with medically complex cases can pull a clinic’s average up materially while the median reveals what the typical patient looks like. It should visualize the patient-level RAF distribution, ideally overlaid across comparable clinics, so that the shape of the curve — not just its center — becomes visible. It should quantify the share of patients with demographics-only RAF or no diagnosis-based contribution, and break RAF down by age band and other demographic segments. It should track how many patients have zero, one, two, three, or four or more HCCs, and follow prior-year to current-year RAF movement at the patient level. Prior-year HCC recapture should be validated by chart review rather than automatic carry-forward, and condition-specific prevalence should be compared with matched benchmark populations. A sound review also samples both low-RAF patients, to look for undercapture, and high-RAF patients, to confirm that documentation supports the reported severity.

The right questions

The wrong question is why one clinic’s RAF is lower than another’s. The better questions are as follows:

• Does a clinic’s RAF distribution fit the age, eligibility, disease burden and utilization profile of its patients?
• Does the low-RAF, demographics-only cluster contain clinical evidence of active conditions that are being evaluated and managed but not captured?
• When clinically similar patients are compared across clinics, are there real documentation or coding gaps in either direction?

That framing distinguishes genuine differences in population health from real opportunities to improve documentation, and it guards against the equally serious error of assuming that a high-RAF clinic must be coding well.

Bottom line

Absolute average RAF is a population characteristic, not a coding grade. Two clinics with very different averages can both be coding accurately — or both be missing important diagnoses — depending on who their patients are. A lower-RAF clinic may in fact be the more complete coder, while a higher-RAF clinic may still be leaving legitimate, well-documented conditions on the table. Judging coding performance requires looking inside the RAF distribution: the shape of the curve, the size of the demographics-only peak, the behavior of matched cohorts, and the clinical reality of individual patients. Only then can an ACO or medical group tell the difference between a healthy panel, an unhealthy panel, complete documentation, and missed diagnoses — and only then can it act on the right ones.

Robert Resnik, M.D., MBA, is a board-certified internal medicine physician practicing in Cary, North Carolina. He earned his medical degree from Eastern Virginia Medical School and completed his residency at East Carolina University. He also holds an MBA from Duke University.