Risk adjustment

When Medicare first allowed beneficiaries to enroll in HMOs in 1982, it paid plans a monthly per-person capitated rate set at 95% of the Adjusted Average Per Capita Cost (AAPCC) for each county. The AAPCC adjusted payments using only five demographic factors: age, sex, Medicaid enrollment (as a proxy for poverty), institutional status (nursing home residency), and working-aged status.

This system had a fatal flaw: it could not distinguish between a healthy 72-year-old who jogs daily and a 72-year-old managing diabetes, heart failure, and COPD. Both received identical payments. Research from the early 1990s proved that managed care plans were systematically enrolling healthier beneficiaries, collecting the full AAPCC payment, and pocketing the difference. The AAPCC explained only about 1% of the variation in actual healthcare costs — essentially useless for predicting individual spending.

County-level rates also varied wildly. In 1997, the annual capitation rate for a 65+ beneficiary ranged from roughly $221/month in Arthur County, Nebraska, to about $767 in Richmond County (Staten Island), New York.

The BBA of 1997 was the watershed moment. Congress created the Medicare+Choice (M+C) program and mandated that CMS implement health-status-based risk adjustment by January 2000. The BBA also required managed care organizations to begin reporting inpatient encounter data starting in 1998, creating the raw data pipeline that would eventually power all future risk adjustment models.

To meet the BBA deadline, CMS implemented the Principal Inpatient Diagnostic Cost Group (PIP-DCG) model in 2000. This was the first diagnosis-based risk adjustment model in Medicare history. It used a beneficiary’s most serious inpatient hospital diagnosis from the prior year, combined with demographics, to predict the following year’s costs.

However, PIP-DCG was intentionally limited — it was a transition tool. Only 10% of total capitation payments were risk-adjusted using PIP-DCG; the other 90% still used the old demographic AAPCC formula. The PIP-DCG model improved predictive accuracy from about 1% R-squared (demographics only) to roughly 6.2%, a six-fold improvement.

But it still had a major shortcoming: it only counted conditions that resulted in a hospital admission. Plans that kept patients healthy and out of the hospital could actually be penalized with lower-appearing risk scores.

In 2004, CMS implemented the full CMS-HCC (Hierarchical Condition Categories) model for Medicare Advantage payments. This was a quantum leap. The model used diagnoses from all care settings — inpatient, hospital outpatient, and physician offices — to predict the following year’s costs. It classified more than 15,000 ICD-9-CM diagnosis codes into 804 Diagnosis Groups, which were then aggregated into approximately 70 Hierarchical Condition Categories.

The hierarchical structure was a deliberate design choice. When a patient has both a less severe and a more severe manifestation of the same disease (e.g., diabetes without complications and diabetes with kidney complications), the model only counts the more severe condition. This prevents double-counting while accurately reflecting disease burden.

The CMS-HCC model increased predictive accuracy to about 11.2% R-squared — eleven times better than demographics alone. The model was phased in from 2004 to 2007.

The MMA of 2003 renamed Medicare+Choice as "Medicare Advantage" effective 2006 and provided additional funding. This is when MA enrollment began its long upward trajectory from roughly 11% of Medicare beneficiaries toward today’s 50%+ penetration.

CMS officially launched the Risk Adjustment Data Validation (RADV) audit program in 2008 to verify that diagnoses submitted by MA organizations were actually supported by medical records.

When the Affordable Care Act established health insurance Marketplaces in 2014, HHS needed a risk adjustment model for the commercial population. They could not simply use the CMS-HCC model because it was designed for the Medicare population (age 65+ and disabled).

HHS developed the HHS-HCC model by adapting the CMS-HCC framework for a younger, commercially insured population. Key differences: HHS-HCC uses concurrent (current-year) rather than prospective (prior-year) scoring; covers all ages (infants through adults); adjusts for actuarial value (plan metal tiers); predicts combined medical and drug spending; and was calibrated on employer-sponsored insurance claims.

The HHS-HCC model launched with 127 HCC codes and 7,768 ICD-10 codes. It uses 15 separate sub-models covering all combinations of age group (infant, child, adult) and plan metal level.

CMS gradually transitioned from using diagnoses submitted through the Risk Adjustment Processing System (RAPS) to using MA Encounter Data System (EDS) records. RAPS was a simplified submission format that only captured diagnoses. EDS captures full encounter-level detail similar to fee-for-service claims. The transition was completed in 2022.

A persistent issue since the CMS-HCC model’s inception has been "coding intensity" — the tendency of MA plans to document conditions more thoroughly than FFS providers, resulting in higher risk scores. MedPAC estimates MA coding intensity inflates risk scores by approximately 10.8%. The ACA mandated a minimum coding intensity adjustment (CIA) of 5.9%. Despite this, MedPAC estimates the government still overpays MA plans by approximately 20% per beneficiary, costing an estimated $84 billion in 2025.

The 21st Century Cures Act (2016) mandated additional changes, including adding variables that count the number of payment conditions and adding conditions for mental health, substance use disorder, and chronic kidney disease.

The prior CMS-HCC model version (V24) was built on ICD-9-CM codes and calibrated using 2014 diagnoses predicting 2015 costs. By the early 2020s, the model was significantly outdated.

V28 was recalibrated using 2018 diagnoses predicting 2019 expenditures. It expanded from 86 to 115 HCCs, while reducing qualifying ICD-10 codes from 9,797 to 7,770. Many previously weighted conditions were dropped because they did not meaningfully drive costs. V28 requires greater documentation specificity.

CMS implemented a three-year transition: 33% V28 / 67% V24 in 2024, 67% V28 / 33% V24 in 2025, and 100% V28 in 2026.

V28 has compressed RAF scores for many MA plans and provider organizations. Conditions that previously generated significant risk score value no longer map to HCCs or carry reduced weights.

Calendar year 2026 marks the completion of the V28 transition. All MA risk scores are now calculated 100% on the 2024 CMS-HCC model (V28). CMS finalized an overall 5.06% average payment increase for 2026 (approximately $25 billion).

On January 26, 2026, CMS released the CY 2027 Advance Notice proposing the most significant structural change since V28 itself: excluding unlinked chart review records ($7.2 billion impact), excluding audio-only telehealth diagnoses, updating calibration data (2023 Dx / 2024 costs), and a base payment increase of just 0.09%.

In May 2025, CMS announced aggressive expansion of the RADV program: expanding the coding review workforce from 40 to approximately 2,000, auditing all ~550 eligible contracts annually, and reviewing 35–200 records per plan. A federal judge vacated the RADV Final Rule in September 2025 but enforcement pressure continues.