State variation in the characteristics of Medicare‐Medicaid dual enrollees: Implications for risk adjustment

Abstract

Objective

To examine between‐state differences in the socioeconomic and health characteristics of Medicare beneficiaries dually enrolled in Medicaid, focusing on characteristics not observable to or used by policy makers for risk adjustment.

Data Source

2010‐2013 Medicare Current Beneficiary Survey.

Study Design

Retrospective analyses of survey‐reported health and socioeconomic status (SES) measures among low‐income Medicare beneficiaries and low‐income dual enrollees. We used hierarchical linear regression models with state random effects to estimate the between‐state variation in respondent characteristics and linear models to compare the characteristics of dual enrollees by state Medicaid policies.

Principal Findings

Between‐state differences in health and socioeconomic risk among low‐income Medicare beneficiaries, as measured by the coefficient of variation, ranged from 17.5 percent for an index of socioeconomic risk to 20.3 percent for an index of health risk. Between‐state differences were comparable among the subset of low‐income beneficiaries dually enrolled in Medicare and Medicaid. Dual enrollees with incomes below the Federal Poverty Level were in better health and had higher SES in states that offered Medicaid to individuals with relatively higher incomes. Duals' average incomes were higher in states with Medically Needy programs.

Conclusions

Characteristics of dual enrollees differ substantially across states, reflecting differences in states' low‐income Medicare populations and Medicaid policies. Risk‐adjustment methods using dual enrollment to proxy for poor health and low SES should account for this state‐level heterogeneity.

1. INTRODUCTION

The Centers for Medicare and Medicaid Services (CMS) uses dual enrollment in Medicare and Medicaid to proxy for the social and clinical vulnerability of patients in its risk adjustment of Medicare payments. For example, spending benchmarks in Medicare Accountable Care Organizations and Medicare Advantage plans are adjusted for the Hierarchical Condition Category (HCC) risk score, which is calculated from Medicare beneficiaries' clinical, demographic, and enrollment characteristics, including dual enrollment status.1, 2, 3 In 2019, CMS began to compare readmission rates and determine penalties in the Hospital Readmission Reduction Program based on each hospital's proportion of Medicare beneficiaries who were enrolled in Medicaid.4, 5 These methods are intended to account for the higher average risk of dual enrollees, which are a reflection of the pathways to eligibility for both programs6, 7, 8, 9: Patients establish eligibility for Medicare if they are elderly or have a qualifying disability, and can receive Medicaid if they have low income and assets, or incur high medical expenses relative to income in states with Medically Needy programs.10, 11

Although higher risk on average, dual enrollees are a heterogeneous group: 38 percent have limitations performing 3 or more activities of daily living, while a nearly equal proportion have none; 45 percent have less than a high school education, while approximately one‐quarter completed some college; and 51 percent originally qualified for Medicare because of a disability, while 47 percent qualified because of age.12, 13 This heterogeneity complicates what it means to adjust for dual enrollment status, particularly if the health and socioeconomic characteristics of dual enrollees differ between providers, regions, or states for which an analyst wishes to compare risk‐adjusted performance.6, 11

Two factors may contribute to differences in the characteristics of dual enrollees between states in particular. First, the composition of low‐income populations differs by geography, resulting in state‐level variation in the characteristics of individuals who may be eligible for Medicaid. For example, of individuals age 65 and older with incomes ≤100 percent of the Federal Poverty Level (FPL), 38 percent in Mississippi live in deep poverty (income < 50 percent of the FPL), while 27 percent in Massachusetts live in deep poverty.14 Second, differences in state policy can affect who receives Medicaid.15 Subject to federal guidelines, states have discretion to set income limits for Medicaid, to offer Medicaid to individuals with catastrophic medical expenses (via Medically Needy programs), and to establish different procedures for determining Medicaid eligibility.10 In Massachusetts, for example, Medicare beneficiaries with incomes ≤ 100 percent of the FPL and those who have high health care expenses and meet Medically Needy spend‐down requirements are eligible for Medicaid. However, Mississippi caps Medicaid eligibility for Medicare beneficiaries at 75 percent of the FPL and does not offer a Medically Needy pathway to Medicaid.16 A consequence of this policy variation is that the income and health status of dual enrollees may differ between states.

In this paper, we used the 2010‐2013 Medicare Current Beneficiary Survey linked to contemporaneous Medicare enrollment and claims data to quantify between‐state differences in the health and socioeconomic characteristics of two populations: (a) low‐income Medicare beneficiaries and (b) low‐income Medicare beneficiaries who were dually enrolled in Medicaid (a subset of the first population). We assessed the extent to which these populations differed on survey‐reported measures of health and socioeconomic status, net of variation in characteristics that are observable in administrative data and commonly used by CMS for risk adjustment. We also examined between‐state differences in the health and socioeconomic characteristics of dual enrollees as a function of state Medicaid policies, controlling for compositional differences in states' low‐income Medicare populations. Our analyses illuminate the extent to which dual enrollees differ between states in ways that are not captured by CMS risk adjustment and highlight the contribution of state policies to differences in states' dually enrolled populations. Our findings can guide efforts to update risk‐adjustment models to account for between‐state differences in the characteristics of dual enrollees.

2. STATE POLICIES AND DIFFERENCES IN DUALLY ENROLLED POPULATIONS

Several state policies may affect Medicaid eligibility and take‐up among low‐income Medicare beneficiaries, and consequently, the characteristics of states' dually enrolled populations. These policies and our underlying data sources are tabulated in Appendix S1.

First, state Medicaid programs set different income eligibility criteria for individuals who meet categorical definitions of need due to age, blindness, or disability—populations that are also eligible for Medicare. In 2013, these income eligibility limits ranged from 66 percent of the FPL in Connecticut to 100 percent of the FPL in 17 states—a difference equivalent to $5300 in income for a married couple in 2013. Most states align their Medicaid eligibility criteria with the Supplemental Security Income (SSI) program's federal benefit rate of approximately 75 percent of the FPL, while 24 states (including the District of Columbia, which we refer to as a state) offered Medicaid to Medicare beneficiaries with incomes above the SSI eligibility limit up to 100 percent of the FPL.16, 17, 18 All else equal, we expect that duals will have higher average incomes in states with more generous Medicaid income eligibility limits. Given the strong association between income and health, we further expect that duals may be in better health in states with more generous eligibility criteria.19, 20

Second, states differ in policies that allow Medicare beneficiaries with high health care costs who would not otherwise qualify for Medicaid to deduct (ie, “spend down”) these costs from income to receive Medicaid.11, 21 In 2013, 36 states had Medically Needy programs, which account for 9 percent of Medicaid enrollment among duals nationally.12 On average, we expect dual enrollees will have relatively higher incomes in states that offer Medically Needy coverage compared to duals in other states. (Medicare enrollment data, which are used for risk‐adjusting Medicare payments and to which we had access via the MCBS, do not differentiate between categorically and Medically Needy duals, although Medicaid enrollment data report separate eligibility pathways.) On the other hand, because individuals must have sufficiently high medical costs to meet spend‐down requirements, we expect that duals may be sicker in states with Medically Needy programs, particularly in states with more stringent Medically Needy eligibility criteria (which require individuals to have higher medical costs or lower income).11

Third, states differ in the ease with which the lowest‐income Medicare beneficiaries can enroll in Medicaid. Consider individuals who qualify for SSI because they have low incomes (≤75 percent of the FPL for a single person) and assets (≤$2000 for a single person or ≤$3000 for a couple). In 2013, 33 states had Section 1634 agreements with the Social Security Administration to automatically enroll SSI recipients in Medicaid, simplifying the application process for SSI eligibles in these states. In the remaining 18 states, applicants must file a separate Medicaid application with their state Department of Human Services. Seven of these states followed SSI criteria in determining Medicaid eligibility but still required SSI recipients to apply separately to Medicaid, while 11 states exercised an option to apply more stringent income or asset limits than the SSI program, per Section 209(b) of the Social Security Act.16 These differences in Medicaid and SSI eligibility determination may affect who enrolls in Medicaid. Economic theory predicts that individuals will enroll in Medicaid if the expected benefit of coverage exceeds the “cost” (eg, administrative burden) of applying.22 Because the value of Medicaid is higher for individuals who require more care and have fewer financial resources, enrollees tend to be poorer and sicker than eligible nonparticipants.23 In states that require SSI recipients to apply separately to Medicaid, SSI‐eligible dual enrollees may be more adversely selected on health and socioeconomic characteristics, since it is only for these higher risk patients that the value of coverage exceeds the higher cost of enrolling. Therefore, all else equal, we expect to see better health and higher socioeconomic status among dual enrollees (of whom 35 percent qualified for Medicaid via SSI12) in states with Section 1634 agreements vs other states.

3. METHODS

3.1. Data and sample

We analyzed 2010‐2013 data from the Medicare Current Beneficiary Survey (MCBS) Cost and Use files linked to contemporaneous enrollment and claims for fee‐for‐service Medicare beneficiaries. The MCBS employs a rotating panel design, surveying individuals for up to four years and replenishing the sample annually for decedents and persons exiting the survey.24 Because we sought to examine differences in the characteristics of low‐income Medicare beneficiaries between states, rather than changes in this population over time, we randomly sampled one year of survey data per respondent, yielding a cross section of individuals pooled over 4 years. This sample encompassed MCBS respondents from 46 states. We excluded respondents in 10 sparsely sampled states (<0.2 percent of our sample) for which we could not reliably estimate state‐level population characteristics (Appendix S2). Thus, our final sample included community‐dwelling and institutionalized Medicare beneficiaries from 36 states. Medicaid policies in the included states were generally similar to policies in the 15 states not represented in our analyses (Appendix S3).

3.2. Analysis population

We examined the characteristics of Medicare beneficiaries with income ≤ 135 percent of the FPL, which we term our low‐income sample, and low‐income beneficiaries dually enrolled in Medicaid (a subset of the low‐income sample). We selected an income threshold of 135 percent of the FPL because Medicare beneficiaries with incomes above the Federal Poverty Level may qualify for Medicaid through income disregards, by meeting Medically Needy spend‐down requirements, or if they require long‐term care.10 This low‐income sample encompassed approximately 90 percent of MCBS respondents who were enrolled in Medicaid.

We used monthly Medicare‐Medicaid dual eligibility codes in linked administrative data to identify beneficiaries receiving ≥1 month of full Medicaid in the study year, whom we defined as dual enrollees (Appendix S2). We focused on Medicare beneficiaries receiving full Medicaid, since most risk‐adjustment methods treat full Medicaid enrollment as a marker of social and clinical risk.8, 25, 26, 27, 28 Moreover, full‐benefit duals account for 72 percent of Medicaid enrollment in the Medicare population and a disproportionate share of Medicare spending.12 We measured income relative to the FPL using annual poverty guidelines published by the Department of Health and Human Services17 and by imputing family size, when missing in the MCBS, from marital status.

In sensitivity analyses, we analyzed beneficiaries with incomes ≤100 percent of the FPL, net of a standard income disregard of $20 per month,29 which we term our poverty sample. By construction, the low‐income sample will include more individuals who qualified for Medicaid via a Medically Needy pathway (ie, spend‐downs), while the poverty sample will include a higher proportion of categorically needy duals who qualified for Medicaid because of their low incomes. The poverty sample encompassed 75 percent of MCBS respondents with a contemporaneous administrative record of Medicaid enrollment. We examined between‐state differences in the characteristics of MCBS respondents in this poverty sample and among a subset of individuals in the poverty sample who were dual enrollees.

3.3. Beneficiary characteristics

Using the MCBS and linked Medicare enrollment and claims data, we identified three sets of variables that captured Medicare beneficiaries' demographic characteristics and different dimensions of their socioeconomic and health status. We selected these variables either because CMS currently uses them for risk adjustment or because prior research demonstrated their relationship to outcomes in pay‐for‐performance programs, and because these variables (eg, survey‐reported measures of socioeconomic status) are unlikely to be affected by a provider's quality of care. The National Academies of Sciences, Engineering, and Medicine identified these variable characteristics in its criteria for including variables in risk‐adjustment models.9 Table 1 describes these variables; Appendix S4 provides additional information about how we constructed consistent health status measures for individuals who responded to the MCBS community vs facility surveys, which use different questionnaires.

Table 1.

Patient characteristics assessed from the Medicare Beneficiary Survey and linked Medicare administrative and claims files

Variable name	Variable construction	Variable description	Source data
Patient characteristics used by CMS for risk adjustment
Age	Continuous	Age in years	Administrative
Gender	Binary	Indicator that the Medicare beneficiary was female	Administrative
Disabled	Binary	Medicare entitlement because of disability	Administrative
End‐stage renal disease	Binary	Medicare entitlement because of end‐stage renal disease	Administrative
Lives in a long‐term care facility	Binary	Beneficiary resided in a long‐term care facility (vs the community) and took the facility health status and functioning survey	MCBS facility screener, health status, and functioning surveya
Hierarchical Condition Category (HCC) score	Continuous	Continuous risk score predicted from diagnoses on claims (measured over a one‐year look‐back period), demographic variables, and enrollment characteristics	Administrative and claims files
Additional socioeconomic and demographic characteristics from the MCBS
Income ($)	Continuous	Annual family income	Demographic questionnaire
Marital status	Binary	Not married	Demographic questionnaire
High school education or less	Binary	High school education or less (vs postsecondary education or higher)	Demographic questionnaire
Additional health status characteristics from the MCBS
Count of self‐reported diagnoses (diseases for which the respondent was ever diagnosed)	Count of 0‐10 diagnoses	Count of diseases for which the respondent was ever diagnosed (self‐report): (1) hypertension, (2) myocardial infarction, (3) stroke, (4) diabetes, (5) cancer (including skin cancer), (6) Alzheimer's disease or dementia, (7) depression, (8) arthritis (including rheumatoid arthritis), (9) osteoporosis, (10) asthma, emphysema, or COPD	Health status and functioning surveyb
Self‐rated health is fair or poorc	Binary	General health is fair or poor (vs good to excellent) compared to others of the same age	Health status and functioning survey
Health is worse compared to prior yearc	Binary	Health is somewhat or much worse (vs the same, somewhat better, or much better) compared to one year ago	Health status and functioning survey
Health limits social activityc	Binary	Health limited social activity some, most, or all of the time (vs none of the time) in the past month	Health status and functioning survey
Difficulty seeing or blindnessc	Binary	Trouble seeing or blindness	Health status and functioning survey
Difficulty hearing or deafnessc	Binary	Trouble hearing or deafness	Health status and functioning survey
Number of difficulties performing activities requiring mobility or agility	Count of 0‐5 activities	Count of the following activities which the respondent reported at least some difficulty: (1) stooping, crouching, or kneeling, (2) lifting or carrying 10 pounds, (3) reaching or extending arms above the shoulder, (4) writing or grasping objects (5) walking ¼ mile or 2‐3 blocks	Health status and functioning survey
Number of difficulties performing instrumental activities of daily living (IADLs)	Count of 0‐5 limitations	Count of the following activities which the respondent reported any difficulty performing or did not perform due to poor health: (1) using the telephone, (2) shopping for personal items, (3) managing money, (4) doing light housework, (5) preparing meals	Health status and functioning survey
Number of difficulties performing activities of daily living (ADLs)	Count of 0‐6 limitations	Count of the following activities which the respondent reported any difficulty performing or did not perform due to poor health: (1) bathing, (2) dressing, (3) eating, (4) getting into out of bed or a chair, (5) walking, (6) using the toiletd	Health status and functioning survey
Obesec	Binary	Body mass index ≥ 30	Health status and functioning survey

^aIndividual resided in a long‐term care facility (eg, a nursing home) and took the facility interview.

^bThe MCBS uses separate health status and functioning questionnaires for individuals residing in long‐term care facilities vs the community. Appendix S4 describes how we constructed consistent health status and functioning measures for community‐ and facility‐dwelling respondents.

^cBinary measures were coded such that a value of “1” corresponded to higher health risk.

^dSee the notes to Appendix S4 for how we constructed these measures for community‐ vs facility‐dwelling respondents.

First, we identified characteristics of beneficiaries that are routinely used to risk adjust Medicare payments. Specifically, we used Medicare administrative data to assess beneficiaries' age, sex, disability status, and presence of end‐stage renal disease (ESRD). We distinguished between beneficiaries residing in long‐term care facilities vs the community based on whether they responded to the MCBS facility vs community interview,30 and we used enrollment data and diagnoses in contemporaneous Medicare claims to construct a beneficiary‐level HCC risk score (using version 21 of the CMS‐HCC risk‐adjustment software).2, 31

Second, we used responses to the MCBS demographic questionnaire to assess three socioeconomic characteristics of respondents: (a) annual family income, (b) marital status, and (c) educational attainment. For the latter two characteristics, we constructed binary measures denoting whether respondents were unmarried or had a high school education or less (vs some postsecondary education or higher). Education and marriage have well‐documented relationships with health and outcomes of care, including hospital readmissions and mortality.20, 32, 33, 34

Third, we used responses to the MCBS health status and functioning questionnaire to assess 10 measures of self‐reported health status and functional impairment: (1) a count of chronic conditions (eg, depression, diabetes, and COPD) for which the respondent ever reported being diagnosed; (2) fair or poor self‐rated health; (3) worse health compared to the prior year; (4) limitations engaging in social activities due to health; (5) difficulty seeing or blindness; (6) difficulty hearing or deafness; (7) difficulties performing activities requiring mobility or agility; (8) difficulties performing activities of daily living (ADLs, such as bathing, eating, or dressing); (9) difficulties performing instrumental activities of daily living (IADLs, such as shopping or completing housework35); and (10) obesity (Body Mass Index ≥ 30). While the CMS‐HCC score is calculated from disease indicators on Medicare claims in the past year, the MCBS asks about individuals' cumulative disease history. A large literature shows that functional limitations, comorbidities (including disease history), and self‐rated health exhibit strong relationships with readmissions,20, 32, 36, 37 mortality,20, 38, 39 spending,28, 40 and other outcomes commonly assessed in pay‐for‐performance programs.41 In addition, risk factors such as hearing and vision impairment may be precipitated by declines in health status and increase individuals' health care needs.42, 43

3.4. Indices of socioeconomic and health status

We constructed indices of socioeconomic and health status from variables in the MCBS. To construct these indices, we estimated a linear regression model that predicted enrollment in Medicaid for person i in year t as a function of CMS risk‐adjustment variables and the additional MCBS‐derived covariates from Table 1:

$${\text{Medicaid}}_{\mathit{it}}={\beta }_0+{\beta }_1{\text{CMS}}_{\mathit{it}}+{\beta }_2\text{MCBS}\_{\text{SES}}_{\mathit{it}}+{\beta }_3\text{MCBS}\_{\text{Health}}_{\mathit{it}}+{\epsilon }_{\mathit{ist}}$$

We estimated this model in our low‐income sample. The regression estimates ${\widehat{\beta }}_2$ and ${\widehat{\beta }}_3$ represent the predicted associations between the MCBS socioeconomic and health status covariates and the probability of Medicaid enrollment, conditional on CMS risk‐adjustment variables. We constructed an SES index from a linear combination of socioeconomic variables, using ${\widehat{\beta }}_2$ to weight the constituent covariates, and a health status index from a linear combination of the health and functional status variables, using ${\widehat{\beta }}_3$ for weights (Appendix S5). We constructed a composite index by summing the health and SES indices.

Each index was standardized to have a standard deviation of 1 and mean equal to the unadjusted proportion of low‐income Medicare beneficiaries who were enrolled in Medicaid. Thus, a one‐unit change in each index represents a 1 standard deviation increase in the probability that a low‐income Medicare beneficiary had Medicaid. Because beneficiaries were more likely to receive Medicaid if they were in poorer health and had social risk factors (eg, lower income, education, and social supports; see Appendix S6), higher values of this index correspond to higher levels of health risk and socioeconomic disadvantage.

3.5. Statistical analyses

We performed two sets of statistical analyses. First, in our low‐income sample, we examined the extent to which the socioeconomic and health characteristics of Medicare beneficiaries and dual enrollees differed across states. We estimated the state‐level standard deviation in each characteristic, net of sampling error, using hierarchical linear regression models with state random effects. Specifically, for each health and socioeconomic characteristic assessed from the MCBS and for our three indices, we estimated a model of the form:

$${\text{characteristic}}_{\mathit{ist}}=\alpha +\left[\theta {\text{CMS}}_{\mathit{it}}\right]+{\text{year}}_t+{\mu }_s+{\epsilon }_{\mathit{ist}}$$

for person i living in state s in year t. In this model, ${\mu }_s$ is a normally distributed state‐level random variable with variance ${\sigma }_s^2$; ${\text{CMS}}_{\mathit{it}}$ is a vector of risk‐adjustment variables used by CMS (age, gender, disability and ESRD status, nursing home residence, and HCC score); and ${\epsilon }_{\mathit{ist}}$ is an error term. We estimated separate models for low‐income beneficiaries and low‐income duals, adjusting each model for survey sampling weights.44, 45 We conducted analyses with and without adjustment for CMS variables (bracketed). When adjusting for the CMS variables, ${\widehat{\sigma }}_s$ represents the between‐state standard deviation in a characteristic, net of variation that is explained by variables observable to—and typically used by—CMS for risk adjustment. From each model, we report a coefficient of variation, defined as the between‐state standard deviation in each characteristic, ${\widehat{\sigma }}_s$, relative to the national average.

Second, motivated by our conceptual framework, we compared the characteristics of low‐income dual enrollees in states categorized by their Medicaid policies for Medicare beneficiaries: Income eligibility limits for individuals who meet categorical definitions of need due to age, blindness, or disability (≤75 percent vs 76 percent‐100 percent of the FPL); the availability of a Medically Needy pathway to Medicaid (or analogous program in 209(b) states that are required to have a spend‐down eligibility pathway); and Section 1634 agreements to automatically enroll SSI recipients in Medicaid.16, 46 We assessed differences in the characteristics of dual enrollees in states with the different Medicaid policies by estimating linear regression models of the form:

$${\text{characteristic}}_{\mathit{ist}}={\beta }_0+{\beta }_1{\text{Policy}}_{\mathit{st}}+{\beta }_2{\text{CMS}}_{\mathit{it}}+{\beta }_3\text{lowinc}\_{\text{char}}_s+{\text{year}}_t+{\epsilon }_{\mathit{ist}}$$

among dual enrollees in our low‐income sample. The dependent variable is a socioeconomic or health characteristic of person i living in state s in year t; $\text{lowinc}\_{\text{char}}_s$ controls for the state‐level average of the corresponding characteristic among Medicare beneficiaries in our low‐income sample (regardless of dual enrollment status); ${\text{year}}_t$ are year fixed effects; and ${\text{Policy}}_{\mathit{st}}$ is a state policy variable. Thus, ${\beta }_1$ represents the mean difference in the health or SES of dual enrollees in states with different Medicaid policies, controlling for compositional differences in states' low‐income Medicare populations and for person‐level risk factors observable to CMS. We estimated separate models for each set of state policies, adjusting each for survey weights and clustering standard errors at the state level. Because of the large number of characteristics we examined, we focused these comparisons on family income and indices of health and SES to reduce Type I error from multiple comparisons.

4. RESULTS

Our unweighted sample consisted of 7222 low‐income Medicare beneficiaries (income ≤ 135 percent of the FPL) from 36 states, of whom 3365 (46.6 percent) were dual enrollees (≥1 month of full Medicaid enrollment in the sampled year; see Appendix S2 for details). When weighted, our sample comprised 30 764 380 low‐income Medicare beneficiaries and 12 034 249 dual enrollees surveyed during the period 2010‐2013. The low‐income sample included 89 percent of MCBS respondents who were dually enrolled in Medicaid.

4.1. Between‐state variation in population characteristics

Table 2 displays population means and state‐level distributions of the health and socioeconomic characteristics of Medicare beneficiaries in our low‐income sample. We found substantial between‐state differences in low‐income beneficiaries' educational attainment, self‐rated health, and functional limitations. For example, the between‐state standard deviation in the proportion of low‐income beneficiaries with a high school education or less (not adjusting for other individual‐level characteristics) was 6.7 percent, meaning that among states within one standard deviation of the national mean (67.4 percent), the prevalence of low education ranged from 60.7 percent to 74.1 percent. The proportion of low‐income beneficiaries who rated their health fair or poor ranged from 37.2 percent to 48.6 percent in states within one standard deviation of the national mean (42.9 percent), consistent with a between‐state coefficient of variation of 13.3 percent. The coefficient of variation for the per‐beneficiary number of limitations performing IADLs, prevalence of vision impairments, and presence of hearing difficulties was 9.4 percent, 22.8 percent, and 26.9 percent, respectively. We found similarly large between‐state differences in indices of these variables (coefficient of variation: 17.5 percent for the SES index and 20.3 percent for the health status index). These between‐state differences were attenuated minimally when we controlled for patient characteristics observable to CMS and used in standard risk adjustment (rightmost column of Table 2).

Table 2.

Population means and between‐state differences in the characteristics of low‐income Medicare beneficiaries

	Population mean	Between‐state standard deviationa		Coefficient of variationb
		Unadjusted	Adjusted for characteristics used in standard CMS risk adjustmentc	Unadjusted	Adjusted for characteristics used in standard CMS risk adjustmentc
Patient characteristics used by CMS for risk adjustment
Age (y)	66.4	1.2	—	1.9	—
Female gender, %	59.2	2.6	—	4.5	—
Disabledd, %	33.0	5.3	—	16.0	—
End‐stage renal diseased, %	1.7	—e	—	—e	—
Lives in a long‐term care facility, %	7.9	1.3	—	16.7	—
HCC scoref: <1, %	52.9	4.9	—	9.2	—
HCC score: ≥1 and <1.7, %	21.2	2.7	—	12.6	—
HCC score: ≥1.7, %	26.0	3.2	—	12.3	—
Additional socioeconomic and demographic characteristics from the MCBS
Income relative to the FPL, %g	81.9	3.7	3.6	4.5	4.4
Income ($)g	12 548	222	230	1.8	1.8
Not married, %	73.3	3.8	3.3	5.2	4.5
High school education or less, %	67.4	6.7	6.6	9.9	9.8
Additional health status characteristics from the MCBS
Count of self‐reported diagnoses, count (0‐10 diagnoses)	2.94	0.31	0.25	10.6	8.6
Self‐rated health is fair or poor, %	42.9	5.7	5.1	13.3	12.0
Health is worse compared to prior year, %	30.1	7.0	6.7	23.1	22.2
Health limits social activity, %	50.5	5.5	5.1	10.9	10.1
Difficulty seeing or blindness, %	9.0	2.1	2.0	22.8	21.9
Difficulty hearing or deafness, %	8.8	2.4	2.4	26.9	26.9
Number of difficulties performing activities requiring mobility or agility (range: 0‐5)	2.05	0.22	0.21	10.8	10.2
Number of difficulties performing IADLs (range: 0‐5)	1.23	0.12	0.14	9.4	11.4
Number of difficulties performing ADLs (range: 0‐6)	1.29	0.10	0.14	7.6	11.2
Obese (BMI ≥ 30), %	33.8	4.3	4.0	12.9	11.7
Composite measures of patient characteristics added from the MCBSh
Socioeconomic status index	0.39i	0.07	0.06	17.5	15.7
Health status index	0.39	0.08	0.11	20.3	29.1
Composite socioeconomic and health status index	0.39	0.09	0.08	23.1	21.3

N = 7222 Medicare beneficiaries in our low‐income sample (income ≤ 135% of the FPL), of whom 3365 were enrolled in Medicaid (Appendix S2). When weighted, the proportion of dual Medicare‐Medicaid enrollees in our low‐income sample was 39%.

Estimates were adjusted for MCBS cross‐sectional sampling weights.

^aWe estimated the between‐state standard deviation in each patient characteristic by fitting a hierarchical linear regression model with state random effects to each characteristic and reporting the standard deviation of the state random effect.

^bRatio of the between‐state standard deviation to the population mean, multiplied by 100.

^cObservable characteristics in Medicare enrollment and claims data that CMS typically uses to risk adjust pay‐for‐performance measures (age, gender, disability status, end‐stage renal disease, facility vs community residence, and the Hierarchical Condition Category risk score).

^dReason for Medicare entitlement was disability or end‐stage renal disease (as opposed to age).

^eNot sufficiently present in the study population to estimate the between‐state standard deviation in the prevalence of this characteristic.

^fHierarchical Condition Category (HCC) risk scores are derived from demographic and diagnostic data in Medicare enrollment and claims files, with higher risk scores indicating higher predicted Medicare spending. We categorized the HCC score due to its non‐normal distribution in our study population.

^gWe report income measured in dollars and relative to the FPL. Income in dollars is one of the component variables of our socioeconomic status and composite health/socioeconomic status indices (see Appendix S5).

^hThese indices were constructed as linear combinations of health and socioeconomic characteristics from the MCBS, with each constituent variable weighted by its association with dual Medicaid enrollment. We estimated these associations by fitting a linear regression model in our sample of low‐income Medicare beneficiaries predicting dual enrollment as a function of base CMS variables and the additional health and socioeconomic characteristics from the MCBS (see Appendix S5). The resulting indices were scaled to have a standard deviation of 1.

ⁱCentered at the survey‐weighted mean of the index among dual Medicare‐Medicaid enrollees in our low‐income sample (income ≤ 135% of the FPL).

Among the subset of low‐income Medicare beneficiaries dually enrolled in Medicaid, the prevalence of these socioeconomic and health risks varied considerably between states, as shown in Table 3. For example, the proportion of low‐income duals with a high school education or less (not adjusting for other beneficiary characteristics) ranged from 69.4 percent to 80.8 percent in states within one standard deviation of the national mean (75.1 percent), consistent with a coefficient of variation of 7.7 percent. The coefficient of variation in the mean per‐beneficiary number of difficulties with IADLs was 11.3 percent, and the coefficient of variation in the proportion of low‐income duals reporting fair or poor health, vision impairments, and hearing impairments was 9.5 percent, 21.1 percent, and 28.3 percent, respectively. In states one standard deviation from the national average, the presence of socioeconomic and health risks among low‐income dual enrollees (as measured by our indices of these variables) differed by 13.0 percent and 21.1 percent of the national average, respectively. These differences persisted after we adjusted for standard CMS risk‐adjustment variables.

Table 3.

Population means and between‐state differences in the characteristics of low‐income Medicare beneficiaries dually enrolled in Medicaid

	Population mean	Between‐state standard deviationa		Coefficient of variationb
		Unadjusted	Adjusted for characteristics used in standard CMS risk adjustmentc	Unadjusted	Adjusted for characteristics used in standard CMS risk adjustmentc
Patient characteristics used by CMS for risk adjustment
Age (y)	62.9	3.2	—	5.1	—
Female gender, %	61.9	3.3	—	5.3	—
Disabledd, %	44.9	9.2	—	20.5	—
End‐stage renal diseased, %	2.2	—e	—	—e	—
Lives in a long‐term care facility, %	17.0	2.7	—	16.0	—
HCC scoref: <1, %	38.3	3.8	—	9.8	—
HCC score: ≥1 and <1.7, %	27.7	1.3	—	4.6	—
HCC score: ≥1.7, %	34.0	3.9	—	11.4	—
Additional socioeconomic and demographic characteristics from the MCBS
Income relative to the FPL, %g	71.0	5.5	5.5	7.8	7.8
Income ($)g	10 346	454	420	4.4	4.1
Not married, %	83.8	3.9	3.6	4.7	4.3
High school education or less, %	75.1	5.7	5.4	7.7	7.3
Additional health status characteristics from the MCBS
Count of self‐reported diagnoses, count (0‐10 diagnoses)	3.04	0.40	0.36	13.0	12.0
Self‐rated health is fair or poor, %	54.1	5.2	4.6	9.5	8.4
Health is worse compared to prior year, %	34.3	4.0	3.0	11.5	8.7
Health limits social activity, %	59.2	4.0	3.9	6.8	6.5
Difficulty seeing or blindness, %	10.4	2.2	2.0	21.1	18.9
Difficulty hearing or deafness, %	9.6	2.7	2.5	28.3	25.9
Number of difficulties performing activities requiring mobility or agility (range: 0‐5)	2.45	0.27	0.26	11.0	10.4
Number of difficulties performing IADLs (range: 0‐5)	1.88	0.21	0.20	11.3	10.4
Number of difficulties performing ADLs (range: 0‐6)	1.79	0.15	0.12	8.4	7.0
Obese (BMI ≥ 30), %	35.9	5.1	4.4	14.3	12.2
Composite measures of patient characteristics added from the MCBSh
Socioeconomic status index	0.74i	0.10	0.09	13.0	12.2
Health status index	0.67	0.14	0.14	21.1	20.6
Composite socioeconomic and health status index	0.81	0.14	0.12	17.5	14.8

N = 3365 Medicare beneficiaries in our low‐income sample who were dually enrolled in Medicaid.

Estimates were adjusted for MCBS cross‐sectional sampling weights.

^aWe estimated the between‐state standard deviation in each patient characteristic by fitting a hierarchical linear regression model with state random effects to each characteristic and reporting the standard deviation of the state random effect.

^bRatio of the between‐state standard deviation to the population mean, multiplied by 100.

^cObservable characteristics in Medicare enrollment and claims data that CMS typically uses to risk adjust pay‐for‐performance measures (age, gender, disability status, end‐stage renal disease, facility vs community residence, and the Hierarchical Condition Category risk score).

^dReason for Medicare entitlement was disability or end‐stage renal disease (as opposed to age).

^eNot sufficiently present in the study population to estimate the between‐state standard deviation in the prevalence of this characteristic.

^fHierarchical Condition Category (HCC) risk scores are derived from demographic and diagnostic data in Medicare enrollment and claims files, with higher risk scores indicating higher predicted Medicare spending. We categorized the HCC score due to its non‐normal distribution in our study population.

^gWe report income measured in dollars and relative to the FPL. Income in dollars is one of the component variables of our socioeconomic status and composite health/socioeconomic status indices (see Appendix S5).

^hThese indices were constructed as linear combinations of health and socioeconomic characteristics from the MCBS, with each constituent variable weighted by its association with dual Medicaid enrollment. We estimated these associations by fitting a linear regression model in our sample of low‐income Medicare beneficiaries predicting dual enrollment as a function of base CMS variables and the additional health and socioeconomic characteristics from the MCBS (see Appendix S5). The resulting indices were scaled to have a standard deviation of 1.

ⁱIndices centered at the survey‐weighted proportion of Medicare beneficiaries in our low‐income sample (income ≤ 135% of the FPL) who were dual Medicaid enrollees.

Although smaller relative to the corresponding population means, we also found appreciable between‐state differences in the average income of low‐income Medicare beneficiaries (between‐state SD: $222; coefficient of variation: 1.8 percent) and dual enrollees (between‐state SD: $454; coefficient of variation: 4.4 percent). We found similarly large between‐state differences in the characteristics of Medicare beneficiaries and dual enrollees in our poverty sample (Appendices S8 and S9).

4.2. Differences among dual enrollees between states categorized by Medicaid policies for the Medicare population

Panel A of Table 4 reports estimates of the association between state Medicaid policies and the average income, SES, and health status of dual enrollees in our low‐income sample. We found no statistically significant differences in the income, SES, or health status of duals in states with higher (ie, 76 percent‐99 percent of the FPL) vs lower (≤75 percent of the FPL) Medicaid eligibility limits for the categorically needy, and no differences between states categorized by policies to automatically enroll SSI recipients in Medicaid (Section 1634 agreements). In states offering a Medically Needy pathway to Medicaid for Medicare beneficiaries, the average income of duals was higher by 3.67 percentage points of the FPL (95% CI: 0.45 to 6.89 percentage points, P = .03), and the overall SES and health status of duals was better (ie, the composite risk score was lower) by 0.15 standard deviations (SDs; 95% CI: 0.23 to 0.07, P = .001) than in states without a Medically Needy program.

Table 4.

Differences in the income, socioeconomic status, and health status of dual enrollees between states categorized by Medicaid policies for Medicare beneficiaries

State Medicaid policy for Medicare beneficiaries	Total number of states with policy in 2013	Income (percentage points of the FPL)	Socioeconomic status indexa	Health status indexa	Composite socioeconomic and health status indexa
		Estimate (95% Confidence Intervalb)
Panel A: Dual enrollees in low‐income sample (income ≤ 135% of the FPL)c
Medicaid income limit for the categorically needy is 76%‐100% of the FPLe	24 states	1.03 (−2.05, 4.12)	−0.04 (−0.13, 0.04)	−0.07 (−0.18, 0.04)	−0.06 (−0.17, 0.04)
State has a Medically Needy programf	36 states	3.67 (0.45, 6.89)	−0.11 (−0.20, −0.02)	−0.11 (−0.19, −0.04)	−0.15 (−0.23, −0.07)
SSI recipients automatically enrolled in Medicaid (Section 1634 stateg)	33 states	−1.53 (−4.65, 1.59)	0.01 (−0.07, 0.08)	0.02 (−0.08, 0.11)	0.02 (−0.06, 0.11)
Panel B: Dual enrollees in poverty sample (income ≤ 100% of the FPL, net of standard income disregards)d
Medicaid income limit for the categorically needy is 76%‐100% of the FPLe	24 states	1.88 (0.27, 3.48)	−0.09 (−0.15, −0.03)	−0.10 (−0.19, 0.00)	−0.11 (−0.18, −0.03)
State has a Medically Needy programf	36 states	2.70 (0.66, 4.75)	−0.11 (−0.17, −0.04)	−0.11 (−0.17, −0.04)	−0.14 (−0.20, −0.08)
SSI recipients automatically enrolled in Medicaid (Section 1634 stateg)	33 states	−0.76 (−3.54, 2.02)	−0.02 (−0.11, 0.08)	0.01 (−0.08, 0.11)	−0.001 (−0.08, 0.08)

This table displays differences in the income, socioeconomic status, and health status of dual enrollees between states categorized by Medicaid policies for Medicare beneficiaries. Each cell is an estimate from a separate regression, in which we modeled the association between the state policy (table rows) and the income, socioeconomic status, or health status (table columns) of dually enrolled Medicare beneficiaries in our low‐income sample (Panel A) or our poverty sample (Panel B). Estimates were adjusted for MCBS cross‐sectional sampling weights; person‐level characteristics used in CMS risk adjustment (age, gender, disability status, end‐stage renal disease, long‐term nursing home residence, and the HCC score); the average of the corresponding income, socioeconomic status, or health status measure assessed among Medicare beneficiaries in the low‐income sample (Panel A) and the poverty sample (Panel B) in the state (to control for compositional differences in states' Medicare populations); and year fixed effects. Full model estimates are reported in Appendix S8. Socioeconomic and health indices reflect the predicted association between patient characteristics and dual enrollment, as described in the notes to Tables 2 and 3 and the text, and are scaled such that a one‐unit change in the index represents a one standard deviation change in socioeconomic disadvantage or health risk.

^aHigher values of these indices indicate greater social disadvantage or health risk.

^b95% confidence intervals were constructed using standard errors clustered at the state level.

^cThe sample size for each set of regression estimates in Panel A was 3,365 dual enrollees from the 2010‐2013 MCBS.

^dStandard income disregard of $20/month. The sample size for each set of regression estimates in Panel B was 2852 dual enrollees from the 2010‐2013 MCBS.

^eVersus states with Medicaid income eligibility limits ≤ 75% of FPL.

^fCompared to states without a Medically Needy program for aged, blind, or disabled persons.

^gVersus states without 1634 agreements to automatically enroll SSI recipients into Medicaid.

Panel B of Table 4 reports analogous estimates for duals in our poverty sample. In this sample, the average income of duals was higher by 1.88 percentage points of the FPL (95% CI: 0.27 to 3.48 percentage points, P = .02) in states with higher vs lower income eligibility limits for the categorically needy, and was higher by 2.70 percentage points of the FPL (95% CI: 0.66 to 4.75 percentage points, P = .01) in states with a Medically Needy program for Medicare beneficiaries. These higher average incomes corresponded to lower levels of health risk and socioeconomic disadvantage. For example, in states that provided Medicaid to categorically needy individuals with incomes as high as 76 percent‐100 percent of the FPL, the average socioeconomic risk score of dual enrollees was lower by 0.09 SDs (95% CI: −0.15 to −0.03, P = .01), the average health risk score was lower by 0.10 SDs (95% CI: −0.19 to 0.00, P = .05), and the composite health and social risk score was lower by 0.11 SDs (95% CI: −0.18, −0.03, P = .01), compared to duals in states with income limits ≤ 75 percent of the FPL. The average socioeconomic and health risk of duals was also lower in states with a Medically Needy pathway to Medicaid vs states without this pathway. We found no appreciable differences in the socioeconomic or health characteristics of dual enrollees across states categorized by the presence of a Section 1634 agreement to automatically enroll SSI recipients in Medicaid.

5. DISCUSSION

Using national data from the Medicare Current Beneficiary Survey, we found substantial state‐level variation in the health and socioeconomic characteristics of Medicare beneficiaries enrolled in Medicaid. These differences were particularly pronounced for measures of health status (notably, self‐rated health, functional limitations, and hearing and vision impairments) and were sizeable for socioeconomic indicators (specifically, education and income) that are typically not observable to or used by policy makers for risk adjustment. Although policy makers and researchers routinely use dual enrollment to proxy for low SES and poor health in risk adjustment,6, 8, 26, 41, 47 our results show that this binary categorization belies substantial heterogeneity in what it means to have Medicaid in different states.

Differences in health and socioeconomic characteristics reflected variation in states' low‐income Medicare populations and state policies that affect which low‐income Medicare beneficiaries receive Medicaid. Specifically, we found that the average income of dual enrollees was higher by 2.7‐3.7 percentage points of the Federal Poverty Level (approximately $440‐$610 in income for a married couple in 2018) in states that that offered a Medically Needy pathway to Medicaid for Medicare beneficiaries. This income difference was greater in our low‐income sample (3.7 percent) than in our poverty sample (2.7 percent)—consistent with expectations that the low‐income sample would include more individuals who qualified for Medicaid via Medically Needy spend‐down criteria rather than eligibility criteria for the categorically needy. The average income of duals in our poverty sample was higher by approximately 1.9 percentage points of the FPL ($310 in income for a married couple in 2018) in states with more generous income eligibility criteria for the categorically needy. The income difference associated with states' categorical eligibility criteria was less pronounced and not statistically significant in our low‐income sample, which likely included more Medically Needy duals.

Consistent with our conceptual framework, duals in our poverty sample were in better health on average in states with more generous income eligibility criteria for the categorically needy. Surprisingly, we also found that duals in states with Medically Needy programs were in better health, despite the fact individuals must have high medical costs relative to income to qualify for Medicaid via a Medically Needy pathway. Two factors might account for this finding. First, Medicaid eligibility via a Medically Needy pathway is not limited to individuals with particular types of health conditions or medical expenses.48 Consequently, some Medically Needy duals may have met spend‐down requirements without using services targeted to individuals with the most complex health needs (eg, long‐term services and supports). Second, the differences we detected could reflect unobserved variation in state policies or population characteristics, since we would not expect favorable selection into Medicaid in states whose Medicaid programs insure individuals with high medical costs.

Finally, we did not find a significant relationship between state policies to enroll SSI recipients in Medicaid (Section 1634 agreements) and the socioeconomic or health status of dual enrollees. This finding suggests that policies that affect the propensity of Medicaid enrollment among eligible individuals are less important determinants of selection than policies that affect who is eligible to receive Medicaid based on their income and medical expenses.

Our findings have important implications for risk adjustment. First, they underscore the limitations of using enrollment in Medicaid—a program whose characteristics vary across states—as a basis for risk‐stratifying Medicare beneficiaries in national payment programs. Ideally, risk adjustment would be based on the underlying social and clinical constructs for which dual enrollment is intended to proxy. Although administrative data sources provide a limited picture of these risk factors,49 research has demonstrated how claims can be used to measure constructs such as frailty,39 disability status,50 institutionalization,30 and via linked Census data, income, education, and social supports.6, 20 Incorporating these measures of medical and social risk may help to account for some state‐level differences in the characteristics of dual enrollees in ways that current risk‐adjustment methods do not. However, some socioeconomic and health status measures may not be collected with sufficient reliability in claims to support their use for risk adjustment.

Second, while dual enrollment in Medicare and Medicaid may be a useful proxy for otherwise unobserved risk factors, risk‐adjustment models could account for the possibility that the relationship between dual status and outcomes (eg, hospital readmissions, mortality, or spending) will be heterogeneous, rather than constant, across states. To quantify this heterogeneity in a risk‐adjustment model, an analyst could add interactions between a person‐level indicator of dual enrollment and a vector of state fixed effects to a risk‐adjustment model.41 An analyst might also consider modeling heterogeneity with respect to specific characteristics of states' Medicaid programs—for example, interacting a person‐level indicator of dual enrollment with a state‐level policy variable, such as the availability of a Medically Needy program or income eligibility limits for Medicaid, as has been done in prior work adjusting hospital readmission rates.20

Third, our findings demonstrate the limitations of risk‐adjustment methods that stratify providers by characteristics of their patients, particularly when these methods disregard relevant heterogeneity in the stratification variables. For example, CMS has begun to compare readmission rates among hospitals grouped by the proportion of Medicare beneficiaries enrolled in Medicaid. However, because these strata are defined nationally and contain hospitals from different states, CMS' approach could cause hospitals serving numerically equivalent proportions of dual enrollees to be compared to one another even if characteristics of their states' Medicaid populations differ. Methods that adjust for person‐level characteristics, and that include an adjustment for state‐level heterogeneity in the characteristics of dual enrollees, can more flexibly account for a richer set of risk factors than stratification.20

Our study had several limitations. First, because the MCBS does not collect detailed information on assets, we were unable to examine state variation in family resources or exclude individuals whose resources would make them ineligible for Medicaid. However, because of Medicaid's eligibility criteria, the vast majority of duals have minimal or no assets.51 Second, sample sizes in the MCBS were insufficient to support reliable comparisons of low‐income Medicare beneficiaries and dual enrollees in 15 states (including the District of Columbia) and may have limited our power to detect differences in population characteristics when we compared the remaining 36 states by their Medicaid policies. The excluded states were also more rural, which could have limited our ability to quantify state‐level differences in the characteristics of rural‐dwelling dual enrollees. However, the 15 excluded states did not differ appreciably on Medicaid policies from the 36 included states (Appendix S3). Moreover, the health and socioeconomic characteristics of low‐income duals varied to a similar extent in the urban (N = 24) and relatively rural (N = 12) states included in our analysis (Appendix S7). Third, our between‐state comparisons of dual enrollees could have been biased by unmeasured differences in state policies or population characteristics. Thus, our estimates should not be interpreted as measuring causal associations between state policies and the characteristics of dually enrolled populations.

6. CONCLUSION

Analysts should consider how dual enrollment in Medicare and Medicaid is used to proxy for underlying health and socioeconomic characteristics given between‐state differences in the characteristics of low‐income Medicare beneficiaries and state policies that affect which Medicare beneficiaries are eligible for Medicaid. Risk‐adjustment methods that use dual enrollment in Medicare and Medicaid to proxy for low SES and poor health should account for this state‐level heterogeneity to support more reliable assessments of risk‐adjusted performance.

Supporting information

 

 

Roberts ET, Mellor JM, McInerney M, Sabik LM. State variation in the characteristics of Medicare‐Medicaid dual enrollees: Implications for risk adjustment. Health Serv Res. 2019;54:1233–1245. 10.1111/1475-6773.13205