Promoting Health Equity In Population-based Payment Is Not A Matter of Predictive Accuracy in Risk Adjustment: Concepts And Evidence

Abstract

The objective of risk adjustment is not to predict spending accurately but to support the social goals of the payment system, which include equity. Setting population-based payments at accurate predictions risks entrenching spending levels that are insufficient to mitigate the impact of social determinants on health care use and effectiveness. Instead, payments must be set above current levels of spending for historically disadvantaged groups to advance equity. In analyses intended to guide such reallocations, we find that current risk adjustment for the community-dwelling Medicare population overpredicts annual spending for Black and Hispanic beneficiaries by $376–1,264. Their risk-adjusted spending is lower than spending for non-Hispanic White beneficiaries despite worse risk-adjusted health and functional status. Thus, continued movement from fee-for-service to population-based payment models that omit race and ethnicity from risk adjustment should result in sizable resource reallocations and incentives that support efforts to address racial and ethnic disparities in care. We find much smaller overpredictions for beneficiaries without a high school degree and for communities with higher proportions of residents who are Black, Hispanic, or have less education, suggesting additional payment adjustments that depart from predictive accuracy are needed.

In contrast to a fee-for-service (FFS) system in which spending is distributed based on service use, a population-based payment system distributes spending based on population characteristics. Accordingly, population-based payment models, as in the Medicare Shared Savings Program (SSP) or Medicare Advantage (MA) program, can facilitate resource reallocations necessary to address health care disparities. Risk adjustment is the mechanism by which payment is allocated.

Traditionally, risk adjustment has been conceived and executed purely as a predictive exercise. Regression is used to predict total annual per-person spending as a function of demographic and clinical characteristics. An individual’s predicted spending is converted to a risk score, which is applied to a base regional rate to determine the prospective payment or benchmark for that individual. The more accurately spending is predicted (i.e., the better the fit of the regression model), the more closely payment matches spending (a proxy for costs), thereby equalizing financial risk across providers or plans serving different populations and limiting incentives to attract favorable risks (patients with overpredicted spending) or avoid unfavorable risks (patients with underpredicted spending).

A commonly voiced concern with the transition to population-based payment is that risk adjustment will fail to account for historically marginalized groups’ presumed higher spending, thereby exacerbating health disparities. Framed as solving a prediction problem, social risk adjustment is thus often thought to achieve its goal by adding social risk factors as predictors to standard risk-adjustment models. Various studies have considered the incremental predictiveness of measurable social risk factors and made recommendations about which to add, but this line of research has focused largely on outcome measures as opposed to spending.^1,2 Implicit in many calls for “improved” risk adjustment is an assumption that social risk factors predict higher spending – that the problem is their omission from predictive models and thus that equity-promoting reallocations can be motivated by predictive accuracy.

However, attempting to support more equitable care by improving the predictive accuracy of risk adjustment is a fundamentally limited strategy because current levels of spending (the target of prediction) are unlikely to be the desired levels of spending for those populations. Individuals who experience social disadvantage may use less health care and have lower spending than others with the same clinical needs.^3–5 For example, they may have less income to spend on health care, have less generous insurance coverage, be less aware of their health care needs because of lower educational attainment, face greater barriers to accessing care (e.g., travel and time constraints), or encounter additional barriers from other manifestations of structural or interpersonal racism. Inclusion of markers of social disadvantage in risk adjustment models may therefore improve predictive accuracy but lower payments for underserved populations relative to models that omit them.

Moreover, current spending for historically marginalized groups may be too low to support equitable care because providers serving those groups may have insufficient resources to improve quality of care delivery or provide the additional supportive services (e.g., case management and outreach) necessary to mitigate the adverse impact of social determinants on health care use and effectiveness.² Many supportive services are not reflected in FFS spending.

Thus, even if the addition of some social factors to standard risk adjustment results in higher population-based payments for populations with higher prevalence of those factors, the adjustments merely recover spending levels under FFS that are believed to be too low to cover the costs of reducing disparities. To address disparities, payment must instead be set above current spending (or an accurate prediction thereof).²

More generally, the objective of risk adjustment is not solely to predict observed spending accurately. Rather, the goal is to support the broader social goals of the payment system – to make the health care system more efficient and equitable.^6,7 If the motivation for payment reform is to spend more efficiently and equitably, then current spending is inherently the wrong target for population-based payments. A reformed payment system should encourage the desired level and distribution of spending, not entrench the status quo.

That risk adjustment presents tradeoffs between fit and other objectives has been well-described. Improving fit inherently weakens the power of incentives in a population-based payment system.⁶ As payments (or benchmarks) are adjusted for more markers of health care use (e.g., diagnoses) or for utilization directly (e.g., lagged indicators of hospitalization), risk-bearing plans or providers save less from curbing unnecessary or avoidable care (reducing use reduces payment). In the extreme, adjusting for use of each service would achieve perfect fit (R²=1.0) but revert payment incentives to FFS. To some extent, we must tolerate incentives encouraging risk selection (deficient fit) to allow the payment system to control spending.

Likewise, advancing the goal of health equity requires payment adjustments that diverge from predictive accuracy. Setting population-based payments (or benchmarks) above an accurate prediction of FFS spending for historically disadvantaged groups worsens fit but mitigates resource disparities that contribute to health disparities and better aligns payment with health care needs (including unmet needs). Deliberately paying above current spending for those groups also protects socially vulnerable patients with underpredicted clinical risk against risk selection and creates incentives for competing providers or plans to attract the underserved with enhanced benefits or services. Several approaches have been developed to set population-based payments at desired rather than accurately predicted levels.⁷

Yet, concerns about inadequate accounting for social determinants in population-based payment models remain largely framed around the predictive accuracy of standard risk-adjustment methods, often appealing to the promise of advanced prediction tools, such as machine learning and artificial intelligence, in proposed solutions.^8–11

To inform payment policy intended to support health equity, in this study we first add individual-level predictors of social disadvantage (race, ethnicity, and educational attainment) to the hierarchical condition categories (HCC) model currently used to risk-adjust payments in MA and benchmarks in the SSP. The results describe existing underpredictions or overpredictions of spending by race, ethnicity, and education. Second, we calculate the associated reallocations across groups achieved by moving to a population-based payment system under current risk-adjusted methods (which omit these social characteristics as predictors). These reallocations equivalently describe the incentives for a risk-bearing entity to attract individuals with these characteristics. Third, we compare the HCC-adjusted differences in spending between groups with HCC-adjusted differences in self-reported health status, functional limitations, and access to care to gauge the extent to which reallocations under current risk adjustment are commensurate with addressing evident disparities. Fourth, we compare results when using area-level, instead of individual-level, versions of the same predictors. Finally, we consider the implications of our findings for the targeting and implementation of population-based payment adjustments that depart from predictive accuracy to support health equity.

STUDY DATA AND METHODS

Study Population and Data

We analyzed Medicare claims from 2012–2017 for 20% annual random samples of FFS beneficiaries and for respondents to the 2012–2017 FFS Medicare Consumer Assessment of Healthcare Providers and Systems (CAHPS) surveys. The FFS Medicare CAHPS survey, administered annually to a national cross-sectional sample of FFS beneficiaries, assesses patient experiences with care and collects sociodemographic and health information not available in Medicare administrative data files. The average CAHPS response rate among beneficiaries meeting inclusion criteria was 42.2% over our study period. We limited each annual 20% or CAHPS sample to beneficiaries continuously enrolled in Parts A and B of FFS Medicare both in the study year (while alive for decedents) and preceding year (to collect diagnoses for the prospective HCC model). For consistency across samples, we excluded long-term nursing facility residents and beneficiaries with end-stage renal disease from our main analyses, as their numbers are small among the largely community-dwelling CAHPS respondents. We conducted separate analyses of these groups for outcomes available for the 20% sample.

Study Variables

We calculated total annual Medicare spending per beneficiary by summing payments across all services reimbursed by Part A or B. From survey data for CAHPS respondents, we assessed indicators of compromised health or access to care: 1) fair/poor general health status; 2) fair/poor mental health; 3) difficulty with one or more activities of daily living (ADL); and 4) difficulty accessing routine, urgent, or specialty care in a timely fashion, as defined by a report of never or sometimes receiving care as soon as needed (versus usually or always).

For both 20% and CAHPS samples, we assessed beneficiaries’ race and ethnicity from the RTI race and ethnicity variable in the Medicare Master Beneficiary Summary File (MBSF).¹² We focused our main analyses on Black, Hispanic, and non-Hispanic White beneficiaries, the RTI variable exhibits stronger concordance with self-reported classification for these groups than others.¹³ For CAHPS respondents, we additionally assessed educational attainment (dichotomized as less than a high school degree vs. high school or more) and self-reported race and ethnicity to explore the validity of estimates derived from the RTI variable.

Using 2015–2019 data from the American Community Survey, we created analogs of these variables at the Census block group level (proportion of residents without a high school degree, proportion Black, and proportion Hispanic). Finally, from the MBSF, we determined beneficiaries’ age, sex, dual eligibility for Medicaid, original reason for Medicare eligibility (aged vs. qualifying disability), county of residence, and 9-digit ZIP code of residence for linking block group level variables.

Statistical Analysis

We fit a linear regression model of total annual per-beneficiary Medicare spending as a function of age, sex, HCC indicators, enrollment segment (aged non-duals, aged duals, disabled non-duals, and disabled duals), interactions between segment and the HCC indicators, and the added predictor of interest (race and ethnicity or education). Whereas typically the HCC model is fit within each segment, this pooled model provided average estimates across community-dwelling segments.

We also included county fixed effects in the model because MA payments and the base rates for the regional component of accountable care organization (ACO) benchmarks are set at the county level (as a function of past FFS spending in the county). HCC risk scores are applied to county base rates to determine payments or benchmarks. Without adjustment for county, we might erroneously conclude, for example, that HCC-adjusted benchmarks would undercompensate ACOs for a group disproportionately living in high-spending counties.

Thus, the model estimates within-county differences in HCC-adjusted FFS spending between groups with different race, ethnicity, or educational attainment. From these differences, we calculated the extent to which the HCC model (which omits these characteristics) over- or under-predicts FFS spending for each group. These over- or under-predictions characterize the selection incentives and payment allocations an ACO operating in a given county would face when serving different groups of beneficiaries under a scenario in which the ACO’s benchmark is based entirely on an HCC-adjusted regional rate, as in MA. (Currently an ACO’s benchmark is based on a blend of a risk-adjusted regional component and the ACO’s historical spending at baseline.) In turn, these estimates tell us how moving from FFS to a population-based payment model would reallocate resources across plans or providers serving different groups. If the HCC model overpredicts spending for historically disadvantaged groups, moving toward population-based payment would increase payment for them (and vice-versa if the model underpredicts spending). Our approach assumes coefficients for HCCs are similar with and without county fixed effects (Appendix Exhibit 1).¹⁴

We then fit the same model to each health status and access to care indicator. These models aided normative interpretation of the results for spending. If, for example, beneficiaries with low education have lower HCC-adjusted spending but worse HCC-adjusted health or access, we can surmise that their lower spending is not commensurate with their health and needs and that higher spending could help address disparities. All analyses used robust variance estimators, clustered at the county level. Analyses of CAHPS data additionally applied survey weights to account for nonresponse.

Limitations

Our study had several limitations. First, our analysis relied on FFS claims. Although our estimates are likely informative for understanding how the HCC model allocates payment and creates selection incentives across groups of MA enrollees, estimates would differ somewhat if based on MA data. For example, spending for beneficiaries with less education may be lower in traditional FFS Medicare in part because they are less likely to have supplemental insurance. We might therefore expect smaller spending differences in MA. In a sensitivity analysis of the FFS CAHPS sample, we additionally adjusted for self-reported sources of supplemental coverage.

Second, while our analysis describes the extent to which the current risk-adjustment system over- or under-predicts FFS spending for historically marginalized groups and communities, it cannot determine the socially optimal level of payment. That depends on social values, the extent of underspending for the underserved, and the extent to which payment increases would be passed through by providers or plans to populations in need – all challenging to know. Nevertheless, our estimates inform where adjustments are needed to increase payment above predictions made by current risk-adjustment methods. For example, a finding that the HCC model predicts spending accurately for a group reporting worse access and health would motivate consideration of an increase in payment above the predicted level for that group. Since the optimal magnitude of such increases cannot be known ex ante, we see no reason for uncertainty to hinder initiation of this direction for payment policy.

Third, the social characteristics we examine were limited to race, ethnicity, and education. We selected these characteristics because they could be ascertained at both individual and block group levels. Although a limited set, these are powerful predictors of disadvantage mediated by a broad range of mechanisms. Moreover, our study is a proof-of-concept analysis that produces an instructive set of varied results and implications across the groups studied. It is not intended to be comprehensive in the predictors examined, as the objective is not to predict better but rather to illustrate conceptual and empirical considerations underlying sound payment policy. Finally, while our analysis can inform payment reallocations to support health equity, it does not assess how additional resources provided under the current or a future system are passed through, as desired, to improve care for specific populations.

RESULTS

Study Population

The sociodemographic characteristics of the 2012–2017 CAHPS samples were similar to those of the 20% samples of community-dwelling Medicare beneficiaries (Exhibit 1).

Exhibit 1.

Characteristics of CAHPS and 20% FFS samples, 2012–2017

Beneficiary Characteristic	CAHPS	20% sample
	(N=512,401 beneficiary-years)	(N=32,721,400 beneficiary-years)
Mean age, y	72.9	72.5
Sex, %
Female	55.0	55.0
Enrollment segment, %
Aged dual	6.2	7.1
Aged non-dual	72.3	68.6
Disabled dual	9.6	11.6
Disabled non-dual	12.0	12.7
RTI race and ethnicity variable, %
American Indian or Alaska Native	0.5	0.5
Asian or Pacific Islander	2.3	2.3
Black	8.5	8.9
Hispanic	4.9	5.4
White, non-Hispanic	82.2	81.4
Other	0.8	0.8
Unknown	0.8	0.7
CAHPS self-reported race and ethnicity, %
American Indian or Alaska Native	1.8	-
Asian or Pacific Islander	2.7	-
Black or African-American	7.7	-
Hispanic or Latino	5.1	-
Multiracial, non-Hispanic	0.4	-
White, non-Hispanic	78.6	-
Unknown	3.6	-
Education, %
No high school degree	13.3	-
High school degree or more	86.7	-
Block-group measures, mean
% without a high school degree	13.9	14.0
% White, non-Hispanic	70.6	70.0
% Black	10.1	10.4
% Hispanic	12.1	12.4
Total annual Medicare spending per beneficiary, $	8,506	9,000
Timely access to needed routine, urgent, or specialty care, %
No problems	83.8	-
1 or more problems	16.2	-
General health, %
Fair/poor	29.3	-
Excellent/very good/good	70.7	-
Mental health, %
Fair/poor	15.1	-
Excellent/very good/good	84.9	-
Activities of daily living, %
No difficulties	62.5	-
1 or more difficulties	37.5	-

Source: Authors’ analysis of enrollment data from the Medicare Beneficiary Summary File from 2012–2017 for FFS CAHPS survey respondents and 20% samples of FFS beneficiaries and of FFS CAHPS survey data

Notes: Descriptive statistics for CAHPS variables are weighted using CAHPS survey weights. The RTI race and ethnicity variable is an enhanced version of the base MBSF race and ethnicity that uses surname and geographic analysis to improve accuracy for Hispanic and Asian populations; classification of Black beneficiaries by both variables is based on self-reported data collected by the Social Security Administration. To support consistent comparisons with the RTI variable classification, all CAHPS respondents who self-identified as Hispanic or Latino in response to an item about Hispanic or Latino descent were categorized as Hispanic or Latino. Those who self-identified as White and one other category of race were assigned to the non-White category. Those reporting two different non-White non-Hispanic categories were classified as multiracial. Estimates from our main analyses for Black and Hispanic beneficiaries were not appreciably changed by alternate categorizations of the CAHPS responses in sensitivity analyses. Activities of daily living include bathing, dressing, eating, getting in or out of chairs, walking, and using the toilet.

Spending

After adjusting for age, sex, enrollment segment, HCCs, and county, total annual Medicare spending per beneficiary was $574 lower for non-Hispanic Black beneficiaries and $1,462 lower for Hispanic beneficiaries than for non-Hispanic White beneficiaries in the 20% samples (Exhibit 2). These estimates suggest substantial overprediction by the HCC model of FFS spending for Black and Hispanic beneficiaries. In turn, population-based payments set by applying HCC risk scores to a county base rate would redistribute payment away from non-Hispanic White beneficiaries (−$198/beneficiary) toward Black (+$376) and Hispanic (+$1264) beneficiaries (Exhibit 3). These payment reallocations equivalently quantify the relative selection incentives an ACO receiving such risk-adjusted population-based payments would face, on average, in a given county; it would have a strong incentive to attract Hispanic and Black residents of that county. Conversely, adding race and ethnicity to the HCC model would lower payments for Black and Hispanic beneficiaries (but would improve the predictive accuracy [fit] of the model).

Exhibit 2.

Medicare spending differences by beneficiary and community-level characteristics after standard risk adjustment

Characteristic	HCC-adjusted difference in per-beneficiary annual Medicare spending, $ (95% CI)
RTI race and ethnicity
White, non-Hispanic	ref
Black	−574 (−691, −456)
Hispanic	−1,462 (−1,609, −1,315)
Education
High school degree or more	ref
No high school degree	−85 (−289, 120)
Block group measures (rescaled to SDs per notes)
Percent without a high school degree	−45 (−68, −23)
Percent Black	−7 (−49, 36)
Percent Hispanic	−115 (−151, −79)

Source: Authors’ analysis of FFS CAHPS survey data, Medicare claims and enrollment data, and American Community Survey data.

Notes: Estimates for individual-level race and ethnicity and block group-level characteristics are from analysis of the 20% FFS samples, while estimates for individual-level education are from analysis of the CAHPS sample. Estimates for each block group measure have been rescaled to reflect a standard deviation change in the measure. For example, an increase of one standard deviation in the block group proportion of residents without a high school degree is associated with a $42 decrease in per-beneficiary Medicare spending.

Exhibit 3. Implied per-beneficiary payment redistribution resulting from transition to population-based payments under standard risk adjustment.

Source: Authors’ analysis of FFS CAHPS survey data and Medicare claims and enrollment data.

Notes: Estimates for individual-level race and ethnicity are from analysis of the 20% FFS samples, while estimates for individual-level education are from analysis of the CAHPS sample. Redistributions reflect the extent to which the HCC model over or underpredicts spending for each group and were calculated as follows. The estimates in Exhibit 2 describe the HCC-adjusted difference in spending between groups. To describe the difference between an HCC-adjusted population-based payment (the average risk adjusted spending for all groups) and spending for a single group, we applied the group population shares (from Exhibit 1) to the estimate in Exhibit 2. For example, beneficiaries without a high school degree have risk-adjusted spending that is $85 lower than those with more education (Exhibit 2). Average risk-adjusted spending is a weighted average of spending for those with more and less education according to the distribution in Exhibit 1 (86.7% have a high school degree or more). Thus, the average is 0.867×$85=$73.7 higher (the estimate in Exhibit 3 within rounding error).

Estimates were less precise but substantively similar in analyses of the CAHPS sample when using the RTI variable (Appendix Exhibit 2),¹⁴ suggesting that analyses of other variables available only for CAHPS respondents also should generalize to the full community-dwelling FFS Medicare population. Within the CAHPS sample, estimates also were substantively similar when assessing race and ethnicity using the RTI variable vs. self-reports (Appendix Exhibit 2).¹⁴ This supplementary analysis additionally revealed that the HCC model substantially overpredicts FFS spending for Asian or Pacific Islander beneficiaries. Estimates for American Indian or Alaska Native beneficiaries varied across samples and data sources, limiting inferences.

In contrast to findings for race and ethnicity, HCC-adjusted spending was minimally lower for beneficiaries with less than a high school degree than for those with more education in the same county (−$85/beneficiary; 95% CI:−$289,$120), implying only a small equity-promoting reallocation ($73) from moving toward population-based payment under the current risk-adjustment system (Exhibits 2–3). Similarly, census block aggregates of race, ethnicity, and education did not predict spending that differed markedly from what the HCC model would predict when applied to a county base rate. As detailed in Exhibit 2, an increase in the proportion of beneficiaries without a high school degree equal to a full standard deviation in the block group distribution predicted $45 lower spending per beneficiary. Similarly scaled estimates for the proportion Black and Hispanic were -$7 and -$115. These findings for area-level predictors suggest that moving from FFS to population-based payment under current risk-adjustment would result in minimal to modest reallocations toward communities with higher proportions of residents of color or with less education and thus give ACOs (or, by extension, MA plans) minimally to modestly stronger incentives to enter, or expand their provider networks in, those communities relative to other communities.

Health Status, Functional Status, and Access to Care

Despite lower HCC-adjusted spending, Black and Hispanic beneficiaries reported significantly worse general and mental health status and more difficulties with ADLs than non-Hispanic White beneficiaries in HCC-adjusted comparisons (Exhibit 4). Disparities in health and functional status were even greater between beneficiaries with vs. without a high school degree despite smaller spending differences. For example, beneficiaries without a high school degree were 10.4 percentage points more likely to report being in fair or poor health (sample mean 29.3%) and 4.3 percentage points more likely to report difficulty with an ADL (mean 37.5%). Black and Hispanic beneficiaries and those with less than a high school degree also all reported worse access to care (Exhibit 4). Findings were mostly similar for Asian or Pacific Islander and American Indian or Alaska Native beneficiaries (Appendix Exhibit 3). Associations between census block variables and health or functional status were mostly similar in direction but smaller in magnitude. Estimates from logistic regression models were similar.

Exhibit 4.

Differences in health status, functional status, and access to care after standard risk adjustment

Characteristic	HCC-adjusted difference in health or access measure, % (95% CI)
	Problem accessing care	Fair or poor general health	Fair or poor mental health	Difficulty with an ADL
RTI race and ethnicity
White, non-Hispanic	ref	ref	ref	ref
Black	5.9 (5.3, 6.5)	3.0 (2.3, 3.7)	2.4 (1.7, 3.1)	1.7 (0.9, 2.5)
Hispanic	7.0 (6.1, 7.9)	7.6 (6.5, 8.6)	5.7 (4.9, 6.6)	2.0 (1.1, 2.9)
Education
High school degree or more	ref	ref	ref	ref
No high school degree	4.3 (3.8, 4.8)	10.4 (9.9, 10.9)	8.4 (7.9, 8.8)	4.3 (3.8, 4.8)
Block group measures
Percent without a high school degree	0.3 (0.1, 0.4)	0.3 (0.2, 0.5)	0.2 (0.0, 0.3)	0.3 (0.1, 0.5)
Percent Black	0.9 (0.7, 1.1)	0.8 (0.6, 1.0)	0.4 (0.3, 0.6)	0.8 (0.6, 1.1)
Percent Hispanic	0.8 (0.6, 1.0)	1.5 (1.2, 1.7)	0.9 (0.7, 1.1)	1.3 (1.0, 1.5)

Source: Authors’ analysis of FFS CAHPS survey data, Medicare claims and enrollment data, and American Community Survey data.

Notes: All results are derived from analysis of CAHPS sample. Estimates for each block group measure have been rescaled to reflect a standard deviation change in the measure.

Supplementary Analyses

Spending estimates by race and ethnicity at the individual level were directionally similar across each community-dwelling population segment when analyzed separately but differed for Black beneficiaries in the end-stage renal disease population and for both Black and Hispanic beneficiaries in the long-term nursing facility resident population (Appendix Exhibit 4).¹⁴ When removing HCCs from the spending model (i.e., adjusting only for age, sex, enrollment segment, and county), the coefficients for Black beneficiaries, Hispanic beneficiaries, and those with less than a high school degree were not attenuated (Appendix Exhibit 5).¹⁴ Results of CAHPS analyses were qualitatively similar after adjustment for supplemental insurance (Appendix Exhibit 6).¹⁴

DISCUSSION

In this study of community-dwelling FFS Medicare beneficiaries, Medicare spending was similar or substantially lower for groups at higher risk of experiencing social disadvantage after adjustment for variables in the current HCC risk-adjustment model. That HCC-adjusted spending was not higher for these groups is consistent with other studies but may run counter to expectations.^4,5 For example, some may extrapolate from evidence of worse risk-adjusted health outcomes for the same groups, assuming that social predictors of worse outcomes should also predict higher spending. Our findings suggest that adding social factors to the HCC model can entrench, rather than reduce, health disparities by lowering population-based payments to more accurately predicted levels of spending.

Health status, functional status, and access to care were consistently worse than predicted by the HCC model for Black, Hispanic, and less educated beneficiaries. The lower or similar HCC-adjusted spending for these groups is therefore not explained by better health but rather is incommensurate with their greater health care needs. Our varied results across groups suggest that moving from FFS to population-based payment under current risk adjustment would reallocate resources to better meet the needs of some groups but not others. Specifically, we find that HCC-adjusted population payments would increase per-beneficiary provider payments for Black and Hispanic beneficiaries by $376–1,264 (approximately 4–14%) above current FFS spending. Although greater increases may be necessary to correct underuse and other quality deficits, these are sizable redistributions that require only continued movement away from FFS toward population-based payments. In the case of ACO models, this requires moving away from benchmarks that incorporate historical spending, which reflect underspending for Black and Hispanic beneficiaries, toward a system of risk-adjusted regional rates.^15,16 It is arguably fortuitous that omission of race and ethnicity from the HCC model results in meaningful implicit reallocations insofar as data on race and ethnicity is imperfect; progress need not await better data. Moreover, if more explicit adjustments were needed, they could face legal challenges.¹⁷

In contrast, HCC-adjusted population-based payments would result in minimal reallocations toward beneficiaries with less than a high school degree. Thus, additional payment reallocations would be needed to better resource efforts to address education-related disparities, which were larger than racial and ethnic disparities in health and functional status. One approach is to use constrained regression to increase the payment weights on HCCs more prevalent among beneficiaries with less education so that payment exceeds current spending for them by a desired amount.¹⁸ An advantage of this approach is that it does not require data for the full population; data on education for a sample (e.g., CAHPS) would suffice. Another approach is to implement post-estimation adjustments (after estimation of the risk-adjustment model) to redistribute payment toward the group of interest. This approach requires data on the full population, which are not currently available at the individual level for education (or many other socioeconomic variables). Accordingly, the ACO REACH model implemented this approach using community-level variables to increase ACO benchmarks above HCC-predicted spending for communities with greater needs.¹⁹

When using analogous sociodemographic predictors at the community level, our analyses produced estimates that were generally similar in direction but smaller in magnitude than those produced by analyses of individual-level predictors. Based on these results, we can conclude that moving toward population-based payment under current risk adjustment would not result in substantial within-county payment redistributions between communities with different racial, ethnic, or educational composition. Thus, as in the case of individual-level education, additional payment adjustments would be needed to reallocate resources towards communities in need.

While payment adjustments at the community level may be considered poorly targeted when the intention is to benefit a subgroup of residents, they may nevertheless be important complements. Individual-level adjustments are critical to establish incentives for MA plans or ACOs to compete for underserved patients by offering them enhanced benefits or care. For plans or ACOs to act effectively on those incentives, however, they must include providers serving those patients’ communities. If patients are not enrolled in MA or aligned with an ACO, they cannot benefit from advantageous population-based payments. Since providers serve communities, payment reallocations at the area level may have greater influence on entry and provider network inclusion decisions made by plans and ACOs.

Furthermore, plans and ACOs may face higher costs relative to FFS spending in historically marginalized communities. These additional costs may include higher transaction costs incurred when contracting with a more fragmented set of providers and/or higher costs of developing the necessary information systems and care management infrastructure to achieve efficiencies under a risk contract. Providers in low-income communities are also less likely to have favorable payer mixes and reserves for financing de novo ACO formation.²⁰ Thus, community-level incentives may be necessary to encourage formation or entry that might otherwise require distortionary incentives at the individual level. For example, consider a community in which 20% of beneficiaries are Black (>2 times the national percentage). An approach with a mix of individual-level and area-level payment redistributions might include a $250 higher payment for all beneficiaries in that community (to encourage entry) plus a $500 per-beneficiary payment increase for Black beneficiaries in all communities (to support reductions in within-community disparities). To achieve the same incentive for entry into that community without an area-level adjustment, the individual-level adjustment for Black beneficiaries would have to be $1750.

Spending for Black, Hispanic, and less educated beneficiaries remained lower after removing HCC adjustments from the model (leaving age, sex, enrollment segment, and county as predictors), indicating that their lower HCC-adjusted spending is not an artifact of differences in coding intensity between groups. Taken together, our findings also suggest that substituting social predictors for manipulable diagnoses in risk-adjustment models would, alone, not be sound strategy. Doing so would mitigate coding incentives but risks reintroducing risk selection incentives and eliminating equity-advancing reallocations implicitly achieved by the current risk-adjustment system. Attention to such tradeoffs will be critical in efforts to address coding incentives in a population-based payment system.

Recommendations for Policy and Research

By departing from predictive accuracy as the singular goal of risk adjustment, a population-based payment system that intentionally sets payments above current levels of FFS spending for groups with greater deficits in health care access or quality (and sets payments below current spending for others) would create incentives for providers or plans to attract those groups and help address resource disparities that contribute to health care disparities. In theory, a combination of intrinsic motivation and competitive pressures should then lead providers or plans to pass the additional resources through to the intended groups in ways that improve their care and outcomes. Based on conceptual consideration and our empirical findings, we anticipate that a combination of additional individual-level and area-level payment adjustments will be needed to establish conditions for meaningful progress. Our findings also suggest that continued expansion of population-based payment models would promote equity across different racial and ethnic groups even under the current risk-adjustment system. Ideally, more comprehensive data would be available to support additional individual-level adjustments, but techniques such as constrained regression could be used in the meantime.¹⁸

The extent to which intended pass-throughs occur is an important topic for research. Studies of MA suggest that competition does indeed promote enhanced offerings to attract enrollees in general.^21,22 It is unclear, however, whether plans respond to higher payments for historically marginalized groups by offering enhancements for those groups specifically. Our results for FFS spending adjusted for supplemental coverage suggest that MA plans receive payments for Black and Hispanic enrollees that are favorable relative to expected expenditures in the absence of such enhancements. Plans should therefore have strong incentives to compete for them with differential benefits (e.g., additional outreach, language-concordant customer service and case management, broader network breadth for prevalent conditions, etc.).

Our FFS estimates are more directly applicable to ACOs. Since provider organizations serve specific communities and exert more direct control over quality of care, ACOs may be positioned to pass through additional payments to patient groups in a more targeted fashion than plans. However, ACOs are statutorily limited in the enhancements they can offer, and competition for patients among ACOs may be weaker – and thus less effective in driving pass-throughs – because patients face higher costs when switching providers than when switching plans. Thus, equity-promoting ACO benchmarks may need to be coupled with mechanisms for passing through tangible benefits directly. For example, Medicare could apply a share of its cut of an ACO’s gross savings to reduce Part B and D premiums for the ACO’s aligned patients. Patients of ACOs disproportionately serving Black and Hispanic patients, for example, would receive greater premium reductions, all else equal, because of savings induced by setting benchmarks 4–14% above spending for Black and Hispanic beneficiaries (as we estimate).

As other indicators of historical disadvantage are considered, coding practices continue to evolve, and risk adjustment is refined (e.g., to limit coding incentives), we recommend expanding and repeating the analytic exercise we have demonstrated, as the results and implications for payment policy may change. Finally, since the optimal distribution of payment cannot be determined from a predictive exercise, the process must be iterative. As such, it will be important to monitor disparities to understand the impact of initial reallocations and inform subsequent adjustments of population-based payments.