The mechanics of risk adjustment and incentives for coding intensity in Medicare

Abstract

Objective

To study diagnosis coding intensity across Medicare programs, and to examine the impacts of changes in the risk model adopted by the Centers for Medicare and Medicaid Services (CMS) for 2024.

Data Sources and Study Setting

Claims and encounter data from the CMS data warehouse for Traditional Medicare (TM) beneficiaries and Medicare Advantage (MA) enrollees.

Study Design

We created cohorts of MA enrollees, TM beneficiaries attributed to Accountable Care Organizations (ACOs), and TM non‐ACO beneficiaries. Using the 2019 Hierarchical Condition Category (HCC) software from CMS, we computed HCC prevalence and scores from base records, then computed incremental prevalence and scores from health risk assessments (HRA) and chart review (CR) records.

Data Collection/Extraction Methods

We used CMS's 2019 random 20% sample of individuals and their 2018 diagnosis history, retaining those with 12 months of Parts A/B/D coverage in 2018.

Principal Findings

Measured health risks for MA and TM ACO individuals were comparable in base records for propensity‐score matched cohorts, while TM non‐ACO beneficiaries had lower risk. Incremental health risk due to diagnoses in HRA records increased across coverage cohorts in line with incentives to maximize risk scores: +0.9% for TM non‐ACO, +1.2% for TM ACO, and + 3.6% for MA. Including HRA and CR records, the MA risk scores increased by 9.8% in the matched cohort. We identify the HCC groups with the greatest sensitivity to these sources of coding intensity among MA enrollees, comparing those groups to the new model's areas of targeted change.

Conclusions

Consistent with previous literature, we find increased health risk in MA associated with HRA and CR records. We also demonstrate the meaningful impacts of HRAs on health risk measurement for TM coverage cohorts. CMS's model changes have the potential to reduce coding intensity, but they do not target the full scope of hierarchies sensitive to coding intensity.

What is known on this topic

• Medicare Advantage (MA) plans use Health Risk Assessments (HRAs) and chart review (CR) records to increase the severity of measured health risk in their covered populations.

• Prior work estimated that health risks for MA plans, measured by Hierarchical Condition Category (HCC) scores, increased by 4.1% in 2015 through CR records.

• This increase in measured health risk results in significant additional Medicare reimbursements to MA plans.

What this study adds

• We computed HCC prevalence and scores, excluding HRA and CR records, for MA enrollees compared to Traditional Medicare (TM) beneficiaries with and without attribution to Accountable Care Organizations.

• We identified a 9.8% increase in the MA cohort and 0.9% (non‐ACO) to 1.2% (ACO) in propensity‐score matched TM cohorts due to CR (MA) and HRA records (TM and MA).

• We identified eight HCC hierarchical groups, driving 69% of the MA coding intensity impact on HCC scores, of which five are not addressed by upcoming HCC methodology changes.

1. INTRODUCTION

Public and private health plans are increasingly contracting with providers based on the total cost of care. Their goal is to create incentives to manage both quality of care and overall resource use. Total‐cost‐of‐care contracting can take the form of a capitated payment, or fee‐for‐service payment with shared savings if cost is below a per‐capita target. Risk adjustment is an important feature of these programs, to discourage provider groups from selecting healthy patients to hold down costs.

The Shared Savings Program and Medicare Advantage (MA) are two important examples of Medicare programs that use risk adjustment. The Shared Savings Program is a component of Traditional Medicare (TM) in which provider groups form Accountable Care Organizations (ACOs) that enter a shared savings arrangement with Medicare. ACOs are paid on a fee‐for‐service basis during the year, with a retrospective review of cost, quality, and patient experience for the panel of patients attributed to the ACO. If the ACO meets cost and quality targets, it is eligible to share in cost savings computed with respect to the risk‐adjusted target cost.

In MA, the beneficiary selects a private MA plan in lieu of TM. Medicare pays the MA plan a per‐capita rate, based on TM spending in the county, adjusted for patient health risk. If the costs anticipated in the plan's bid are less than this per‐capita payment, plans can use the savings to reduce patient cost sharing or increase the scope of covered benefits to improve their market share.

The risk adjustment method for both TM ACOs and MA plans uses diagnoses recorded in claims or encounter data to compute Hierarchical Condition Category (HCC) scores. ¹ The HCC scoring mechanism is normalized to a score of 1.0 for the average TM beneficiary. Scores deviating from 1.0 represent the beneficiaries' expected spending relative to this average. Each year, the Centers for Medicare and Medicaid Services (CMS) releases software for the model, including diagnosis mapping and model coefficients. The intensity with which diagnoses are recorded will affect the software's computed health risk.

Because coding intensity increases reimbursement for MA plans and TM ACOs, these programs have an incentive to ensure all diagnoses are recorded. They may give providers financial motivation to complete coding, or build technology that facilitates identification of comorbidities during coding. We focus on two coding intensity mechanisms that can be identified in claims and encounter data: the use of chart reviews (CRs) (MA) and health risk assessments (HRAs) (MA and TM) to document diagnoses in the patient's history.

1.1. The HCC model

Diagnoses are recorded in TM claims and MA encounter data using the International Classification of Diseases, Tenth Revision (ICD10) coding system. The 2019 HCC software ² maps ICD10 codes from 2018 patient encounters to diagnosis codes from the Ninth Revision (ICD9), and then maps the ICD9 codes to one of 79 chronic conditions. Some of these chronic conditions are grouped into hierarchies, and indicators of less severe conditions are dropped when a condition of higher severity is identified. For example, if a patient has an ICD10 code of E09.9, this is mapped to “diabetes without complications.” If they also have an ICD10 of E09.40, this is mapped to “diabetes with chronic complications.” When the hierarchies are imposed, only the more severe “diabetes with chronic complications” is retained. Each beneficiary has a vector of 79 indicator variables, summarizing the HCCs assigned to them. While the ICD10 mapping is regularly updated, the same set of 79 HCCs and coefficients has been used since the 2017 software release, and is still used in the 2023 mid‐year update.

Demographic and HCC indicators are multiplied by coefficients to compute the beneficiary's HCC score. The beneficiary's coefficients depend on whether they are eligible for Medicare through disability or old age, whether they are community‐dwelling or institutionalized, and whether they have dual Medicare–Medicaid coverage. The model also includes interaction terms, so some diagnoses have additional leverage through these interactions. The largest number of interactions in the current (v24) model applies to an HCC of congestive heart failure (CHF). CHF interacts with HCCs in the diabetes hierarchy, the lung hierarchy that includes chronic obstructive pulmonary disease (COPD), the heart arrhythmia HCC, and the kidney disease hierarchy. In the extreme scenario, these four interactions would more than triple the impact of CHF alone.

1.2. HCC changes for 2024

CMS published a notice of proposed rulemaking on February 2, 2023, titled “Advance Notice of Methodological Changes for Calendar Year 2024 for Medicare Advantage Capitation Rates and Part C and Part D Payment Policies.” ³ A new HCC risk adjustment model version (v28) is a key component of this notice, and a subsequent publication ⁴ confirmed the new model would be phased in during 2024–2026. The major changes include: First, diagnosis‐to‐condition mapping will be based directly on ICD10 codes rather than through ICD9 codes. This results in 115 HCC indicators rather than 79, adding granularity to the disease classifications. In addition, certain HCCs subject to variation in coding intensity will be eliminated. These include protein‐calorie malnutrition (HCC21 in v24), angina pectoris (HCC88), and atherosclerosis of arteries of the extremities with intermittent claudication (not defined in v24). Finally, the coefficients will be constrained to be equal across the three diabetes HCCs (no complications, HCC17; with chronic complications, HCC18; with acute complications, HCC19), and across three of six heart failure HCCs (not in v24, though v24 includes a broad CHF category, HCC85). CMS estimates that the new model will reduce payments to MA plans by 3.12%, excluding trends in coding practices.

1.3. Mechanisms for increasing coding intensity

MA plans have an opportunity to review medical records to ensure that providers did not accidentally omit a diagnosis from encounter records. These reviews are more important when the providers' reimbursement does not incent detailed coding of the patients' secondary diagnoses. MA plans to make corrections to add or (rarely) delete a diagnosis through CR records. In addition, both MA and TM providers may record additional diagnoses through a HRA during a wellness visit or a home visit for this purpose.

Meyers and Trivedi ⁵ estimated that diagnoses recorded only on CR records increased the average HCC score in MA by 4.1% in 2015. They estimated that this increase translated to $2.3 billion of additional payments to MA plans. They did not examine the impact of HRA records on risk measurement or payments. The Office of the Inspector General estimated a much larger $6.7 billion in additional payments to MA plans in 2017, associated with diagnoses recorded only on CR records. ⁶ In a subsequent report they identified an additional $2.6 billion in 2017 MA payments associated with diagnoses recorded only on HRA records. ⁷ Jung et al found that HRA and CR records increased MA plans' HCC scores by 9%–10% from 2016 to 2019. ⁸ However, they also found that more than half of this increase was associated with increased resource use, suggesting that some of the additional conditions identified by HRA and CR records are actively managed. No study has examined how HRAs change risk scores for TM beneficiaries, although some suggest intensive coding incentives exist for TM ACOs. ⁹

1.4. Focus of this study

We measured how specific mechanisms for coding intensity—HRA and CR records—affect health risk measured by HCC scores, and we identified the hierarchical groups of HCCs that exhibited the greatest impact from coding intensity. We summarized how hierarchical groups contributed to the overall mean HCC score, and how this contribution varied across types of Medicare coverage, including MA plans, TM beneficiaries attributed to ACOs (“TM ACO”), and TM beneficiaries not attributed to an ACO (“TM non‐ACO”). The TM non‐ACO cohort provides an important comparison because they are not exposed to the same coding intensity incentives experienced by MA plans and TM ACOs. The coding intensity incentives are highest for MA plan enrollees, where plans are reimbursed by a fixed per‐capita payment. TM ACO providers have a baseline fee‐for‐service reimbursement, and most of these contracts provide the opportunity to share the savings without imposing downside risk, ¹⁰ which moderates the incentive for coding intensity.

In addition, we estimate the impact of upcoming HCC model changes, to the extent possible with the 2019 software (v22, which uses the v24 coefficients), and identify where the new (v28) model may still be vulnerable to high levels of increased coding intensity through CR and HRA records.

2. METHODS

2.1. Data and computed variables

Our sample was drawn from the CMS 2019 standard 20% sample, excluding those currently eligible for Medicare due to end‐stage renal disease. Because health risk measures are based on lagged data, we used 2018 TM or MA coverage to select the sample. To increase the comparability of diagnosis data in TM and MA, we restricted the sample to beneficiaries who had continuous 2018 coverage in Medicare Parts A, B, and D, and to those with continuous enrollment in either TM or MA during 2018. We further subdivided TM beneficiaries into those attributed at year‐end 2018 to an ACO (Shared Savings Plan or Next Generation), and those who were not. Thus, we had three coverage cohorts: MA, TM, ACO and TM non‐ACO. The data used to identify diagnoses include 2018 inpatient facility, outpatient facility, and carrier files.

Following the methods used by Reid, Le, Seitz, Chiarenzelli, ⁷ we identified two types of HRA records from the carrier files. Any patient encounter for an annual wellness visit (Healthcare Common Procedure Coding System [HCPCS] codes G0438, G0439) or initial preventive physical exam (HCPCS code G0402) was considered an HRA record. Similarly, a home visit (HCPCS codes 99341–99345, 99347–99350, with place of service code 04, 12, 14, 16, or 33) was considered an HRA record, but only if there was only one HRA visit of any type during the year. These HRA records are identified for both TM and MA patient visits. CR records exist only in MA encounter data.

We used the CMS software for 2019 ² to compute HCC indicators and HCC scores from diagnoses in 2018 claims and encounter data. While the HCC algorithm is calibrated to have an average score of 1.0 for the 2019 TM population, our selection criteria—in particular the Part D requirement—led us to expect the TM study sample would have an average risk score higher than 1.0. We computed HCC indicators and scores for baseline records only, and then computed incremental changes in HCC prevalence and scores as we added HRA and CR records. The incremental change can include the removal of an HCC indicator if a new record shifts a patient from that HCC to one of greater severity within a hierarchy.

We also obtained characteristics of the patients and their neighborhoods for propensity score matching (discussed below). The CMS beneficiary file supplied beneficiary age, sex, race and ethnicity, dual Medicare‐Medicaid status, and an indicator of death in 2019. In addition to the demographic information available in the CMS files, we obtained beneficiary neighborhood variables by linking beneficiary ZIP codes to the American Community Survey 5‐year average data (2015–2019) ¹¹ and Rural–Urban Commuting Area codes, ¹² and beneficiary county to the Area Health Resource File. ¹³ From the American Community Survey, we obtained the percentage of adults in the ZIP code with a 4‐year college degree, the percentage of households living under the Federal Poverty Limit, and the percentage age 5 or older speaking English only. Rural ZIP codes were identified using Rural–Urban Commuting Area codes. We used the Area Health Resource File data to identify health care resources in the county including hospital beds per 1000 residents, physicians per 1000, and skilled nursing facility beds per 1000.

2.2. Statistical methods

We computed average HCC scores and prevalence of HCC indicators for the unmatched and propensity‐score matched samples. We also computed the contribution of each hierarchical group to the overall mean HCC scores, and subdivided the impact of HRA and CR records on HCC scores by hierarchical diagnosis group within the model. For this decomposition, we identified the change in prevalence of each HCC indicator due to the added records and used the coefficients specific to the individual's disability and dual Medicaid status to compute incremental changes. The CMS beneficiary data did not allow us to identify institutional status, so we used community coefficients for all individuals. Thus, our computed impact does not reflect the structure of the institutional coefficients.

2.3. Propensity score matching

To facilitate cross‐cohort comparisons, we developed propensity‐score matched samples of MA enrollees, TM ACO, and TM non‐ACO beneficiaries. Because many policy applications may require broader samples than our propensity‐score matched cohorts, we provide full results for both matched and unmatched samples in Tables S‐3a and S‐3b.

We computed the propensity score as the logistic probability of being enrolled in an MA plan, using all patient and neighborhood characteristics described above and the lagged HCC score from base records only. We matched beneficiaries within counties, using the nearest neighbor method without replacement to find a TM ACO match and then a TM non‐ACO match for that MA enrollee. We used a 0.01 caliper (less than 1/10th of the propensity score's standard deviation) for the matching process. We defined the matched sample as triads with county‐level matches between an MA enrollee and both a TM ACO and a TM non‐ACO beneficiary. Matched triads were identified for 24% of the MA enrollees.

3. RESULTS

Table 1 summarizes beneficiary and neighborhood characteristics for the unmatched and matched samples. Before matching, the TM ACO cohort was older and had the highest fraction who were white, college‐educated and spoke English only. The MA cohort had the highest level of racial and ethnic diversity. The TM non‐ACO cohort enrolled the highest fraction of beneficiaries with dual‐eligible status. Propensity score matching improves the balance across the coverage cohorts, although the TM ACO group still has lower rates of dual Medicare–Medicaid eligibility. Tables S‐1a through S‐1c summarize the CMS 20% sample before the requirement of continuous Parts A, B, and D coverage is imposed, and restate Table 1 results for the unmatched and matched samples, including statistical tests of differences across cohorts.

TABLE 1.

Characteristics of beneficiary/enrollee sample and their neighborhoods.

	Unmatched sample with continuous A/B/D in diagnosis year			Matched sample with continuous A/B/D in diagnosis year
	MA plans	TM ACO	TM non‐ACO	MA plans	TM ACO	TM non‐ACO
Sample size	3,937,626	1,599,943	2,731,009	892,156	892,156	892,156
Beneficiary characteristics
Beneficiary age end of payment year (mean [st dev])	73.3 [9.9]	73.4 [11.1]	71.1 [12.8]	73.4 [10.9]	73.8 [11.0]	72.1 [12.4]
Female (%)	57.0%	58.8%	56.8%	58.2%	59.2%	57.8%
Distribution by race (%)
White, non‐Hispanic	68.1%	82.9%	77.7%	82.4%	84.7%	81.5%
Black, non‐Hispanic	12.0%	7.7%	9.2%	8.9%	6.8%	8.9%
Other race, non‐Hispanic	6.1%	4.9%	6.4%	5.4%	5.2%	5.6%
Hispanic	13.8%	4.4%	6.6%	3.3%	3.4%	4.0%
Beneficiary died during the year (%)	3.6%	3.9%	4.3%	4.4%	4.0%	4.5%
Zip code characteristics
ACS: Percent population age 25+ with 4‐year degree or higher (mean)	30.2	33.6	30.9	34.6	34.8	34.0
ACS: Percent population age 5+ speaking English only (mean)	79.3	84.3	83.0	84.1	84.0	83.6
ACS: Percent households under the Federal Poverty Limit (mean)	13.5	11.9	13.4	11.8	11.6	12.1
RUCA: Rural area (%)	16.5%	21.6%	30.2%	20.9%	20.2%	20.3%
County characteristics
AHRF: Hospital beds per 1000 population in county (mean)	2.98	3.06	2.98	3.03	3.03	3.02
AHRF: MDs per 1000 population in county (mean)	3.50	3.49	3.18	3.55	3.55	3.55
AHRF: SNF beds per 1000 population in county (mean)	5.15	6.05	5.84	5.93	5.92	5.92
HCC score category (%)
Non‐dual aged	74.3%	76.5%	65.2%	68.7%	77.1%	70.2%
Non‐dual disabled	5.8%	3.1%	4.2%	6.6%	3.0%	3.9%
Dual partial benefits aged	5.3%	3.1%	4.0%	6.4%	3.1%	3.4%
Dual partial benefits disabled	2.2%	2.2%	3.4%	3.2%	2.1%	2.8%
Dual full benefits aged	8.9%	8.5%	13.0%	10.1%	8.6%	11.0%
Dual full benefits disabled	3.6%	6.5%	10.2%	4.9%	6.2%	8.8%
New enrollee

Abbreviations: ACO, Accountable Care Organization; ACS, American Community Survey 5‐year average (2015–2019); AHRF, 2019 Area Health Resource File; HCC, Hierarchical Condition Category; MA, Medicare Advantage; RUCA, Rural–Urban Commuting Areas; TM, Traditional Medicare.

3.1. Coding intensity and HCC scores

Table 2 summarizes the average 2019 HCC scores by type of coverage for the unmatched and matched samples. The TM ACO cohort had the highest average risk score (1.305) in the unmatched sample, using only base records to compute HCC scores. The average base risk score for TM non‐ACO beneficiaries (1.248) was slightly higher than that for MA plan enrollees (1.219). TM HCC scores were above 1.0 due to our requirement that the sample have Part D coverage. Without the restriction to continuous Parts A, B, and D coverage, our computed average TM HCC score was 1.04 (Table S‐1a). After propensity score matching, average HCC scores from base records were similar for MA (1.290) and TM ACO (1.298) cohorts, with risk measures indicating better health in the TM non‐ACO cohort (1.229). This relationship could indicate more intense coding in base records in the programs where the incentive exists, or it could reflect differences in underlying health risk not accounted for in the matching process.

TABLE 2.

Impact of coding intensity on 2019 average Hierarchical Condition Category (HCC) scores.

Medicare coverage cohort a	2019 Incremental HCC scores			Total score from all records
	From base records	From HRA records	From CR records
Unmatched population with continuous Part A/B/D coverage in 2018
MA cohort	1.219	0.042 (+3.5%)	0.070 (+5.8%)	1.306 (+9.3%)
TM ACO cohort	1.305	0.016 (+1.2%)		1.311 (+1.2%)
TM non‐ACO cohort	1.248	0.010 (+0.8%)		1.231 (+0.8%)
Propensity‐score matched population with continuous Part A/B/D coverage in 2018
MA cohort	1.290	0.046 (+3.6%)	0.080 (+6.3%)	1.409 (+9.8%)
TM ACO cohort	1.298	0.016 (+1.2%)		1.312 (+1.2%)
TM non‐ACO cohort	1.229	0.011 (+0.9%)		1.242 (+0.9%)

Note: Readers who understand the calibration of the HCC methodology may be surprised that our average total score for the TM cohorts are greater than 1.2 rather than averaging 1.0. This difference is driven by the coverage requirements we imposed, including the requirement that the sample has full Part D coverage. Averages for the sample without coverage restrictions can be seen in Table S‐1a.

Abbreviations: ACO, Accountable Care Organization; CR, chart review; HRA, health risk assessment; MA, Medicare Advantage; TM, Traditional Medicare.

^a Coverage cohorts are defined by 2018 enrollment, so that prior‐year diagnoses used in the 2019 HCC calculations come from a consistent coverage source.

In the matched sample, incremental health risk due to HRA records increased in line with incentives to maximize risk scores: +0.9% for TM non‐ACO, +1.2% for TM ACO, and +3.6% for MA. This was very similar to the impact of HRAs in the unmatched sample. When both HRA and CR records are included, the matched MA plan enrollees saw a 9.8% increase in average HCC scores relative to the score from base records. This was slightly higher than the 9.3% increase in the unmatched MA cohort.

In Table S‐2, we break out the matched 2018 enrollment cohorts into subsets defined by 2019 enrollment decisions. Total scores within the cohort have the expected slope, increasing in magnitude as 2019 enrollment goes from MA to TM ACO to TM non‐ACO. However, a greater proportion of the final health risk measure is captured through base records for MA enrollees that choose to switch to TM coverages: the increment due to CR records and HRAs declines from +9.9% for those staying in MA to +3.5% for those switching to TM non‐ACO.

3.2. Mean HCC scores by hierarchical group across cohorts

Fifty of the 79 HCCs are clustered into hierarchies. This results in 45 hierarchical groups, each comprising 1–8 HCCs. For example, three diabetes HCCs form one hierarchical group. Table 3 compares the relative contribution of each hierarchical group to the mean HCC score from each cohort (MA, TM ACO, TM non‐ACO), using the matched sample to facilitate cross‐cohort comparisons. Table 3 includes the 10 hierarchical groups with the highest contribution to the total score. The ordering of hierarchical groups is consistent across cohorts with one major exception: neoplasms, ranked 5th in MA plans, is among the top 2 hierarchical groups in both TM cohorts.

TABLE 3.

Mean Hierarchical Condition Category (HCC) scores by hierarchical group by plan type (matched cohorts).

Hierarchical group indicators a	Hierarchical group description	MA cohort		TM ACO cohort		TM non‐ACO cohort
		Contribution rank	% Contribution to total HCC score b	Contribution rank	% Contribution to total HCC score b	Contribution rank	% Contribution to total HCC score b
HCC106‐HCC108, HCC161, HCC189	Vascular	1	7.49%	2	6.47%	1	6.32%
HCC85	Congestive heart failure	2	6.92%	3	6.32%	3	5.96%
HCC17‐HCC19	Diabetes	3	6.55%	4	6.20%	4	5.70%
HCC110‐HCC112	Lung1	4	6.01%	5	5.25%	5	5.06%
HCC8‐HCC12	Neoplasms	5	5.63%	1	6.90%	2	6.14%
HCC57‐HCC58	Psych	6	4.36%	7	3.63%	6	3.97%
HCC96	Specified heart arrhythmias	7	3.90%	6	4.35%	7	3.85%
HCC134‐HCC137	Kidney	8	3.04%	8	3.01%	8	3.15%
HCC40	Rheumatoid arthritis and inflammatory connective tissue disease	9	2.44%	9	2.57%	9	2.33%
HCC82‐84	Arrest	10	2.31%	10	2.12%	10	2.19%

Note: Detailed descriptors of hierarchical groups: Arrest group includes: Respirator Dependence/Tracheostomy Status, Respiratory Arrest, Cardio‐Respiratory Failure and Shock. Congestive Heart Failure group is a single HCC. Diabetes group includes: Diabetes with Acute Complications, Diabetes with Chronic Complications, Diabetes without Complication. Kidney group includes: Dialysis Status, Acute Renal Failure, Chronic Kidney Disease Stage Five, Chronic Kidney Disease Stage Four. Lung1 group includes: Cystic Fibrosis, Chronic Obstructive Pulmonary Disease, Fibrosis of Lung and Other Chronic Lung Disorders. Neoplasms group includes: Metastatic Cancer and Acute Leukemia, Lung and Other Severe Cancers, Lymphoma and Other Cancers, Colorectal Bladder and Other Cancers, Breast Prostate and Other Cancers and Tumors. Psych group includes: Schizophrenia, Major Depressive Bipolar and Paranoid Disorders. Rheumatoid Arthritis and Inflammatory Connective Tissue Disease is a single HCC. Specified Heart Arrhythmias group is a single HCC. Vascular group includes: Atherosclerosis of the Extremities with Ulceration or Gangrene, Vascular Disease with Complications, Vascular Disease, Chronic Ulcer of Skin Except Pressure, Amputation Status Lower Limb/Amputation Complications.

Abbreviations: ACO, Accountable Care Organization; CR, chart review; CVD, cardiovascular disease; HRA, health risk assessment; MA, Medicare Advantage; TM, Traditional Medicare.

^a HCC variable names and hierarchical groups from the current (v24) HCC Model.

^b Computed as ((mean HCC score) − (mean HCC score computed without the HCC group))/(mean HCC score), using all claims or encounter records.

The CHF group (HCC85) is associated with 6.9% of the mean MA HCC score. As noted earlier, a CHF diagnosis has a large contribution to the overall health risk measure due to interactions with the diabetes, lung, heart arrhythmia, and kidney hierarchies. These CHF interactions compose 46% of the CHF contribution to the MA mean HCC score (3.2% of 6.9%); both TM cohorts had the same proportion of the CHF contribution due to interaction terms (46%). Across the whole sample, 84% of those with a CHF diagnosis also had at least one diagnosis forming a CHF interaction term.

The concentration of the top‐10 groups' contribution to the mean HCC score declines with incentives for coding intensity: the top‐10 groups compose 48.7% of the mean HCC score for the MA cohort, 45.8% of the score for the TM ACO cohort, and 44.7% of the score for the TM non‐ACO cohort.

Tables S‐3a (matched sample) and S‐3b (unmatched sample), provide the weights for the complete list of 45 hierarchical groups.

3.3. Mean HCC scores and impact of coding intensity by hierarchical group for MA plans

To study the sensitivity of mean HCC scores to HRA and CR records in greater detail, we focus on the MA cohort, using the unmatched sample to facilitate the comparison of our results with previous findings. The rightmost columns of Table 4 restate from Table 3 the contribution of selected hierarchical groups to the mean HCC score for the unmatched MA cohort. The leftmost columns of Table 4 summarize the groups' sensitivity to coding intensity through added diagnoses from HRA and CR records. In Table 4, we selected the eight hierarchical groups most sensitive to coding intensity, plus two additional groups related to the upcoming 2024 model changes. The complete results for all 45 hierarchical groups are available in Tables S‐3a (matched sample) and S‐3b (unmatched sample).

TABLE 4.

Mean Hierarchical Condition Category (HCC) scores and impact of coding intensity by hierarchical group for MA plans and estimated model change effects (unmatched cohort).

Hierarchical group indicators a	Hierarchical group description	Coding sensitivity rank	% Increase to base HCC score due to HRA/CR records b	Contribution rank	% Contribution to total HCC score c
Eight hierarchical groups most sensitive to coding intensity
HCC106‐HCC108, HCC161, HCC189	Vascular	1	1.47%	1	7.75%
HCC57‐HCC58	Psych	2	1.01%	6	4.67%
HCC85	Congestive heart failure	3	0.94%	3	6.54%
HCC110‐HCC112	Lung1	4	0.90%	4	5.86%
HCC17‐HCC19	Diabetes	5	0.68%	2	7.50%
HCC22	Morbid obesity	6	0.61%	10	2.32%
HCC40	Rheumatoid arthritis and inflammatory connective tissue disease	7	0.44%	9	2.54%
HCC54‐HCC55	Substance abuse	8	0.34%	14	1.70%
Additional groups affected by upcoming model changes
HCC86‐HCC88	Heart (includes angina pectoris)	15	0.15%	15	1.03%
HCC21	Protein‐calorie malnutrition	16	0.15%	18	0.82%

Note: Detailed descriptors of hierarchical groups: Congestive Heart Failure group is a single HCC. Diabetes group includes: Diabetes with Acute Complications, Diabetes with Chronic Complications, Diabetes without Complication. Heart group includes: Acute Myocardial Infarction, Unstable Angina and Other Acute Ischemic Heart Disease, Angina Pectoris. Lung1 group includes: Cystic Fibrosis, Chronic Obstructive Pulmonary Disease, Fibrosis of Lung and Other Chronic Lung Disorders. Morbid Obesity group is a single HCC. Protein‐Calorie Malnutrition is a single HCC. Psych group includes: Schizophrenia, Major Depressive Bipolar and Paranoid Disorders. Rheumatoid Arthritis and Inflammatory Connective Tissue Disease is a single HCC. Substance Abuse group includes: Drug/Alcohol Psychosis, Drug/Alcohol Dependence. Vascular group includes: Atherosclerosis of the Extremities with Ulceration or Gangrene, Vascular Disease with Complications, Vascular Disease, Chronic Ulcer of Skin Except Pressure, Amputation Status Lower Limb/Amputation Complications.

Abbreviations: ACO, Accountable Care Organization; CR, chart review; HRA, health risk assessment; TM, Traditional Medicare.

^a HCC variable names and hierarchical groups from the current (v24) HCC Model.

^b Computed as ((mean HCC score with HRA/CR increase in this group only)‐(mean HCC score base records))/(mean HCC score base records).

^c Computed as ((mean HCC score) − (mean HCC score computed without the HCC group))/(mean HCC score), using all claims or encounter records.

The eight most sensitive hierarchical groups were responsible for two‐thirds of the total increase in mean HCC scores due to diagnoses added by HRA or CR records (6.4% of 9.3%) in the unmatched MA cohort. The top three groups include five HCCs in the vascular group (1.5% increase), two HCCs in the psych group (1.0% increase), and the CHF HCC (0.9% increase). Thirty‐nine percent of the CHF‐related increase (0.4% of 0.9%) was associated with CHF interaction terms.

The incremental changes in measured risk due to HRAs and CRs are driven by shifts in HCC prevalence, multiplied by the model coefficients. Readers interested in more detail should turn to Table S‐4 to see detailed changes in the prevalence of the 79 individual HCCs due to new diagnoses identified in HRA and CR records. These changes include both new diagnoses and shifts from lower to higher severity levels within a hierarchical group. We list each HCC's Community, Non‐dual, Aged coefficient to help the reader understand the relative weight of each HCC. We note where interaction terms increase the impact of changes beyond the main‐effect coefficients.

3.4. Impact of upcoming model changes

CMS estimated that, absent trends in coding practices, the upcoming model changes will reduce MA plan HCC scores by 3.1%, on average. ³ Their advance notice details changes in the model structure. Using the 2019 software, we were able to estimate the impact of three of the upcoming changes, including the compression of the three diabetes HCCs to a single coefficient, which we modeled as a weighted average of the current (v24) coefficients, using the TM prevalence rates for HCC17‐HCC19 to compute the weighted average; elimination of HCC21, protein‐calorie malnutrition; and elimination of HCC88, angina pectoris.

The impact of these three changes on the mean HCC scores is summarized in Table 5 for the matched cohorts. We estimated that these changes together would reduce the average MA HCC score by 2.0%. As expected, the diabetes and angina pectoris changes have a smaller impact on the TM cohorts than the MA cohort. However, the elimination of protein‐calorie malnutrition did not cause the expected monotonic decline from MA to TM non‐ACO cohorts. We were unable to estimate the impact of compressing coefficients for three of the six HCCs in a new hierarchical group for heart failure, or the elimination of a new HCC for atherosclerosis of arteries of the lower extremities with intermittent claudication, not defined by v24 of the model.

TABLE 5.

Estimated model change effects (matched cohort).

Upcoming model changes	Impact on total HCC score (%)
MA cohort
Use same coefficients for diabetes (HCC17‐HCC19) a	−0.51
Drop protein‐calorie malnutrition (HCC21) b	−1.09
Drop Angina Pectoris (HCC88) b	−0.43
TM ACO cohort
Use same coefficients for diabetes (HCC17‐HCC19) a	−0.19
Drop protein‐calorie malnutrition (HCC21) b	−0.88
Drop Angina Pectoris (HCC88) b	−0.34
TM non‐ACO cohort
Use same coefficients for diabetes (HCC17‐HCC19) a	−0.01
Drop protein‐calorie malnutrition (HCC21) b	−1.01
Drop angina pectoris (HCC88) b	−0.29

Abbreviations: ACO, Accountable Care Organization; HCC, Hierarchical Condition Category.

^a Computed as the difference between current individual versus weighted average coefficients for diabetes, where average coefficient is weighted by prevalence in the standard 20% sample of TM beneficiaries, where all claims and encounter records are used.

^b Computed as ((mean HCC score) − (mean HCC score computed without the HCC group))/(mean HCC score), where all claims or encounter records are used.

4. DISCUSSION

We increase understanding of coding intensity on health risk measures in Medicare by computing the impact of HRA and CR records on HCC scores in MA plans, and demonstrate that HRA records have a meaningful impact on health risk measurement in TM coverage cohorts as well. MA plans have a much higher rate of identifying new diagnoses in HRAs (3.6% increase in mean HCC score) than either TM ACOs (+1.2%) or TM non‐ACOs (+0.9%). When HRA and CR records are included for the matched MA coverage cohort, we estimate a 9.8% increase (3.6% HRA + 6.3% CR with rounding error) in average HCC scores. This is higher than the 4.1% CR‐only increase identified by Meyers and Trivedi in 2015, ⁵ but recent work documents that the impact of adding diagnoses through HRA and CR records is increasing over time. ⁸

By decomposing the HCC score into the contribution from each hierarchical group, we are the first to identify the condition categories that have the greatest contribution to health risk measurement. We are also the first to decompose coding sensitivity to HRA and CR records by hierarchical group, to identify potential focus for future model revisions. We identified eight hierarchical groups that compose 6.4 of the 9.3 percentage‐point increase (69%) due to coding intensity in the unmatched MA cohort.

4.1. Study limitations

Our study is not without limitations. We measured increases in coding intensity resulting from provider and plan use of HRAs and CR records, but we didn't examine other sources. Patient choices about Medicare coverage and in accessing care can affect both the true risk mix and the intensity of diagnostic coding. MA plans and TM coverage paired with private Medicare supplement plans tend to have lower cost sharing than basic TM coverage, often in exchange for an additional premium from the beneficiary. This tends to attract sicker beneficiaries, who are willing to pay a premium to reduce financial barriers at the point of care. On the other hand, in exchange for lower copayments or lower premiums, MA enrollees may agree to receive their care from tighter networks of providers. Restrictions on provider choice tend to attract healthier beneficiaries, who worry less about freedom of provider choice. The net impact of these two forces on true health risk is ambiguous. In addition, lower cost sharing means beneficiaries may increase the number of points of contact with the health care system, increasing the probability that more diagnoses will be recorded and measured health risks will increase.

It is also important to realize that, while a TM ACO or MA plan may have total‐cost‐of‐care incentives, these incentives may not trickle down to individual providers. Providers paid by capitation or salary may have little incentive for complete and accurate diagnosis coding unless bonuses or other incentives drive coding intensity.

Finally, while propensity score matching markedly improved the balance of observable patient and neighborhood characteristics, and our process of matching within counties assured balance in geographic influences, meaningful unobserved confounders may remain.

4.2. Policy implications

The upcoming changes released by CMS in February 2023 were estimated to reduce HCC scores by 3.1% for MA plans. ³ Our partial replication of this analysis identified a 2.0% reduction in mean HCC scores, consistent with CMS's findings. However, we note that the changes to the HCC methodology address only three of the eight hierarchical groups (vascular disease, heart failure, and diabetes) identified as highly sensitive to coding intensity.

In their commentary on the changes, ¹⁴ the Medicare Payment Advisory Committee (MedPAC) examined the rate of coding intensity in HCCs impacted by the model changes. MedPAC measured coding intensity for a specific HCC as a relative rate: (MA prevalence rate)/(TM prevalence rate). They computed these relative rates by MA plan decile. It is clear from their analysis that the top two deciles of MA plans are driving much of the coding intensity within these HCCs. However, model changes in response to HCCs where these relative rates are high may not have a meaningful impact on MA plan payments. If the overall prevalence of a condition is low, or the model coefficient is relatively small, the bottom‐line financial impact of a particular hierarchical group may be low, even if MA coding rates relative to TM are high. Our analysis identifies the hierarchical groups where coding intensity has the largest bottom‐line impact.

Importantly, none of the analyses of the upcoming changesby our research group, MedPAC ¹⁴ or CMS ³ —incorporates behavioral changes based on the new hierarchical groups. CMS did estimate that the impact of coding trends would more than offset the savings due to the new model, resulting in an increase of about 1% in net MA plan payments. ¹⁴ However, this was a projection of historic trends and did not include behavioral shifts induced by model changes. For example, the use of a single coefficient across all severities of diabetes reduces the incentive to record diagnoses that shift a patient from “diabetes without complications” to “diabetes with chronic complications.” However, the increased coefficient at the lowest severity level dramatically increases the incentive to identify individuals who have diabetes without complications. In the matched MA sample, 7.3% had a diagnosis of low‐severity diabetes in base records, with an additional 1.1% (+14.9%) added by HRA and CR records (Table S‐4). Clearly, the potential exists to search more intensively for patients with diabetes not yet documented. Similarly, the elimination of angina pectoris from the model may incent more frequent documentation of the next higher diagnosis in the hierarchical group, unstable angina, and other acute ischemic heart disease. Finally, the increase from 79 to 115 HCCs subdivides some v24 HCCs into multiple v28 HCCs. This increased granularity provides incentives for providers and plans to increase the intensity of coding by moving individuals to more severe HCCs within a hierarchical group. Close monitoring of these behavioral shifts and increased attention to the hierarchical groups most sensitive to coding intensity are warranted as CMS continues to refine their risk adjustment model.

Supporting information

Appendix S1. Supporting information.

Carlin CS, Feldman R, Jung J. The mechanics of risk adjustment and incentives for coding intensity in Medicare. Health Serv Res. 2024;59(3):e14272. doi: 10.1111/1475-6773.14272