ACO Attribution and Post-Acute Skilled Nursing Facility Outcomes for People Living with Dementia

Abstract

Objectives:

Under the Accountable Care Organization (ACO) model, reductions in healthcare spending have been achieved by targeting post-acute care, particularly in skilled nursing facilities (SNFs). People with Alzheimer’s disease and related dementias (ADRD) are frequently discharged to SNF for post-acute care and have may be at particular risk for unintended consequences of SNF cost reduction efforts. We examined SNF length of stay (LOS) and outcomes among ACO-attributed and non-ACO attributed ADRD patients.

Design:

Observational serial cross-sectional study.

Setting and Participants:

20% national random sample of fee-for-service Medicare beneficiaries (2013–2017) to identify beneficiaries with a diagnosis of ADRD and with a hospitalization followed by SNF admission (n=263,676).

Methods:

Our primary covariate of interest was ACO (n=66,842) and non-ACO (n=196,834) attribution. Hospital readmission and death were measured for 3 time periods (<30, 31–90, and 91–180 days) following hospital discharge. We used two-stage least-squares regression to predict LOS as a function of ACO attribution, and patient and facility characteristics.

Results:

ACO-attributed ADRD patients have shorter SNF LOS than their non-ACO counterparts (31.7 v. 32.8 days; p<0.001). Hospital readmission rates for ACO v. non-ACO differed at ≤30 days (13.9% v. 14.6%; p<0.001) but were similar at 31–90 days and 91–180 days. No significant difference was observed in mortality post-hospital discharge for ACO v. non-ACO at ≤30 days, however slightly higher mortality was observed at 31–90 days (8.4% v. 8.8%; p=0.002) and 91–180 days (7.6% v. 7.9%; p=0.011). No significant association was found between LOS and readmission with small effects on mortality favoring ACOs in fully adjusted models.

Conclusions and Implications:

Being an ACO-attributed patient is associated with shorter SNF LOS but is not associated with changes in readmission or mortality after controlling for other factors. Policies that shorten LOS may not have adverse effects on outcomes for people living with dementia.

BRIEF SUMMARY:

Policies that impact post-acute care could have negative consequences for people with ADRD. ACO-attributed ADRD patients had a shorter SNF stays but readmission and mortality were no worse compared with non-ACO attributed patients.

INTRODUCTION:

Accountable care organizations (ACOs) focus on delivering high-quality care and reducing healthcare spending. Effects of ACOs in aggregate show small positive outcomes on readmissions^1–3 and spending,⁴ often related to post-acute care use.^5,6 However, outcomes vary across specific diseases.^7–12 Post-acute care, particularly in a skilled nursing facility (SNF), has been an area of interest for ACOs, as hospitals’ participation in an ACO has been associated with lower readmission rates, SNF length of stay (LOS), and Medicare spending on SNF.^1,13,14

People with Alzheimer’s disease and related dementias (ADRD) are frequently hospitalized and discharged to SNF and may be differentially impacted by ACO actions affecting the SNF setting. One-third of all fee-for-service Medicare beneficiaries receiving post-acute care have a diagnosis of ADRD, and more than 7 in 10 receive this care in a SNF.¹⁵ ADRD has been associated with longer Medicare-covered SNF stays,¹⁶ and high-intensity functional exercise is beneficial in improving balance and strength for older individuals with dementia.^17–20 However, ACOs’ influence on SNF use may reduce the LOS or type of therapy for their attributed beneficiaries, which could have a negative impact on health outcomes for people with dementia. Hence, examining SNF outcomes associated in ACOs for ADRD patients who have unique needs is imperative. The effect of ACO attribution on rehabilitation for patients with ADRD, who may require longer periods of rehabilitation, remains unknown.

In addition to interest in ACO outcomes themselves, newer payment models also may reduce length of rehabilitation making the issue raised for people with ADRD an ongoing concern. The newly implemented Patient-Driven Payment Model (PDPM) for SNF has changed SNF payment from being time-based (therapy or nursing minutes delivered) to using disease-based risk adjustment payments. The change in the payment model has led to a decline in the total number of therapy minutes delivered.^21,22 Studies using length of stay in SNF rehabilitation as a proxy for the amount of rehabilitation received can inform payment model decision-making where vulnerable groups, such as people with ADRD, are of concern. New payment models that influence the amount and type of rehabilitation provided post-hospital stay in a SNF could negatively impact individuals with ADRD. We aim to learn from prior policy interventions (i.e., ACO) whether people with ADRD experience worse outcomes related to the incentives that reduce post-acute care intensity. Our overarching question is: Under the ACO model, do people with ADRD receive less SNF rehabilitation measured by the length of stay (LOS) and have worse outcomes, hospital readmission and mortality, than in a non-ACO context?

METHODS

Data Sources and Study Population

We used a national random 20% sample of Medicare beneficiaries (2013–2017), which has national data on patients who go to hospitals and SNFs. From these data, we obtained patient demographics, enrollment, comorbidity, hospital stay characteristics, provider readmission decisions, mortality, timing and location of stays. Additionally, we used the annual ACO provider file.

The study population included Medicare beneficiaries enrolled in fee-for-service (FFS) Medicare, who had a hospitalization followed by a SNF stay, and had a diagnosis of ADRD. Beneficiaries had to be continuously enrolled in Parts A & B FFS Medicare 12 months prior and six months after an index hospitalization. Index hospitalizations were included if they did not occur within 30 days of any other hospital discharge and also had to be followed by a SNF admission within one day of the index hospitalization discharge from 2013 to 2017. We also restricted to beneficiaries who turned 66 by the end of the year, who were discharged alive, not discharged against medical advice, and not discharged the same day that they were admitted. We used a previously validated claims algorithm applied to Medicare Provider and Analysis Review (MedPAR), Outpatient, Home health, Hospice, and Carrier Research Identifiable Files (to identify beneficiaries with an ADRD diagnosis within 365 days prior to the index hospitalization^23,24. For beneficiaries with more than one stay meeting these criteria, we included their first stay in the given hospital discharge year of interest. We excluded the few cases with non-positive spending values for the SNF stay. Our final sample consisted of 263,676 ADRD beneficiaries. A visual illustration of the cohort construction process is provided in the supplemental file Figure S1.

Exposure and Outcome Variables

Accountable Care Organization (ACO) Status

Our primary covariate of interest was an indicator representing whether a patient was attributed to an ACO versus a non-ACO. Beneficiaries were attributed to an ACO using the methodology from the Centers for Medicare & Medicaid Services (CMS) Shared Savings and Losses and Assignment Methodology (Version 4) updated annually.²⁵ The first step in the assignment process was to identify primary care services using a set of Healthcare Common Procedure Coding System (HCPCS) codes delivered by primary care physicians (PCPs) and specialists. Charges from these services were then aggregated for PCPs and specialists separately to the tax identification number (TIN) level, which identifies distinct organizations in ACOs and non-ACOs. Beneficiaries were then assigned to a TIN, belonging to an ACO or non-ACO, based on primary care services delivered by PCPs. If a beneficiary did not have primary care services from their PCP, then they were assigned through primary care services provided by specialists.

Patient Outcomes

We calculated two critical patient outcome measures for the 30-day, 90-day, and 180-day periods following the index hospital discharge. We calculated whether each beneficiary was readmitted to a hospital within the given periods following hospital discharge. We also determined whether each beneficiary died within each given period following hospital discharge using mortality data from the Master Beneficiary Summary File (MBSF).

Covariates

Patient sex, race/ethnicity, age, and dual eligibility were obtained from the MBSF. We created a hierarchical condition category (HCC) measure for the year prior to admission to adjust for beneficiary comorbidity. This measure is commonly used for risk adjustment in CMS programs and incorporates 72 unique health conditions that are identified using ICD diagnosis codes.²⁶ For each patient in the data, we obtained all relevant diagnoses across all their inpatient and outpatient claims for each year of the analysis. We then applied code produced by CMS to construct a single score for each patient in the year of their index hospitalization. Index hospital LOS and whether the stay included any intensive care unit days were determined from MedPAR.

Analyses

We first examined for differences in age, race/ethnicity, sex, dual eligibility, comorbidities (based on HCC), and hospital stay characteristics in our study sample between beneficiaries attributed to ACOs compared to beneficiaries not attributed to ACOs using descriptive statistics.

Next, we applied an instrumental variable approach and used two-stage least squares to test for an association of ACO attribution with readmission and mortality through its effect on SNF LOS. Specifically, to understand how changes in SNF LOS may affect patient hospital readmission and mortality, we used ACO status as an instrument for the endogenous SNF LOS variable. We first ran a least-squares regression to predict LOS as a function of ACO attribution and patient characteristics (age, sex, race/ethnicity, dual eligibility, HCC community score), hospital stay variables (whether the patient had an ICU stay, index hospital LOS) as well as quarter-year fixed effects, which is also the first stage of the two-stage least squares analysis. This first-stage regression is shown in Table S1 of the online supplementary files. In the second stage, we tested the hypotheses that longer LOS affects readmission and mortality rates. Within each model, we included beneficiary characteristics (e.g., sex, race/ethnicity, age, dual eligibility, comorbidity), hospital LOS, whether the beneficiary had any intensive care unit days during their index hospital admission, as well as SNF fixed effects to account for non-time varying SNF-specific effects and quarter-year fixed effects. As a sensitivity test, we conducted the analysis without the SNF fixed effects. The second stage regression is shown in Table S2 of the online supplementary files. This study was approved by the University of Michigan Institutional Review Board.

RESULTS

Patient Characteristics

Our final sample consisted of 263,676 patients with ADRD; 66,842 patients were attributed to an ACO and 196,834 to a non-ACO. The sample was primarily female (66%), and the average age was 84.5 years. The racial composition of the sample was 83% White, 10% Black, 5% Hispanic, and 3% of other race. Twenty-seven percent of patients were dually eligible, and 27% had an ICU stay during the index hospitalization. The average HCC score was 3.11, and the average LOS for the index hospitalization was 5 days (Table 1).

Table 1:

Medicare Fee-for-Service Beneficiaries with Alzheimer’s Disease and Related Dementias Admitted for Post-Acute Skilled Nursing Facility Stay by Accountable Care Organization Attribution, 2013–2017

	Total		ACO		Non-ACO		p-value
Sex, n (%)
Female	173,416	65.8	44,013	65.9	129,403	65.7
Male	90,260	34.2	22,829	34.2	67,431	34.3
Race, n (%)
White	219,856	83.4	56,614	84.7	163,242	82.9	***
Black	24,969	9.5	5,740	8.6	19,229	9.8
Hispanic	11,915	4.5	2,740	4.1	9,175	4.7
Other	6,936	2.6	1,748	2.6	5,188	2.6
Age at End of Reference Year, Mean [SD]	263,676	84.52 [7.42]	66,842	84.67 [7.31]	196,834	84.47 [7.45]	***
Dual Eligibility in Hospital Discharge Month, n (%)
Not dually eligible	193,134	73.3	51,123	76.5	142,011	72.2	***
Dually eligible	70,542	26.8	15,719	23.5	54,823	27.9
HCC Community Score, Mean Score [SD]	263,676	3.11 [2.28]	66,842	3.13 [2.25]	196,834	3.10 [2.29]	***
Any ICU Days During Hospital Stay, n (%)
No ICU Days	193,413	73.4	48,867	73.1	144,546	73.4
Any ICU Days	70,263	26.7	17,975	26.9	52,288	26.6
Index Hospital Length of Stay, Median Days [SD]	263,676	5.00 [4.77]	66,842	5.00 [4.49]	196,834	5.00 [4.86]
Skilled Nursing Facility Length of Stay, Median Days	263,676	26.0	66,842	26.0	196,834	26.0	***
Skilled Nursing Facility Length of Stay, Mean Days [SD]	263,676	32.52 [24.51]	66,842	31.72 [23.87]	196,834	32.79 [24.71]
Hospital Readmission Post-Hospital Discharge, n (%)
None	154,843	58.7	39,475	59.1	115,368	58.6	*
Occurred Within 30 Days	38,077	14.4	9,306	13.9	28,771	14.6	***
Occurred Within 31–90 Days	40,051	15.2	10,233	15.3	29,818	15.2
Occurred Within 91–180 Days	30,705	11.7	7,828	11.7	22,877	11.6
Death Post-Hospital Discharge, n (%)
None	212,882	80.7	54,364	81.3	158,518	80.5	***
Occurred Within 30 Days	7,353	2.8	1,812	2.7	5,541	2.8
Occurred Within 31–90 Days	22,816	8.7	5,590	8.4	17,226	8.8	***
Occurred Within 91–180 Days	20,625	7.8	5,076	7.6	15,549	7.9	**

^***p<0.001,

^**p≤0.01,

^*p<0.05

SOURCE: Authors’ analysis of data from the Medicare Research Identifiable Files, 2013–17.

Overall, the mean patient characteristics were similar across the two cohorts (see Table 1). The ACO cohort was slightly older (mean age 84.7 vs 84.5 years; p<0.001) and differed in racial composition (85% vs 83% White, 9% vs 10% Black, 4% vs 5% Hispanic; p<0.001) and dual eligibility (24% vs 28%; p<0.001) compared to the non-ACO cohort. The ACO cohort had a slightly higher average HCC score (3.13 vs 3.10; p<0.001) and had a similar percentage of ICU stays during the index hospitalization compared to the non-ACO cohort.

Patient Outcomes

Overall, the mean SNF LOS was 32.5 days. Hospital readmission rates at 0–30, 31–90, and 91–180 days from index hospitalization were 14.5%, 15.1%, and 11.6%, and mortality rates at 0–30, 31–90, and 91–180 days from index hospitalization were 2.9%, 8.7%, and 7.8%, respectively.

When comparing the crude cohort data, ACO-attributed patients had shorter SNF LOS than non-ACO-attributed patients (31.72 vs 32.79 days; p<0.001). Thirty-day hospital readmission rates were lower for ACO vs non-ACO (13.9% vs 14.6%; p<0.001) but were similar at 31–90 days and 91–180 days. No significant difference was observed in 30-day mortality post-hospital discharge for ACO vs non-ACO; however, slightly lower mortality was observed in the ACO group at 31–90 days (8.4% vs 8.8%; p=0.002) and 91–180 days (7.6% vs 7.9%; p=0.011) compared to the non-ACO group (Table 1).

The results after two-stage regression modeling demonstrate that ACO-attributed patients had a shorter SNF LOS by 0.71 days (p<0.001) compared to ADRD beneficiaries not attributed to ACOs (Table 2). The strength of the instrument in predicting LOS, an important feature of any instrumental variable analysis, was measured by an F-statistic of 29.8. This indicates that the instrument was strong. However, no association was found between SNF LOS and hospital readmission at intervals up to 6 months after hospital discharge. The probability of 30-day mortality was not associated with LOS, but the probability of mortality at 90 and at 180 days was slightly higher with longer LOS (0.6 percentage point increase in mortality within 90 days per each additional day, p=0.04; 1.3 percentage point increase in mortality within 180 days per each additional day, p=0.001). These results mean that the association of ACO attribution with clinical outcomes is small, with no association with readmissions and lower mortality at later time points, contrary to our hypothesis. In a sensitivity analysis that did not include facility fixed effects, which captured unmeasured differences between each SNF, there were small magnitude, but statistically significant, differences in readmission and mortality rates, both favoring ACO-attribution. The difference in results between the models with and without SNF fixed effects on readmission suggests that SNF-specific factors among the network of SNFs used by ACO-attributed beneficiaries, such as higher quality, may drive the association of ACO-attribution and lower readmission. However, mortality was slightly lower for ACO-attributed ADRD beneficiaries even when we control for differences across SNFs.

Table 2.

Association of accountable care organization attribution among Medicare beneficiaries with Alzheimer’s disease and related dementias with skilled nursing facility length of stay, readmissions, and mortality

Effects of ACO attribution on SNF rehabilitation length of stay
	SNF length of stay change (days)
In ACO	−0.705*** days
Effects of SNF rehabilitation length of stay on key patient outcomes
Time period following hospital discharge	Change in event probability for each day
Hospital readmission
0–30 Days	0.0043
0–90 Days	0.0023
0–180 Days	0.0019
Mortality
0–30 Days	0.000
0–90 Days	0.006*
0–180 Days	0.013***

^***p<0.001,

^**p≤0.01,

^*p<0.05.

SOURCE: Authors’ analysis of data from the Medicare Research Identifiable Files, 2012–17.

NOTES: These results were generated using Instrumental Variables (2SLS) regression. Models included age, sex, race, Medicaid enrollment, comorbidity, hospital length of stay and if any ICU days, SNF facility fixed effects, and quarter-year fixed effects.

DISCUSSION

In a large national sample of Medicare beneficiaries, we observed that ACO-attributed patients with a diagnosis of ADRD had shorter SNF LOS than non-ACO-attributed ADRD patients. The difference in LOS, however, was not associated with hospital readmission but was associated with slightly lower mortality within 180 days after the index hospitalization, rather than higher which was the hypothesized effect. Given the observational method and potential for residual confounding, we interpret the finding of lower mortality among ACO-attributed ADRD patients with caution. Therefore, we did not find evidence to support the concern that ACO or policy interventions that reduce post-acute rehabilitation length of stay for people with dementia may have unintended negative consequences.

Post-acute SNF use has been an important area for payment policy reform. Prior research has demonstrated that Medicare ACOs have influenced behavior change in post-acute SNF use. The impact of reforms may have heterogeneous effects across vulnerable populations, such as people with ADRD. People with ADRD are a critical group among users of the Medicare post-acute SNF benefit because they account for a third of SNF stays²⁷, have longer lengths of stay,¹⁶ and are three times more likely to end up as a long stay nursing home resident.²⁸ Yet, the hip fracture literature supports that people with dementia, despite the presence of cognitive impairment, can improve with rehabilitation and return to premorbid status.^29,30 As a result, it is important to demonstrate that the implementation of ACOs does not worsen major outcomes as an unintended consequence for this vulnerable population by reducing spending and length of stay in SNFs. In this study, we did not observe negative consequences on readmission and detected a small positive mortality effect. As noted previously, we interpret the small benefit with caution due to the potential for residual confounding due to the potential for unmeasured factors such as socioeconomic or marital status. Overall, these results lessen the concern for negative unintended consequences on readmissions or mortality.

There are limitations in this study that need to be kept in mind. While using an instrumental variable approach reduces the effects of selection bias, this study is observational, which limits causal interpretation, meaning that there may be mechanisms in addition to SNF LOS contributing to the ACO vs non-ACO findings. There remains the possibility of residual confounding due to unmeasured differences between the study groups despite including baseline comorbidity (i.e., HCC score) and hospital-illness severity. Another limitation is that we did not have access to functional status or cognitive status data for this study which would have allowed both additional covariate adjustment and the ability to report functional recovery which is the main goal of rehabilitation. Lastly, these results are not generalizable to Medicare Advantage.

Policy Implications

While these results are reassuring for the effects of the ACO payment model on people with ADRD, they also provide relevant information about other new payment models. The Patient Driven Payment Model (PDPM) is one reform that changed payment for Medicare SNF stays from being funded based on the number of rehabilitation minutes delivered to disease-based risk-adjustment payment.^21,31 This change has led to a decrease in the amount of rehabilitation delivered, raising concerns about unintended consequences for people living with dementia. Our study provides some reassurance that in another value-based payment model that reduced the amount of rehabilitation delivered to people living with ADRD, measured by SNF length of stay, there were not large untoward outcomes associated with the reduction in rehabilitation. Regardless of ACO affiliation, clinicians providing care at SNFs should carefully assess the unique needs of patients, particularly those with ADRD, and implement individualized care plans to optimize health outcomes.

CONCLUSION AND IMPLICATIONS

Being an ACO-attributed patient with ADRD is associated with shorter skilled nursing facility length of stay, which does not directly affect readmissions and has a small association with lower mortality.