Outcomes of Medicaid home‐ and community‐based long‐term services relative to nursing home care among dual eligibles

Abstract

Objective

To provide the first plausibly causal national estimates of health outcomes for older dual‐eligible recipients of Medicaid HCBS relative to nursing home care and to explore possible mechanisms for the effect.

Data Sources

We use 2005 and 2012 Medicaid Analytic eXtract (MAX), a national compilation of Medicaid claims, merged with Medicare claims to identify hospital admissions, our main outcome variable.

Study Design

We model the effects of HCBS using a longitudinal instrumental variables framework. To address the endogeneity of HCBS receipt, we instrument for it using the county percentage of nonelderly long‐term care users who receive HCBS. The percentage of nonelderly users is highly predictive of HCBS use for an elderly beneficiary, but because the instrument was derived from a separate population, the exclusion restriction is unlikely to be violated.

Population Studied

1,312,498 older adults (65+) dually enrolled in Medicaid and Medicare and are using long‐term care. We also examine heterogeneity of effects by race/ethnicity and the presence of dementia.

Principal Findings

HCBS users have 10 percentage points higher (P < .01) annual rates of hospitalization than their nursing home counterparts when selection bias is addressed; rates of potentially avoidable hospitalizations are 3 percentage points higher (P < .01). These differences persist across races, dementia status, and intensity of HCBS spending.

Conclusions

Shifting Medicaid long‐term care funding for older adults from nursing homes to HCBS, while well‐motivated, results in the unintended consequence of substantially higher hospitalization rates for older dual eligibles. The quality and/or quantity of services may be inadequate for some HCBS recipients. Hospitalizations are costly to Medicare but also to the HCBS recipient in terms of stress and risks. Although consumer preferences to remain at home may outweigh poor outcomes of HCBS, the full costs and benefits need to be considered. HCBS outcomes—not just expansion—need more attention.

What This Study Adds

• Our study provides the first plausibly causal national estimates of health outcomes for older dual eligibles receiving HCBS relative to nursing home care.

• Dual eligibles receiving HCBS are 10 percentage points more likely than those in nursing facilities to be hospitalized. This is a key unintended consequence of HCBS that needs closer scrutiny.

• Preferences to stay at home need to be weighed against resulting higher hospitalization rates.

1. INTRODUCTION

Across the health care system, there is increasing interest in shifting care from expensive institutional settings to home‐ and community‐based settings. This interest has intensified during the COVID‐19 pandemic, during which nursing homes have disproportionately experienced cases and deaths, raising the relative risks of institutional care. ¹ , ² In long‐term care (also referred to as long‐term services and supports), state Medicaid programs have dramatically expanded the funding of home‐ and community‐based services (HCBS) as an alternative to institutional long‐term care. In 1996, only 19% of Medicaid long‐term care expenditures were for HCBS relative to nursing facilities; by 2016, HCBS expenditures accounted for 57%. ³ Because people generally prefer to age in their homes, and because institutions are expensive, the advisability of this shift toward home‐based care rarely elicits debate. However, little is known about the health outcomes of HCBS relative to nursing facility care.

States have expanded Medicaid funding for HCBS through state plans and through 1915(c) waivers. ³ State plan services constitute entitlements available to anyone in the state who qualifies. Waiver services may be county‐specific, with caps on enrollment. Under the waiver mechanism, Medicaid beneficiaries obtain access to HCBS only if their needs meet a nursing facility level of care, with the explicit goal providing an alternative to institutionalization. Policy makers take pains to avoid a “woodwork” effect where people who otherwise would not receive services are encouraged by the availability of HCBS to sign up. Some expansion to healthier individuals has nonetheless occurred, but the increased intensity of funding for sicker individuals—consistent with the goal of keeping people out of nursing facilities—has been the more pronounced trend. ⁴

Many waivers are meant to provide HCBS options for frail, elderly beneficiaries with long‐term care needs. In this study, we focus on the subset of the elderly who are enrolled in both Medicare and Medicaid, known as “dual eligibles,” who are often physically and cognitively impaired, are disproportionately from racial or ethnic minority groups, and have high use of health care services and costs. In 2013, duals accounted for 15% of Medicaid enrollment but 32% of Medicaid expenditures and 20% of Medicare enrollment but 34% of Medicare expenditures. ⁵ Medicaid programs may see HCBS options as a way to reduce spending on dual eligibles. At the same time, these beneficiaries often face challenges navigating care across the Medicaid and Medicare programs and have fewer social supports than the average Medicare or Medicaid beneficiary, making them vulnerable to poor outcomes.

The benefits of HCBS to long‐term care recipients and their families, especially in terms of preferences to age in place and improved quality of life, may be substantial. However, health outcomes may be worse under HCBS than in nursing facilities for several reasons. HCBS generally entails lower intensity care relative to the round‐the‐clock care available in a nursing facility. Furthermore, HCBS shifts some of the care burden from trained, paid staff to largely untrained family or friends who must fill the critical gaps in care intensity. Home environments may not be safe or appropriately designed to accommodate needs. Home care workers may face challenges implementing high‐intensity treatments in the home environment, and informal caregivers may not be well trained to handle clinical issues. Thus, HCBS could lead to worse health outcomes relative to nursing facility care.

The challenges of HCBS may have particularly large implications for racial and ethnic minorities, who are disproportionally represented among Medicaid long‐term care users. ⁶ Minority groups tend to use fewer institutional services and more home care and informal care. ⁷ , ⁸ , ⁹ , ¹⁰ , ¹¹ Additionally, when minorities use long‐term care services, they tend to receive lower quality care. ⁹ , ¹² , ¹³ , ¹⁴ Prior descriptive research indicates that spending on HCBS among blacks is significantly lower than among whites nationally, and hospitalization rates tend to be higher among black HCBS users than white HCBS users. ⁷

Any negative outcomes of HCBS use are also likely to be exacerbated for sicker care recipients, especially those with Alzheimer's disease and other dementias. Home‐based care might not be optimal to meet the complex needs of individuals with dementia, which could eventually increase the risk of negative outcomes. Also, caring for individuals with dementia can be more stressful than for other conditions. Costs to caregivers of persons with dementia, such as lost work productivity and caregiving‐related health problems, are substantial. ¹⁵ Caregiver stress exists even with nursing facility placement ¹⁶ but is especially burdensome in the home. As informal caregivers are a critical part of the care team under HCBS, the presence of dementia may exacerbate caregiver burden such that outcomes for the care recipient suffer. Descriptive evidence indicates that hospitalization rates are higher for HCBS users with dementia than those without dementia. ⁷

Despite increasing interest in home‐based care and the decades‐long shift in Medicaid funding toward HCBS, rigorous evidence on the outcomes of HCBS relative to institutional care is sparse. There have been evaluations of specific types of HCBS programs or demonstrations which found that HCBS use is beneficial to care recipients, but these results are not likely to be broadly generalizable to elderly, dual‐eligible HCBS users. ¹⁷ , ¹⁸ , ¹⁹ A recent report examines high‐cost HCBS users defined by Medicaid spending only, thereby focusing largely on under‐65 nondual Medicaid beneficiaries. ²⁰ Several studies have documented high rates of potentially avoidable hospitalizations among HCBS users, but did not provide comparisons. ²¹ , ²² In two key studies that compared rates of potentially preventable hospitalizations among elderly, dually eligible HCBS users relative to nursing facility residents using propensity score methods, results suggested that HCBS users were more prone to hospitalization than similar nursing facility residents ²³ , ²⁴ ; however, these studies were limited to individuals in seven states in 2003‐2005 and did not examine differences by key subgroups. Finally, one study used more recent, national data and compared rates of hospitalization by setting, but the analysis was purely descriptive and did not account for potential selection bias into care setting. ⁷

Our study builds on prior literature by providing rigorous estimates of the outcomes of HCBS use relative to nursing home care that control for selection bias, using national claims data and examining several key subgroups of dual eligibles. Furthermore, we examine the intensity of HCBS spending as a possible mechanism. Our study informs policy makers in evaluating the effects of current policy that increases access to HCBS relative to nursing facility care.

2. METHODS

2.1. Data and variables

We use the 2005 and 2012 national Medicaid Analytic eXtract (MAX) linked with Medicare claims to build our analysis dataset. We chose this span of time to allow for sufficient changes in HCBS policy to be modeled. The MAX is a claims‐based dataset created by the Centers for Medicare and Medicaid Services from data submitted quarterly by states. The first year for which HCBS is coded consistently across states using the community long‐term care (CLTC) flags is 2005. We use 2012 because at the time of our analysis it was most recent year all states were included in the MAX, but we conduct a robustness check using the subset of states available for 2014 to assess whether results are different over time. We identify Medicaid long‐term care program enrollment, service use, and expenditures from the MAX person summary and claims files. We then link to the Medicare Provider Analysis and Review (MedPAR) file to identify hospitalizations at the individual level. The Master Beneficiary Summary Files (MBSF) are used to identify demographic characteristics (age, sex, race, and ethnicity), Medicare managed care enrollment, original reason for Medicare eligibility, and chronic conditions.

To identify the study target population, we first identify long‐term care users who are dually enrolled in Medicare and Medicaid for all of 2005 or 2012 and are aged 65 or older. Following the prior literature, ²¹ , ²⁵ we identify long‐term care users through Medicaid waiver enrollment and fee‐for‐service (FFS) claims. Institutional long‐term care use is identified as having any FFS claim in the MAX Long Term Care Files. HCBS is identified from enrollment in a 1915(c) HCBS waiver for at least one month and/or use of waiver services and/or state plan services using FFS claims. Although we require only one month of service use, the vast majority of the sample (83%) used either HCBS or institutional care for all 12 months, and we conduct a robustness check excluding those with fewer months. We do not include as HCBS users those who used transportation services only, given conceptual ambiguity about whether transportation should be included and because some transportation services were related to ambulance care and could create reverse causality issues. Among older duals using long‐term care, we then exclude long‐term care users (less than 10%) who receive both institutional care and HCBS during a calendar year, consistent with prior literature, ²³ as these individuals are measurably sicker, likely undergoing a transition, and generally have higher hospitalization rates than those using HCBS or institutional care alone. (We conduct a robustness check including them and assigning them to HCBS.) Our target sample is, thus, older duals using HCBS or institutional care on an ongoing basis.

We make several exclusions to the main analysis sample. We limit to long‐term care users who have complete demographic information and who were alive for the whole year to equalize exposure time, and to those whose county can be identified.

The remaining two categories of exclusions are exclusions by county or by state. The county‐level exclusions are required in order to implement our IV design, for which we need data on the instrument, the treatment, and the outcomes for both time points, and we need the instrument to work in a consistent direction (higher levels of the instrument make individual‐level HCBS use more likely). All counties in Alaska and Connecticut are omitted for the latter reason. A few additional state‐level exclusions are required in order to avoid sample selection issues between the two years; states that have large managed long‐term care populations (Arizona, Hawaii, New Mexico, Tennessee) or underwent large changes in these programs (Arkansas, New Jersey, Ohio) may have a selected sample remaining in fee‐for‐service or a differentially selected sample between the two years. Our final analysis sample includes 664 202 observations from 2005 and 648 296 from 2012 for a total of 1 312 498 observations in a balanced panel of 1 171 counties. See Appendix Table S1 for a detailed tabulation of the sample restrictions.

The key outcome we examine is admission to the hospital, including separate measures for all hospitalizations and those that are potentially avoidable. Hospital admissions have long been considered a negative marker of quality, especially for frail older adults who may experience adverse consequences from the transfer itself and a risk of nosocomial infection. They have also come under increased scrutiny for the costs imposed on Medicare. ²⁶ We identify hospitalizations using the MedPAR files and as potentially avoidable using the Agency for Healthcare Research and Quality (AHRQ) Prevention Quality Indicator algorithms. ²⁷ We present both outcomes because identifying potentially preventable hospitalizations and potentially unavoidable ones is subject to error, even with established algorithms. We define each hospitalization measure dichotomously to indicate whether a hospitalization occurred during the year. We include only hospitalizations that occurred after the onset of long‐term care use. Because the vast majority of the sample uses services the entire year, an annual hospitalization measure is reasonable; we conduct a robustness check excluding those who used services for fewer months.

Key variables for stratification in the analyses are race, dementia diagnosis, and intensity of HCBS spending. Beneficiary race is identified using the MBSF RTI race variable. ²⁸ An individual is identified as having dementia if he/she met the claims criteria for the Alzheimer's disease and related disorders or senile dementia chronic conditions flag in either the MAX or Medicare chronic conditions supplement files. Medicaid expenditures for HCBS are calculated from the MAX state plan and HCBS waiver claims. We use the medical component of the Consumer Price Index to adjust for inflation and the Medicare wage index to adjust for geographic differences in purchasing power. We then divide spending across the two years into quartiles to examine outcomes by the intensity of HCBS spending.

Finally, we include in our analyses a rich set of covariates, including age, sex, race, original reason for Medicare eligibility, and a long list of indicator variables for comorbidities.

2.2. Research design

We use a longitudinal instrumental variables framework. Because HCBS users may be different, presumably healthier, than nursing facility users in unobserved ways, a research design that did not take this potential endogeneity into account could result in biased estimates. To address the endogeneity of HCBS receipt, we instrument for it using the county‐level percentage of nonelderly Medicaid long‐term care users who receive HCBS. This instrument likely reflects changes in both Medicaid long‐term care policy and the population‐level affinity for home‐based care over time, as the percent of people using HCBS across age groups tends to be correlated. ²⁹

A valid instrumental variable predicts exogenous variation in treatment that is not correlated with the outcomes of interest except through treatment. The change in the percentage of nonelderly HCBS users is highly predictive of the change in HCBS use for an older beneficiary, but because the instrument was derived from a separate population, the exclusion restriction is unlikely to be violated; in other words, a change in the percent of HCBS use among the nonelderly is unlikely to be directly related to hospitalizations for older adults. This is especially true in that nonelderly long‐term care users tend to have different health conditions leading to their long‐term care use and hospitalizations, such as intellectual and developmental disabilities or mental illness. ³⁰ One might worry that county‐level health care infrastructure or quality might constitute a separate pathway from the percent of nonelderly adults using HCBS to hospitalizations among older adults, but the probability of this violation is arguably small in a county fixed‐effects framework; the 2005‐2012 percent change in nonelderly adults using HCBS would have to be correlated with a change in the health infrastructure or quality during the same time period which would have to be correlated with a change in hospitalization rates among older adults using long‐term care. In a robustness check, we use as an instrument changes in the county‐level percentage of HCBS use among older adults; this is a common application of an area‐level IV to predict individual treatment within the same population. While conceptually slightly weaker, the first‐stage power of this alternative IV to predict individual HCBS use among older adults is stronger, not surprisingly.

For comparison, we begin by estimating the effects of long‐term care setting on hospitalization using a county fixed‐effects regression framework without instrumental variables. We estimate the following equation:

$$\mathrm{Logit}(Pr(Y_{\mathit{ict}}=1))={\beta }_1{\mathrm{HCBS}}_{\mathit{ict}}+{\delta }_2{X}_{\mathit{ict}}+{\mathrm{County}}_c+{\mathrm{Year}}_t+{ϵ}_{\mathit{ist}}$$ (1)

The subscripts i, c, and t represent individual, county, and time (year). $X_{\mathit{ict}}$ represents controls for the following individual‐level characteristics: age, sex, race, original reason for Medicare enrollment, and chronic conditions. In the absence of endogeneity of long‐term care setting, ${\mathrm{HCBS}}_{\mathit{ict}}$, and hospital use ($Y_{\mathit{ict}}$), the coefficient ${\beta }_1$ can be interpreted as the average treatment effect of HCBS versus nursing home care on hospitalization within the population of older adult dual eligibles that use either HCBS or nursing home care. While this fixed‐effects regression controls for time‐invariant county‐ and state‐level factors and time effects that affect all counties similarly, it is still subject to omitted variables bias if there are unobserved, time‐varying characteristics of counties or individuals that impact both the individual's choice of long‐term care setting and their probability of hospitalization.

To address this possible endogeneity of long‐term care setting, we next implement the instrumental variables approach using two‐stage residual inclusion methods, given our dichotomous treatment and outcomes. ³¹ Again, we estimate a county‐level fixed‐effects regression, which controls for all time‐invariant omitted variables at the county and state levels of the following form:

Stage 1: $\mathrm{Logit}(Pr({\mathrm{HCBS}}_{\mathit{ict}}=1))={\alpha }_1{Z}_{\mathit{ict}}+{\alpha }_2{X}_{\mathit{ict}}+{\mathrm{County}}_c+{\mathrm{Year}}_t$ (2)

Stage 2: $\mathrm{Logit}(Pr(Y_{\mathit{ict}}=1))={\beta }_1{\mathrm{HCBS}}_{\mathit{ict}}+{\delta }_2{X}_{\mathit{ict}}+{\mathrm{County}}_c+{\mathrm{Year}}_t+\widehat{{\phi }_{\mathit{ist}}}$

The instrument, $Z_{\mathit{ict}}$, is used in the first stage to predict individual‐level HCBS use (${\mathrm{HCBS}}_{\mathit{ict}}$), controlling for individual‐level characteristics, $X_{\mathit{ict}}$, and county and time fixed effects, ${\mathrm{County}}_c\mathrm{and}{\mathrm{Year}}_t,$ respectively. The residual from the first‐stage regression, $\widehat{{\phi }_{\mathit{ist}}}$, is then included in the second stage regression of the outcome on treatment, ${\mathrm{HCBS}}_{\mathit{ict}}$, including the full set of controls. The coefficient on treatment in the second stage, ${\beta }_1$, is the estimated, plausibly causal effect of HCBS versus nursing home use on hospital utilization. This means that the estimated effect of individual‐level HCBS use is identified for those individuals who used HCBS rather than nursing facility care due to the growth of Medicaid HCBS in their county of residence between 2005 and 2012. In other words, it is the “local average treatment effect ³² ” for the subset of long‐term care users induced to use HCBS due to state and county HCBS expansions, as opposed to those who would have used HCBS or would have used a nursing facility regardless of HCBS expansion. This local average treatment effect is an especially relevant effect for policy makers, as it informs outcomes for those most affected by expansion of HCBS policies.

3. RESULTS

Over the two analysis years, more than two‐thirds of the sample used only HCBS during the year and the rest used only nursing facility services (Table 1). As expected, those using HCBS tend to be younger, healthier, and less likely to be white than institutional users. These observed differences between nursing home users and HCBS users suggest that there may also be unobserved differences between the two groups, underscoring the need for an instrumental variables analysis.

TABLE 1.

Sample characteristics by care setting and year

	Overall	2005		2012
		Inst.	HCBS	Inst.	HCBS
N	1 312 498	198 122	466 080	187 372	460 924
Age (%)
65‐69	14.0	7.6	16.0	10.7	16.2
70‐74	17.0	10.9	19.6	12.6	18.6
75‐79	19.3	15.8	21.9	14.8	20.0
80‐84	19.8	21.0	20.2	18.3	19.6
85‐89	16.0	20.8	13.0	20.2	15.1
90+	13.9	23.9	9.3	23.5	10.5
Race (%)
White	49.8	70.4	42.8	66.1	41.3
Black	21.1	19.9	21.8	20.9	21.0
Hispanic	15.1	6.5	18.2	9.1	18.1
Asian	12.2	2.5	15.0	3.1	17.3
Others	1.8	0.7	2.2	0.8	2.3
Gender—Female (%)	73.9	77.2	74.4	73.5	72.2
Eligibility—Aged (%)	17.0	4.2	22.3	4.7	22.3
Dementia = Yes (%)	47.4	79.1	28.9	84.1	37.6
Anemia = Yes (%)	46.7	51.8	40.6	60.5	45.1
Asthma = Yes (%)	7.9	3.4	8.6	4.6	10.5
Atrial_Fibrillation = Yes (%)	9.6	10.7	8.1	13.2	9.3
Cataract = Yes (%)	22.8	25.0	24.2	28.0	18.4
Chronic Kidney Disease = Yes (%)	22.6	15.5	15.8	32.5	28.4
Heart Failure = Yes (%)	37.5	43.2	37.5	41.7	33.4
COPD = Yes (%)	21.2	19.7	20.8	23.7	21.4
Depression = Yes (%)	26.4	33.5	17.1	46.7	24.6
Diabetes = Yes (%)	46.5	37.8	44.5	47.1	51.9
Glaucoma = Yes (%)	12.9	8.6	14.1	10.4	14.6
Hyperlipidemia = Yes (%)	44.3	16.3	44.1	39.2	58.6
Hyperplasia = Yes (%)	6.0	2.6	5.2	6.0	8.4
Hypertension = Yes (%)	78.1	67.9	76.4	81.6	82.9
Hypothyroidism = Yes (%)	17.0	19.5	12.6	25.3	17.0
Ischemic Heart Disease = Yes (%)	48.5	42.2	52.2	44.3	49.3
Osteoporosis = Yes (%)	13.0	12.5	11.1	14.7	14.4
RA OA = Yes (%)	48.8	38.4	48.8	45.5	54.5
Stroke = Yes (%)	10.2	18.2	8.0	15.8	6.8
Others = Yes (%)	8.8	7.7	8.7	8.6	9.5

Abbreviations: COPD, chronic obstructive pulmonary disease; HCBS, home‐ and community‐based services; Inst, institutional care; RA OA, rheumatoid arthritis/osteoarthritis.

The results of our longitudinal, county fixed‐effects estimation of Equation (1) indicate that older dual eligibles receiving HCBS are more likely to be hospitalized than those using nursing facility care by 7 percentage points; this magnitude grows to 10 percentage points once selection bias is accounted for using instrumental variables (Table 2). HCBS users are 3 percentage points more likely to experience a potentially avoidable hospital admission than users of institutional care.

TABLE 2.

Marginal effects of home‐ and community‐based services use on hospitalizations

Variables	Any hospitalization		Potentially avoidable hospitalization
	Without IV	With IV	Without IV	With IV
	ME (95% CI)	ME (95% CI)	ME (95% CI)	ME (95% CI)
Care setting = HCBS	0.07 (0.05, 0.10)**	0.10 (0.08, 0.12)**	0.03 (0.02, 0.04)**	0.03 (0.02, 0.04)**
Age
65‐69 (reference)
70‐74	−0.03 (−0.03, −0.02)**	‐0.02 (−0.03, −0.02)**	‐0.00 (−0.01, 0.00)**	0.00 (−0.01, −0.00)**
75‐79	−0.04 (−0.04, −0.03)**	−0.04 (−0.04, −0.03)**	−0.00 (−0.01, 0.00)**	−0.00 (−0.01, −0.00)**
80‐84	−0.04 (−0.05, −0.04)**	−0.04 (−0.05, −0.03)**	−0.00 (−0.01, 0.00)	−0.00 (−0.00, 0.00)
85‐89	−0.04 (−0.05, −0.03)**	−0.04 (−0.05, −0.03)**	0.00 (−0.00, 0.01)	0.00 (−0.00, 0.00)
90+	−0.04 (−0.05, −0.03)**	−0.04 (−0.04, −0.03)**	0.00 (0.00, 0.01)	0.00 (0.00, 0.01)**
Gender = Female	0.01 (0.01, 0.02)**	0.01 (0.01, 0.02)**	0.01 (0.01, 0.01)**	0.01 (0.01, 0.01)**
Race
White (reference)
Black	0.03 (0.02, 0.05)**	0.03 (0.01, 0.05)**	0.01 (0.01, 0.02)**	0.01 (0.01, 0.02)**
Hispanic	0.03 (0.01, 0.06)*	0.03 (0.00, 0.06)*	0.01 (0.00, 0.03)*	0.01 (−0.00, 0.03)
Asian	−0.03 (−0.05, −0.02)**	−0.04 (−0.05, −0.02)**	−0.01 (−0.02, −0.01)**	−0.01 (−0.02, −0.01)**
Others	−0.02 (−0.03, −0.01)**	−0.03 (−0.04, −0.02)**	−0.01 (−0.01, −0.01)**	−0.01 (−0.01, −0.01)**
Eligibility = Disabled/Others	−0.02 (−0.03, −0.02)**	−0.03 (−0.03, −0.02)**	−0.01 (−0.01, −0.01)**	−0.01 (−0.01, −0.00)**
Dementia	0.05 (0.04, 0.05)**	0.05 (0.05, 0.06)**	0.02 (0.02, 0.02)**	0.02 (0.02, 0.02)**
Mean of the dependent variable (%)	29.4	29.4	8.5	8.5
First‐stage F‐statistic		46.0		46.0

Further adjusted by 1) month of first service, 2) year fixed effects, 3) county fixed effects, and 4) all comorbidities listed in Table 1. Standard errors in IV models not adjusted for inclusion of predicted residual in two‐stage residual inclusion procedure and may be slightly underestimated.

Abbreviations: CI, confidence interval; HCBS, home‐ and community‐based services; IV, instrumental variable = county percent HCBS use among nonelderly; ME, marginal effects.

^* P < .5; **P < .01.

Figure 1 shows that the hospitalization penalty from receiving HCBS is similar across race groups and by dementia status once selection bias is taken into account. Although blacks are slightly more likely than whites to experience a hospitalization from a nursing facility, the additional risk of hospitalization from HCBS is similar. Likewise, those with dementia are more likely than those without dementia to experience a hospitalization from a nursing facility, but both increase their risk by similar levels from receiving HCBS.

FIGURE 1.

Marginal effects of home‐ and community‐based services use on hospitalizations by race and Dementia. HCBS, home‐ and community‐based services users; Inst, institutional care users; PQI, prevention quality indicator = potentially avoidable hospitalization

Figure 2 examines a possible mechanism for higher hospitalization rates in HCBS, the intensity of care as measured by HCBS spending. In each case, we estimated our instrumental variables regression using a single quartile of HCBS spending as the treatment group compared with all nursing home residents. We expected the differences in hospital admission rates to be higher when the intensity of HCBS spending was lower, if intensity of HCBS spending was driving our effect. However, we find no meaningful differences in hospitalization rates across HCBS spending quartiles.

FIGURE 2.

Instrumental variables estimates of marginal effects of home‐ and community‐based services use on hospitalizations by spending quartile. Q1 is the lowest spending quartile and Q4 is the highest, after adjusting for the medical component of the Consumer Price Index and the Medicare wage index. Cut points for the quartiles are 4 881, 10 431, and 18 265. HCBS, home‐ and community‐based care users; Inst, institutional care users; PQI, prevention quality indicator = potentially avoidable hospitalization

Finally, Table 3 summarizes the results of multiple robustness checks, presented in more detail in the appendix. First, we re‐estimate our main 2005‐2012 analysis using only states for which we have 2014 data and then estimate our model using 2014 as our endpoint. These results show that the effect size is substantially smaller in the 2005‐2014 analysis than in our main model, but it is driven largely by the subset of states used, because the 2012 estimate on this subset is also substantially smaller than our national estimate. Thus, the HCBS effect does not appear to be changing in a meaningful way from year to year, indicating that our use of 2012 is reasonable. Second, we re‐estimate our model using a different instrumental variable, the county‐level percent of older adults using HCBS, which gives a stronger first‐stage F‐statistic (2834.9). Nonetheless, the estimated marginal effect is very similar. Finally, we examine several of our sample restrictions. We re‐estimate our model on 83% of the sample that used services during the entire year to check whether differences in exposure time from the remaining 18% might bias the results; the estimate is very similar. Including people who use both HCBS and institutional care during the year produces a smaller estimate, while including people who died produces an estimate very similar to our main model. Overall, these tests support the robustness of our estimates.

TABLE 3.

Robustness checks: instrumental variables estimates of marginal effects of home‐ and community‐based services use on hospitalizations under alternative samples/assumptions

	Any hospitalization	Potentially avoidable hospitalization
	ME (95% CI)	ME (95% CI)
Base model for comparison	0.10 (0.08, 0.12)**	0.03 (0.02, 0.04)**
Choice of years
2005 and 2012 results using only states available in 2014 a	0.05 (0.00, 0.09)*	0.01 (−0.00, 0.03)
2005 and 2014 results for subset of states a	0.03 (0.01, 0.05)*	0.02 (0.01, 0.04)*
Alternative IV b	0.10 (0.08, 0.12)**	0.03 (0.02, 0.04)**
Sample
Used long‐term care the entire year	0.10 (0.08, 0.12)**	0.03 (0.02, 0.04)**
Include both HCBS and institutional care users c	0.07 (0.05, 0.10)**	0.03 (0.02, 0.04)**
Include people who died	0.11 (0.09, 0.13)**	0.04 (0.03, 0.05)**

Standard errors in IV models not adjusted for inclusion of predicted residual in two‐stage residual inclusion procedure and may be slightly underestimated.

Abbreviations: CI, confidence interval; HCBS, home‐ and community‐based services; IV, instrumental variable; ME, marginal effects.

^aIncluded states = CA, GA, IA, ID, LA, MI, MN, MO, MS, PA, UT, VT, WV, and WY; adjusted for age, gender, race, eligibility with year, and county‐level fixed effects.

^bWe used the county‐level percentage of HCBS use among long‐term care users aged 65 and older as our instrumental variable (F‐value = 2834.9).

^cWe included those who used both HCBS and institutional care, and treated them as HCBS users.

^* P < .5; **P < .01.

4. DISCUSSION

We show that among older, dual‐eligible long‐term care users who are at the margin of using institutional care versus HCBS, those using HCBS experience a substantially higher probability of hospital admission, whether measured as any hospital admission (10 percentage points higher) or one that is potentially avoidable (3 percentage points higher). High rates of hospital admission among HCBS users are consistent with prior descriptive evidence, but the magnitude of the effect on any hospitalization is substantially larger when selection bias is addressed. Our results are robust to a variety of assumptions and changes.

Whereas prior descriptive evidence showed differences by race and dementia status, with Blacks and persons with dementia experiencing higher rates of hospital admission under HCBS, our instrumental variables analysis indicates that the HCBS effect is effectively the same across races and dementia status, once observed and unobserved differences in health status are accounted for. The underlying rates of hospitalization from institutional settings are different (with blacks and persons with dementia experiencing higher rates), so an equal differential from using HCBS still results in uneven hospitalization rates across these groups. Blacks also tend to rely on HCBS more than whites, so if these individual‐level results are extrapolated to the population, then the HCBS effect will add to more additional hospitalizations from HCBS among Blacks than whites.

Our study has several limitations. First, we use data from 2005 to 2012. However, the underlying relationships we find are generally consistent with descriptive studies from much earlier, and our robustness check suggests that using 2014 data would not have changed our conclusions. Second, in claims data we cannot identify several potentially important drivers of the success of HCBS, such as the extent of family and caregiver support; we can balance them through our instrumental variables design but cannot study them directly; these should be the subject of future research. Third, while we examine several key areas of potential heterogeneity—race, dementia, and intensity of HCBS spending—there are others that could be important. For example, the effects of HCBS relative to nursing homes may differ by the quality and availability of nursing homes and HCBS providers in a county. Fourth, we examine one key outcome that is easily identified in claims data—hospital admissions—but other outcomes may be equally important, such as functional status and quality of life. Fifth, our instrumental variables design estimates a local average treatment effect based on beneficiaries at the margin of using HCBS or institutional care. Our estimates therefore do not apply to individuals who use HCBS but are healthy enough that they would otherwise get no formal care, or to those who are sick enough that HCBS is not an option. Finally, while our study is based on rigorous econometric methods designed for causal inference and we believe our instrument satisfies key assumptions, any violations of our assumptions could cause bias in our results.

There are several implications of the patterns we find. First, high hospitalization rates among HCBS users may mean that HCBS is inadequate in quality and/or quantity. The inadequacy may be due to too few or not the right services, insufficient frequency of care, low‐quality providers, or inadequate attention to coordinating services, which may be harder with HCBS than in a nursing facility setting where staff are available to help around the clock. This would imply that an additional influx of resources could alleviate the problem, especially resources that might enable greater access to acute‐care services at home. However, we found no meaningful difference in the HCBS effect across quartiles of HCBS spending, consistent with some prior literature. ³³ The intensity of HCBS spending therefore may not be relevant to reducing hospital admissions, or the differences in spending from quartile to quartile in our data are not large enough to make a meaningful difference.

Second, and related, the nature of HCBS may mean that hospital admissions are difficult to avoid even when the provided services are generous. Home care providers, and the informal caregivers who provide supplementary services and support, will probably never have the laboratory, diagnostic, and staffing infrastructure that might enable a nursing facility resident to avoid hospitalization. Even in the nursing home setting, a decline in condition often results in a transfer to the hospital, but the tipping point is likely to be lower at home. This is especially likely if family members are making the decision, whose threshold for transfer may be lower than that of medical professionals. When a care recipient in the home setting exhibits a decline in health, a transfer to the hospital is likely to seem like the most reasonable course of action.

Third, any discussions of whether HCBS saves money should consider Medicare hospital spending, rather than just Medicaid spending. Furthermore, accounting of full social costs of these programs must also include monetary and nonmonetary costs to caregivers and to care recipients for adverse outcomes. Hospitalizations present the risk of physical deterioration and iatrogenic disease to care recipients. Relative to facility care, HCBS places a greater responsibility on families to support the formal care being received, and this mostly unpaid support is likely to be one reason that Medicaid finds it cheaper to fund HCBS than institutional care. Yet, it is well‐established that caregivers incur substantial costs in the form of reduced labor market participation and poorer health. ³⁴ , ³⁵ At the same time, measurement is still in need of substantial development to better capture the nonmonetary benefits to HCBS, such as increased satisfaction from remaining in one's home.

Our study provides compelling evidence that for older dual eligibles, Medicaid's expansion of funding for HCBS as an alternative to nursing facility care needs more careful scrutiny, despite its intuitive appeal. If HCBS is promoted as a preferred setting of long‐term care to Medicare and Medicaid beneficiaries who need a high intensity of care or who lack appropriate support at home, unintended consequences can result. The intensity of services may be inadequate for some HCBS recipients, and careful targeting of services should consider the potential for these unintended consequences. In the end, strong preferences to stay at home and increased quality of life from HCBS may outweigh these negative consequences, but a holistic view of policy must consider both.

Supporting information

Supplementary Material

Appendix S1

Konetzka RT, Jung DH, Gorges RJ, Sanghavi P. Outcomes of Medicaid home‐ and community‐based long‐term services relative to nursing home care among dual eligibles. Health Serv Res. 2020;55:973–982. 10.1111/1475-6773.13573