Comparative Effectiveness of Inpatient Rehabilitation Versus Skilled Nursing Facilities for Stroke and Hip Fracture Patients

Abstract

Background:

Prior research suggests discharge to inpatient rehabilitation facilities (IRF) leads to improved outcomes for stroke and hip fracture patients relative to skilled nursing facilities (SNF), while incurring greater costs. However, these estimates are likely biased by non-random patient selection.

Methods:

We used a quasi-experimental design to compare post-acute care outcomes among Medicare beneficiaries hospitalized for stroke or hip fracture in fifty-five US hospitals that closed their IRF units between 2009-2017. Primary and secondary outcomes were 30-, 90-, and 180-day readmission and mortality, and successful community discharge.

Results:

Among 10,761 stroke and 13,963 hip fracture hospitalizations, IRF discharge declined sharply, offset by increases to SNF and home health. Relative to IRF, SNF discharge was associated with no significant differences in readmissions, but an increase in 90-day mortality for stroke (+6.5%, 95% CI 1.5% to 11.4%) and hip fracture (+5.8%, 95% CI 2.5% to 9.0%). Successful community discharge did not differ for patients redirected to SNF, but stroke patients redirected to home health had significantly higher rates of successful discharge (DID estimate: +6.8%; 95% CI 0.1% to 13.5%). The protective effect of IRF was concentrated within twenty days post-discharge.

Conclusions:

Following hospitalization for stroke and hip fracture, discharge to an IRF was associated with lower mortality relative to SNF. However, given the potential for unmeasured confounding, this association should be interpreted with caution. Careful post-acute care referral protocols are critical to ensure good patient outcomes.

INTRODUCTION

Post-acute care (PAC), including inpatient rehabilitation facilities (IRFs), skilled nursing facilities (SNFs), and home health agencies, provides essential therapies to patients after hospital discharge. In 2022, Medicare spent over $53 billion on these services, yet the optimal setting for patient recovery remains uncertain.¹

Patient selection into these settings is not random. IRFs are subject to Medicare’s “60% Rule,” which concentrates patients with specific conditions like stroke and hip fracture, while also admitting based on medical acuity and the ability to tolerate intensive therapy. This process creates a significant risk of confounding by indication in comparative effectiveness research, and uncertainty remains about the optimal PAC setting for delivering this therapy. For hip fracture patients, one study found that those discharged to IRF experienced lower one-year mortality relative to patients discharged home (IRF: 13.6%; Home: 19.3%), but IRF does not appear to reduce mortality or improve physical function compared to SNF, though an exception has been noted for patients with dementia.^2–5 For stroke patients, studies have found consistently that relative to SNF, discharge to IRF increases the probability of successful discharge to the community, improves gains in physical function, and reduces mortality.^6–12 Given the significant cost differences and mixed evidence on IRF’s added value across conditions, there is a critical need for unbiased, comparative evidence to guide policymakers and providers in determining the most appropriate PAC setting.

To correct for the selection bias inherent in PAC discharge decisions, some studies have used distance to facilities as an instrumental variable (IV).^3,4,13–15 However, such instruments are imperfect. Distance-based IVs are not perfectly random, as patient and provider locations are often correlated with socioeconomic factors that independently influence health outcomes.

In this study, we exploit a natural experiment resulting from the closures of hospital-based IRF units, which generated sharp changes of PAC discharge patterns at these hospitals. Hospital-based IRF units, which account for 73% of all IRFs (849 out of 1,159 IRFs in 2020), operate with much lower profit margins (less than 2% on average compared to 25% for free-standing IRFs) due to higher fixed costs and lower patient volume, which puts some of them at greater risk of closure.¹⁶ These closures provide a stronger instrumental variable than distance-based approaches because they are primarily driven by financial considerations at the hospital level rather than local patient characteristics. ¹⁷ Additionally, IRFs are far less common (fewer than 1,100 facilities nationwide) than SNFs (over 15,000), making alternative IRF options after a closure much more limited.¹⁸ For this reason, IRF closures result in increased discharges to SNF or HHA and act as a strong instrumental variable that provides a robust and unbiased evaluation of the comparative effectiveness of IRF relative to SNF care.

METHODS

Data Sources and Study Population

This quasi-experimental study exploits hospital-based IRF unit closures as a natural experiment to estimate changes in post-acute care outcomes when patients switch from IRF to SNF or home health (HH) following hospitalization. We identified fifty-five hospitals which closed their IRF units between 2009 and 2017, while maintaining similar volume (+/−25%) of stroke and hip fracture hospitalizations and decreased their rate of discharge to IRF by at least 25% (relative to the pre-closure period). We excluded hospitals near free-standing IRFs or with access to another hospital’s IRF unit. The study population included all continuously enrolled Traditional Medicare (TM) beneficiaries who were hospitalized within two years before or after an IRF-unit closure for hip fracture or stroke and subsequently discharged to IRF, SNF, home with HH, or home without HH, excluding patients who experienced in-hospital mortality during the two years before or after the IRF unit closure. The Medicare Beneficiary Summary File (MBSF) was used for enrollment information and death dates; hospitalizations were identified using the MedPAR file, and PAC utilization was measured using Medicare SNF, IRF, HHA and LTCH claims and patient assessment data.

Exposure

Exposure consisted of the specific post-acute discharge setting, including IRF, SNF, home, and home with HH before and after IRF-unit closure.

Outcomes

Our primary outcome is all-cause hospital readmissions within 30-, 90-, and 180-days post-hospital discharge. Our secondary outcomes are mortality within 30-, 90-, and 180-days post-hospital discharge and successful community discharge. We defined successful community discharge using the principles of the parallel CMS Quality Reporting Program measures for post-acute care. A successful discharge was defined as remaining alive and in the community (i.e., no rehospitalization or admission to a SNF, IRF, LTCH, or long-stay nursing home) for the 30-day period following the end of the post-acute care episode, for both institutional and home-based services.^19,20

Covariates

Our analysis controlled for 224 covariates across the following domains: beneficiary demographics and Medicare enrollment information from the MBSF, chronic conditions from the Chronic Conditions Warehouse (CCW), diagnoses, procedures and other hospitalization characteristics from MedPAR, and residential settings prior to hospital admission. Diagnosis and procedure codes from the hospitalization were classified into clinically meaningful groups using Clinical Classifications Software (CCS) from the Healthcare Cost and Utilization Project (HCUP).²¹ We calculated the Elixhauser mortality score using the index hospitalization diagnoses.^22,23

Identification Strategy and Statistical Analysis

Overview

Sharp changes in post-acute discharge setting patterns following IRF-unit closures serve as a quasi-natural experiment, as cases formerly discharged to IRF were diverted to alternate settings such as SNF or home with HH. Our approach restricts analysis to patients discharged from the same hospitals and residing in the same counties pre- and post-IRF closure, thus controlling for area-level socioeconomic factors that might otherwise confound the relationship between discharge setting and outcomes. We seek to identify the patients whose discharge setting choice was influenced by IRF closures (those who used SNF or HH but would have used IRF if it was available) to minimize the threat of selection bias when comparing patient outcomes across PAC settings and across hospital/regions.

Cross-Temporal Matching

Our analytic strategy uses a propensity-score based method to identify the relevant comparison groups of patients pre and post IRF closure. We applied a cross-temporal matching design approach that identifies three groups of patients present in an instrumental variables (IV) framework. Compliers (patients who, due to the closure, were redirected from IRF to SNF or HH), Always Takers (those who would use IRF regardless), and Never Takers (those who would never use IRF).²⁴ Typical IV estimators provide a (local) average treatment effect estimate for the so-called Compliers or marginal patients for whom treatment depends on the instrument. However, the usefulness of this effect estimate for informing clinical practice and policy decisions has been questioned, since the group of Compliers is not identified.²⁵ Our cross-temporal matching approach, on the other hand, enables identification of these groups through a three-step process, detailed in the Supplemental Appendix, allowing us to examine differences across patient groups that can inform possible interventions.

The multinomial propensity score models of discharge to IRF, SNF, and home with HH (with home without HH as reference group) were estimated using the extreme gradient boosting (XGBoost) machine learning algorithm.²⁶ Matching was conducted within hospital clusters and with replacement using a vector-matching approach which leverages all three dimensions of the generalized propensity scores vector.²⁷ Balance within matching groups was assessed across 224 clinically relevant covariates using standardized mean differences with an a-priori defined threshold of 10%.^28–30 The overlap in propensity score distributions across comparison groups further confirmed the robustness of our matching procedure (Supplementary Figure 1). Additional matching approach details are available in the Supplemental Methods.

Difference-in-Differences Model

To estimate the comparative effectiveness of discharge to IRF versus SNF or HH, we use a difference-in-differences (DID) model, which compares changes in outcomes over time between a treatment group affected by an intervention and a control group unaffected by it, allowing us to isolate the effect while accounting for time trends. Compliers, who were observed under different PAC treatments in the pre- and post-closure periods, served as the treatment group, while matched Never Takers served as the control group to account for secular trends in outcomes over time. For the community discharge days outcome (a count variable), we employed negative binomial regression due to overdispersion in the data (variance exceeding the mean). For other outcomes, we used linear probability models, adjusting for all covariates from the propensity score matching to address residual imbalance. ^31,32 The DID coefficients represent the average treatment effects of IRF relative to SNF among the Compliers. We repeated this estimation approach to produce DID estimates of IRF relative to HH.

To visualize the comparative effects of IRF versus SNF and HH, we created figures displaying the pre- and post-closure values for both treatment and control groups, alongside the adjusted DID estimates with confidence intervals.

Additional Analyses

To address potential declines in hospital quality contemporaneous with IRF closures, we conducted falsification tests using beneficiaries hospitalized at the same hospitals during the study period with diagnoses that rarely require IRF discharge: congestive heart failure (CHF) and chronic obstructive pulmonary disease (COPD). We specifically selected these conditions because, unlike stroke and hip fracture, they are not among Medicare’s qualifying conditions for IRF care and thus patients with these diagnoses should not be significantly affected by IRF unit closures. These analyses aimed to determine whether health outcomes for these patients changed after IRF unit closures in a similar manner to that observed for stroke and hip fracture patients.

RESULTS

Changes in PAC Setting Discharge Rates

Our final analytic sample included 10,761 stroke cases and 13,963 hip fracture cases (Table 1). During the two years prior to IRF-unit closure, 28.9% of the 5,264 stroke cases were discharged to IRF, 26.4% to SNF, 13.5% to HH, and 31.2% to home. Following the closures, IRF discharges declined to 9.8% (−19.1pp), offset by increases in discharges to SNF (+11.9pp) and HH (+5.0pp). Hip fracture cases (N=7,154 during pre-period) experienced a similar trend, with IRF discharges decreasing from 33.3% to 11.2% (−22.1pp), almost entirely offset by increased discharges to SNF (+21.4pp).

Table 1:

Characteristics of Stroke Patients Discharged to PAC Before and After IRF-Unit Closure

	Discharge Setting
	IRF		SNF		HHA		Home
	Pre	Post	Pre	Post	Pre	Post	Pre	Post
Hospitalizations, N (%)	1,523	541	1,390	2,108	710	1,014	1,641	1,834
	28.9%	9.8%	26.4%	38.3%	13.5%	18.4%	31.2%	33.4%
Change		−19.09%		11.94%		4.96%		2.19%
Age	77.9	75.3	82.1	81.4	78.5	78.4	74.4	74.3
Sex: Female (%)	53.5	49.0	65.8	62.9	56.8	59.4	47.0	46.5
Race: White (%)	83.5	76.5	85.7	85.8	82.0	79.8	86.1	84.6
Race: Black (%)	8.4	11.8	7.8	6.3	9.7	10.8	6.1	6.9
Enrolled: OASI (%)	81.7	80.2	85	84.4	79.4	78.4	78.5	78.8
Enrolled: Disability (%)	17.7	18.1	14.1	15.3	19.7	19.8	20.2	20.2
Medicaid Dual (%)	17.7	20.7	27.2	23.7	24.4	23.5	16.3	16.9
Hospital LOS	4.7	5.5	6.5	5.6	3.8	3.5	3.0	2.8
ICU (%)	44.5	52.3	38.5	39.6	35.5	36.2	35.2	39.1
General (%)	19.0	23.7	16.7	18.3	9.7	10.5	12.2	12.8
Intermediate Care (%)	23.0	27.4	20.6	20.9	24.8	24.9	21.5	25.7
CCU (%)	22.4	25.3	19.8	19.3	21.7	19.1	17.8	16.3
Elixhauser mortality score	3.8	3.4	5.1	4.9	3.7	3.6	2.1	1.9
CCW, N of Chronic Conds.	9.2	8.8	10.7	10.6	10.2	10.2	8.2	8.4
Pre Hospit.: Nursing Home (%)	1.4	0.6	18.3	13.9	0.7	0.8	0.2	0.1
Stroke Type: Hemorrhagic (%)	10.3	10.0	10.9	10.4	7.0	7.2	5.2	5.9
CCW: ADRD (%)	19.9	18.7	48.1	41.0	28.2	29.4	13.9	13.2
CCW: CHF (%)	45.2	38.8	58.1	53.5	51.8	51.7	36.0	34.5
CCW: COPD (%)	35.4	29.8	39.9	40.6	42.4	40.8	31.4	32.8
CCW: Depression (%)	27.6	29.0	37.0	38.6	36.1	34.9	24.6	25.1
CCW: Cataracts (%)	67.0	59.5	77.6	74.9	70.6	69.2	60.4	59.4
CCW: Osteoporosis (%)	19.3	15.5	30.9	30.7	24.2	22.8	17.5	16.4
CCW: Anemia (%)	59.2	58.0	71.2	71.2	68.2	69.6	49.2	50.2
CCW: Hypertension (%)	95.4	95.0	96.6	96.4	95.5	97.6	92.3	92.5
Paralysis (%)	44.1	53.6	33.7	39.7	16.9	20.7	12.4	15.8
Other Gastro. Disorders (%)	19.2	24.4	24.5	25.5	13.4	11.6	8.3	9.1
Other nervous system (%)	54.9	68.2	49.9	57.1	41.1	52.1	44.0	48.9
UTI (%)	12.9	11.8	21.2	19.5	13.4	10.9	7.4	6.1
TCI (%)	0.8	0.2	1.4	1.1	2.3	2	3.2	2.3
Connective Tissue Disorders (%)	30.1	35.3	23.7	32.4	24.6	28.6	23.7	27.4
DM w/o Comp. (%)	25.3	22.0	25.2	24.2	26.5	27.2	24.3	26.2
Echocardiogram (%)	7.5	8.1	5.5	4.6	6.2	5.3	9.0	7.0
Gastrostomy; temp/perm (%)	2.4	4.3	8.6	5.6	0.8	0.7	0.7	0.6
Resp. Intubation and MV (%)	2.2	3.9	3.2	2.9	1.1	1.0	1.4	1.6
Other Therapeutic Proc. (%)	5.3	7.6	4.6	5.8	2.8	5.8	4.1	5.9

Abbreviations: OASI = old-age and survivors’ insurance; LOS = length of stay (days); ICU = intensive care unit; CCU = cardiac care unit; CCW = chronic conditions warehouse; ADRD = Alzheimer’s disease & related dementias; COPD = chronic obstructive pulmonary disease; CHF = congestive heart failure; UTI = urinary tract infection; TCI = transient cerebral ischemia; DM = diabetes mellitus; MV = Mechanical Ventilation.

Patient Characteristics

Prior to matching, patient characteristics differed substantially across post-acute care settings for both the stroke and hip fracture cohorts (Table 1, Supplementary Table 1). Consistent with prior research and clinical criteria, patients discharged to IRF were generally younger, had a lower burden of comorbidity (e.g., lower Elixhauser scores and lower prevalence of ADRD), and were less likely to be dually-eligible for Medicaid compared to patients discharged to SNF.

Post-Matching Balance

Following propensity score matching, all comparison groups were well-balanced, as detailed in Supplementary Table 2 and Supplementary Table 3.

Changes in Patient Outcomes

Figure 1 presents our comparison between IRF and SNF for both stroke and hip fracture patients. For stroke patients, no significant differences in readmission rates were observed between IRF and SNF at any time point in the adjusted DID estimates. However, mortality rates increased substantially when patients were redirected from IRF to SNF, with adjusted DID estimates showing increases of 6.8 percentage points at 30 days (p<.01) and 6.5 percentage points at 90 days (p<.05) from baseline rates of 3.6% and 9.2%, respectively (Table 2).

Figure 1:

IRF vs SNF Care Outcomes: Pre/Post Values and DID Estimates

*DID =Difference in differences, unadjusted mean values and DID point estimates (bars) and 95% confidence interval (black lines)

Table 2:

Unadjusted Pre and Post IRF-closure Patient Outcomes and DID treatment effect estimates.

	IRF vs SNF							IRF vs HH
	CO		NT		Difference in Differences			CO		NT		Difference in Differences
	Pre	Post	Pre	Post	Unadj.	Adjusted (Percentage Points, 95% CI)		Pre	Post	Pre	Post	Unadj.	Adjusted (Percentage Points, 95% CI)
Stroke
Mortality (%)
30-day	3.55	8.17	14.29	12.67	6.23	6.81**	(2.06, 11.56)	3.57	0.54	4.33	2.53	−1.23	−0.45	(−3.91, 3.00)
90-day	9.23	13.78	23.50	22.47	5.58	6.45*	(1.38, 11.52)	8.57	5.54	9.57	6.32	0.21	1.09	(−4.77, 6.95)
180-day	15.91	19.67	29.72	30.76	2.73	3.94	(−1.70, 9.58)	14.82	10.00	14.98	12.27	−2.11	−0.41	(−6.48, 5.65)
Readmission (%)
30-day	13.49	13.28	14.52	13.94	0.36	−0.22	(−5.56, 5.12)	13.93	15.00	12.82	13.72	0.17	1.04	(−6.83, 8.92)
90-day	29.40	25.00	28.23	27.42	−3.60	−4.11	(−11.46, 3.23)	30.54	24.46	25.27	28.34	−9.14	−7.56	(−17.39, −2.27)
180-day	41.62	38.07	37.21	37.67	−4.01	−4.71	(−11.32, 1.90)	42.14	35.89	36.46	38.63	−8.42	−7.7	(−19.26, 3.85)
Successful Discharge	76.56	75.78	72.24	73.16	−1.70	−1.53	(−7.49, 4.44)	77.68	83.93	84.48	83.21	7.51	6.79**	(0.09, 13.49)
N (Weighted)	704	704	868	868				554	554	560	560
Hip Fracture
Mortality (%)
30-day	1.62	5.00	6.57	5.46	4.50	4.12**	(1.47, 6.77)
90-day	6.00	10.85	15.20	14.01	6.03	5.77***	(2.47, 9.06)
180-day	10.00	15.23	21.77	20.56	6.44	6.25**	(2.21, 10.29)
Readmission (%)
30-day	9.69	14.15	12.40	10.90	5.97	5.78**	(1.86, 9.70)
90-day	23.62	26.77	25.30	24.59	3.87	3.09	(−1.98, 8.17)
180-day	33.38	35.23	34.67	34.54	1.97	1.22	(−4.74, 7.18)
Successful Discharge	81.28	77.1	73.97	74.88	−5.09	−4.39	(−9.86, 1.08)
N (Weighted)	1,175	1,175	3,757	3,757

Note: “Compliers” (CO) are patients redirected from an inpatient rehabilitation facility (IRF) to a skilled nursing facility (SNF) or home health (HH) post-closure. “Never Takers” (NT) are control patients who used SNF or HH in both periods. The adjusted difference-in-differences (DID) estimate represents the average treatment effect of this redirection, controlling for all covariates. CI indicates confidence interval.

^*p<0.05,

^**p<0.01,

^***p<0.001

Hip fracture patients redirected to SNF experienced significantly higher 30-day readmission rates (5.8 percentage point increase, p<.01), with no significant differences at later time points. Mortality rates showed significant increases at all time points for hip fracture patients redirected to SNF: 4.1 percentage points at 30 days (p<.01), 5.8 percentage points at 90 days (p<.001), and 6.3 percentage points at 180 days (p<.01) from baseline rates of 1.6%, 6.0%, and 10.0%, respectively. For the successful community discharge outcome, we found no statistically significant differences for either stroke patients (DID estimate: −1.53%; 95% CI −7.49% to 4.44%) or hip fracture patients (DID estimate: −4.39%; 95% CI −9.86% to 1.08%) redirected to SNF.

Supplementary Figure 2 presents the comparison between IRF and HH for stroke patients. Pre-closure mortality rates were generally comparable between IRF and HH settings. The adjusted DID estimates showed no significant changes in mortality or readmission rates when patients were redirected from IRF to HH. However, in contrast to the SNF comparison, stroke patients who received home health instead of IRF had a significantly higher likelihood of a successful community discharge (DID estimate: +6.79%; 95% CI 0.09% to 13.49%).

As a sensitivity analysis to examine the lag-time and duration of the IRF treatment effects, we estimated DID coefficients for mortality at ten-day increments following hospital-discharge. This analysis (Figure 2 and Figure 3) reveals that the increased mortality effect is larger during the first 20 days for both conditions (average IRF length-of-stay was 13 days) and remained stable until day 90, declining marginally after that for stroke patients. For hip fracture patients, the increased mortality continued to grow until day 120 where it stabilized.

Figure 2:

Adjusted DID (SNF relative to IRF) Mortality Estimates, by Days Post-Discharge (Stroke)

* DID = Difference in differences, point estimates (circle dots) and 95% confidence intervals (dotted lines) every ten-day increment up to 180 days following hospital discharge

Figure 3:

Adjusted DID (SNF relative to IRF) Mortality Estimates, by Days Post-Discharge (Hip Fracture)

* DID = Difference in differences, point estimates (circle dots) and 95% confidence intervals (dotted lines) every ten-day increment up to 180 days following hospital discharge

Patient Characteristics Associated with Increased Mortality Among Patients Substituting SNF for IRF

Given the observed increase in mortality rate for both stroke and hip fracture patients that received SNF care in place of IRF, we conducted a post-hoc regression model to identify characteristics strongly associated with higher 30-day mortality in the SNF Compliers group using all main covariates and their interactions with the post-period indicator. For the stroke cohort, patients with hemorrhagic stroke (compared to ischemic stroke), other ill-defined heart disease, disease of white blood cells and other liver disease had statistically significant increases in 30-day mortality in the (SNF) post period, with dysphagia significant at the 10% level (Supplementary Table 4A). For the hip fracture cohort, predictors of increased 30-day mortality were secondary malignancies, coagulation and hemorrhagic disorders, pneumonia and aspiration pneumonitis, Parkinson’s disease, myocarditis and cardiomyopathy, coronary atherosclerosis and other heart diseases, acute renal failure, other (non-hip) fractures, and having received ICU care in the hospital (Supplementary Table 4B).

Our falsification test analyses comparing persons hospitalized for CHF and COPD, two low IRF discharge rate conditions, in the same hospitals during the same two years pre and post IRF closure showed no significant changes in 30-day mortality rates (CHF: 12.1% pre vs. 12.5% post, or −0.18% (p-value=0.61) difference after adjusting for annual trend; COPD: 4.4%% pre vs. 4.2% post, −0.26% (p-value=0.35) adjusted difference) indicating no significant general changes in hospital mortality after the IRF unit closure.

DISCUSSION

This study leverages the quasi-natural experiment created by the closure of hospital-based IRF units to examine our primary outcome, all-cause readmissions. We found that for patients redirected from IRF to SNF, there was no corresponding decrease in rehospitalization for stroke patients and a significant increase in 30-day readmissions for hip fracture patients. This finding is particularly striking despite the competing risk of our key secondary outcome: these same patients experienced substantially higher mortality rates. The elevated mortality would be expected to lower the readmission rate, suggesting the underlying risk of adverse events for patients redirected to SNF care was substantially elevated. This conclusion is further supported by the lack of difference in successful community discharge between IRF and SNF.

For stroke patients discharged to HH instead of IRF, no significant differences in mortality or readmission outcomes were observed. Notably, these patients had a significantly higher likelihood of a successful community discharge compared to those discharged to an IRF, suggesting a potential benefit to avoiding an institutional stay for this outcome. Because our outcome window begins after the cessation of PAC services for all settings, this finding represents a more stable return to the community for patients discharged home with home health.

Our analyses have two distinct advantages over previous IRF comparative effectiveness studies. First, the abrupt loss of access to IRF care due to IRF-unit closures provided a stronger instrumental variable than distance-based IVs between patient’s residence and PAC providers in prior studies, which can be confounded by with socioeconomic and geographic factors. Second, our design enabled clear identification and characterization of Compliers (patients that switched discharge setting due to the IRF closure), Always Takers, and Never Takers. This approach allowed us to identify and profile Compliers – SNF users who were generally younger and less clinically complex compared to Never Takers that are discharged to SNF when IRF care is available (Supplementary Table 2 and Supplementary Table 3). Additionally, identifying the Compliers allowed us to examine which patient characteristics were most strongly associated with increased mortality when redirected to SNF care.

Our study included falsification tests with CHF and COPD patients to further validate our findings. Given that CHF and COPD patients are seldom discharged to IRFs, these populations served as controls to examine if health outcomes experienced comparable changes following IRF closures as seen in stroke and hip fracture patients. We observed no significant differences in mortality outcomes for CHF and COPD patients pre- and post-closure, supporting the hypothesis that the increased mortality for stroke and hip fracture patients was likely due to the loss of IRF access rather than broader hospital or community-level factors.

The diverging mortality patterns between stroke and hip fracture patients reflect disease-specific rehabilitation mechanisms. For stroke, IRF’s early protection (peaking at 20 days) aligns with preventing acute complications (e.g., aspiration pneumonia), while hip fracture patients benefit from sustained mobility rehab reducing frailty-related risks (e.g., falls).^5,33,34 These differences explain why IRF’s effects attenuate by 180 days for stroke but persist for hip fracture. Additionally, care transitions between hospitals and SNFs are often hindered by poor interoperability and a lack of shared accountability, resulting in medication errors, poor communication, and delays in therapy initiation. ^35–37 Hospital-based IRF patients receive intensive rehabilitation in a highly staffed environment with fast access to specialty care in the event of an acute exacerbation. Prior research has suggested that for patients with mild to moderate stroke severity, intensive rehabilitation reduces mortality risk significantly, and more than half of stroke deaths are the result of severe medical and neurological complications. ^33,34,38,39 SNFs are less likely to have timely access to specialists (e.g., neurologists, intensivists) and care support (high nursing staffing ratios) needed to address patients experiencing these complications. In our stroke group of matched Compliers, 20-day specialty care utilization after hospital discharge was substantially greater when discharged to IRFs (neurology, 21% vs 10%; radiology, 39% vs 13%; physical medicine & rehabilitation, 59% vs 14%).³⁸

The cost of IRF care, however, remains a significant consideration, as it can be more than twice as high as SNF care (e.g., average cost for stroke: $22,765 in IRF versus $11,261 in SNF).³ As more Medicare beneficiaries are cared for by Medicare Advantage and Medicare value-based programs, the high cost differential creates strong incentives for reducing IRF utilization for post-acute care. Our findings underscore that these financial incentives must be balanced with clinical needs to ensure high-risk patients can access the level of care required to reduce mortality.

This study has three main limitations. First, despite our robust quasi-experimental design and matching using a comprehensive set of patient characteristics, residual confounding cannot be entirely ruled out. For example, IRFs typically select patients capable of tolerating at least three hours of daily therapy, a selection criterion not fully captured in administrative data that may influence our findings. Furthermore, there is a strong likelihood that patients discharged to SNFs are more clinically complex or have less social support in ways not fully captured by our data, which could influence the observed mortality differences. Second, while we focused on essential outcomes for stroke and hip fracture patients, functional and cognitive improvement were not evaluated due to differences in assessment tools across PAC settings (IRF-PAI, MDS, and OASIS), making direct comparison of functional outcomes challenging. A significant limitation of our study is the absence of data on several other important clinical predictors of hospital readmission and mortality, including pre-admission measures of frailty, living situation (alone, with relative/friend, or facility), cognitive assessments, and advance care plans. Additionally, our study period overlapped with Medicare’s Hospital Readmissions Reduction Program, which may have influenced discharge patterns independent of IRF closures. Future research using newer standardized assessment data could partially address this limitation. Third, our results may not be fully generalizable to free-standing IRFs, as hospital-based IRFs may serve a different case mix and provide more care than their free-standing counterparts.¹⁶

In summary, our study found that for patients redirected from IRF to SNF care, there was no consistent benefit regarding hospital readmissions, our primary outcome. This finding, however, was coupled with a clear survival disadvantage, as discharge to an IRF was associated with a clear survival advantage, especially in the critical period immediately following hospital discharge. This observed association may be explained by the greater intensity of rehabilitation, continuity of care, and faster access to specialists in the hospital-based IRF setting. Healthcare providers should consider the potential consequences of payment incentives that discourage IRF use, particularly given our findings. Optimizing PAC triage to identify patients most vulnerable to adverse outcomes in SNF settings could improve mortality outcomes. Future research using more detailed functional data is needed to refine patient selection criteria, as our findings may still be influenced by residual confounding.

Key Points:

• Following closures of hospital inpatient rehabilitation facility (IRF) units, patients who would have received IRF care but were instead discharged to skilled nursing facility (SNF) had no significant change in ninety-day all-cause hospital readmissions but experienced a notable increase in ninety-day mortality: +6.5 percentage points for stroke (baseline: 9.2%) and +5.8 percentage points for hip fracture (baseline: 6.0%).

• There were no significant differences in the likelihood of a successful community discharge for patients redirected from IRF to SNF. However, stroke patients redirected to home health were significantly more likely to have a successful community discharge (+6.8 percentage points).

• Patients who would have received IRF care but instead were discharged to home with home health (HH) had no significant changes in mortality or readmission rates.

• Despite the quasi-experimental design, these findings may be influenced by unmeasured confounding, as administrative data cannot fully capture patient-level factors like rehabilitation tolerance.

Why does this paper matter?

While redirecting patients from IRF to SNF care did not consistently reduce hospital readmissions, it was associated with a significant increase in mortality. This underscores that IRF access has a critical impact on patient survival, a factor that must be carefully weighed in post-acute care referral and payment policies.