Relative contributions of hospital versus skilled nursing facility quality on patient outcomes

Abstract

Background

Hospitals and health systems worldwide have adopted value-based payment to improve quality and reduce costs. In the USA, skilled nursing facilities (SNFs) are now financially penalised for higher-than-expected readmission rates. However, the extent to which SNFs contribute to, and should thus be held accountable for, readmission rates is unknown. To compare the relative contributions of hospital and SNF quality on readmission rates while controlling for unobserved patient characteristics.

Methods

Retrospective cohort study of Medicare beneficiaries, 2010–2016. Acute care hospitals and SNFs in the USA. Medicare beneficiaries with two hospitalisations followed by SNF admissions, divided into two groups: (1) patients who went to different hospitals but were discharged to the same SNF after both hospitalisations and (2) patients who went to the same hospital but were discharged to different SNFs. Hospital-level and SNF-level quality, using a lagged measure of 30-day risk-standardised readmission rates (RSRRs). Readmission within 30 days of hospital discharge.

Results

There were 140 583 patients who changed hospitals but not SNFs, and 183 232 who changed SNFs but not hospitals. Patients who went to the lowest-performing hospitals (highest RSRR) had a 0.9% higher likelihood of readmission (p=0.005) compared with patients who went to the highest-performing hospitals (lowest RSRR). In contrast, patients who went to the lowest-performing SNFs had a 2% higher likelihood of readmission (p<0.001) compared with patients to went to the highest-performing SNFs.

Conclusions

The association between SNF quality and patient outcomes was larger than the association between hospital quality and patient outcomes among postacute care patients. Holding postacute care providers accountable for their quality may be an effective strategy to improve SNF quality.

INTRODUCTION

Hospitals and health systems worldwide are transitioning toward value-based payment structures. For example, England established the Payment for Results programme for hospitals to reduce costs and motivate high-quality efficient care.¹ France adopted a pilot programme known as Financial Incentive for Quality Improvement, which was expanded in 2016 to include all acute care hospitals.² Germany and Belgium have also taken up pay-for-performance policies for hospitals within the past 5 years.³ In the USA, with the passage of the Affordable Care Act in 2010, Medicare announced its focus on value-based payment and began financially penalising hospitals in 2012 with higher-than-expected readmission rates, which represent one dimension of quality.⁴ Since the announcement of this policy, readmission rates in the USA have significantly declined.⁵ Furthermore, recent evidence suggests the care provided by the hospital can increase or mitigate readmission risk; that is, the same patients are more likely to be readmitted if they are admitted to a hospital with a higher-than-expected readmission rate as compared with a hospital with a lower-than-expected readmission rate.⁶

With the success of these efforts, Medicare is now focusing on reducing readmission rates from skilled nursing facilities (SNFs). More patients than ever before are being discharged to SNFs,⁷ and readmission rates from SNFs remain stubbornly high.⁸ In October 2018, Medicare began financially penalising SNFs with higher-than-expected readmission rates.⁹ Despite enthusiasm for the goal of reducing unnecessary readmissions from SNFs, some have raised questions about whether SNF readmissions are truly under the control of SNFs or whether they are more likely attributable to the quality of care received at the hospital from which the patient was discharged.¹⁰ If hospital quality is the principal driver of readmissions, it may not be prudent to hold SNFs accountable for readmissions.

Prior work has attempted to characterise the relative contributions of hospital and SNF quality on patient outcomes among patients with hip fracture and found that SNF quality explained more variation in patient outcomes than did hospital quality.¹¹ However, it is unclear whether these findings generalise to the broader non-hip-fracture SNF population. Furthermore, prior research has compared outcomes of patients at high-quality and low-quality SNFs without adequately accounting for differences in patient characteristics that exist across these different types of facilities.¹² Thus, these associations could be driven by patient characteristics rather than the effect of the SNF’s quality. Specifically, if high-quality SNFs care for patients who are unobservably healthier than those cared for by low-quality SNFs, this association could be spurious.

Our objective was to compare the relative contributions of hospital and SNF quality on patient readmission rates while controlling for unobserved patient characteristics. To do so, we sampled patients who were hospitalised twice and discharged to SNFs following both hospitalisations, but who either changed hospitals or changed SNFs. We then used each patient as his or her own control, and thus attributed differences in readmission rates between the first and second hospitalisations to either a change in hospital quality or SNF quality, thus providing new evidence for the extent to which hospital quality and SNF quality impact readmission rates.

METHODS

Data

We used 100% Medicare Provider and Analysis Review files from 2010 to 2016 to identify hospital discharges from acute care and critical access hospitals for fee-for-service Medicare beneficiaries with subsequent SNF admissions. We supplemented these data with hospital characteristics from the 2016 American Hospital Association’s Annual Survey and SNF characteristics from the 2016 Nursing Home Compare data, Medicare’s online public reporting tool for nursing homes.¹³ We obtained patient characteristics of age, sex, race and dual-eligible status from the Medicare Beneficiary Summary File and calculated indicators of 31 comorbidities based on the Centres for Medicare and Medicaid Services (CMS) Hospital Readmission Reduction Programme specifications.^14,15

Study Sample

Our study sample was at the hospital–SNF stay level, and each individual has two observations (corresponding with two hospital–SNF stays). To create this sample, we started with all hospital discharges for patients who were 66 years old or greater, admitted after 1 January 2010 and discharged before 31 October 2016. We exclude patients younger than 66 years in order to have a full year of data to measure patient comorbidities prior to hospitalisation. We also exclude patients enrolled in Medicare Advantage and those discharged to hospice.

We then limited our sample to beneficiaries with at least two inpatient claims with discharges to SNF following the hospitalisations. For patients with more than two discharges, we selected the first two, creating a sample with two observations per patient. Next, we divided our sample into two groups, which were analysed separately: (1) patients who had two observations at different hospitals but at the same SNF and (2) patients who had two observations at the same hospital but at different SNFs. This allowed us to determine the contributions of hospital performance among patients in group 1 (while holding the SNF constant) and the contributions of SNF performance among patients in group 2 (while holding the hospital constant).

We further limited the sample to claims with hospital admission dates that were more than 30 days but less than 365 days apart to minimise the contribution of changes in underlying health conditions over longer periods of time. To minimise bias introduced by patients who are long-term SNF residents and may have specific post-discharge referral patterns, we excluded patients who had a long-term care stay in the 100 days preceding the index hospital admission (based on the nursing home Minimum Data Set). Additional details on the creation of the sample are available in online supplementary appendix figure 1.

Study variables

Our outcome variable was a discharge-level measure of readmission within 30 days of hospital discharge. We followed Medicare’s definition of hospital-wide readmission from the Hospital Readmission Reduction Programme, which includes unplanned readmissions to any acute care hospital within 30 days of discharge from an index hospitalisation.¹⁶ Planned readmissions that were excluded were those for bone marrow or solid organ transplant, maintenance chemotherapy, rehabilitation or a potentially planned procedure not performed to treat an acute condition or a complication of previous care.

Our independent variables of interest were hospital-level quality and SNF-level quality, using a lagged measure of 30-day risk-standardised readmission rates (RSRRs) for each. Importantly, readmission rates represent only one domain of quality but are publicly reported and used to determine value-based payments. We calculated annual hospital-level and SNF-level RSRRs based on admission-level data and specifications from the CMS for these measures as used in their public reporting programmes.^14,15 We used the methods specified by CMS to calculate RSRRs as a ratio of the predicted readmission to expected readmissions, multiplied by the nationally observed readmission rate. In addition, hospital-wide RSRRs were adjusted for age and 31 comorbidities defined by CMS.¹⁷ For SNF RSRRs, we adjusted for age, sex, length of stay during the prior proximal hospitalisation, dual-eligible status and 31 comorbidities. We measured hospital and SNF RSRRs in the year prior to the discharge of interest in order to minimise endogeneity potentially introduced by measuring a hospital or SNF’s RSRR over the same period of time that we are measuring patient-level readmissions. By measuring the RSRR in the prior year and the risk for readmission in the year of the observed claim, the RSRR is not determined by our primary outcome. Additional details on calculation of hospital and SNF RSRRs are available in online supplementary appendix figure 2.

Finally, we categorised hospitals and SNFs into quartiles based on performance on their RSRRs calculated previously (quartile 1 included the highest-performing hospitals or SNFs, those with the lowest RSRR; quartile 4 included the lowest performing hospitals or SNFs, those with the highest RSRR).

Analysis

First, we compared patient characteristics of age, sex, race, dual-eligible status and common clinical comorbidities between the two groups: (1) patients who had two observations at different hospitals but at the same SNF and (2) patients who had two observations at the same hospital but at different SNFs. We also compared hospital-level and SNF-level characteristics between the two groups.

Next, we estimated the risk-adjusted probability of readmission in the two groups of patients using a linear probability model. For the first group (patients who changed hospitals), 30-day readmission after each hospitalisation was estimated as a function of indicator variables for the performance quartile for the hospital from which the patient was discharged. Conversely, for the second group (patients who changed SNFs), 30-day readmission was estimated as a function of indicator variables for the performance quartile for the SNF to which the patient was discharged. In all regressions, we included patient-level fixed effects to control for all time-invariant patient characteristics and year fixed effects to control for secular trends in readmissions. Models also included time-varying patient characteristics, including age, dual-eligible status and comorbidities. Robust standard errors were clustered at the patient level.

Because each patient is observed twice during the study period, including patient-level fixed effects (or a dummy variable for each patient) allowed each patient to serve as his/her own control. This addresses the unobserved patient characteristics that typically confound the estimated relationship between provider quality and patient outcomes. Thus, in the first group of patients, we constrain time-invariant patient characteristics and SNF characteristics to be the same across the two hospital discharges, allowing hospital performance to vary. In the second group of patients, we constrain time-invariant patient characteristics and hospital characteristics to be the same, only allowing SNF performance to vary. Because hospital and SNF performance (or quartiles of RSRRs) might vary over our study period, in all regressions we controlled for the hospital’s performance at the time of discharge in the cohort of patients who were discharged twice from the same hospital; and the SNF’s performance at the time of discharge for the cohort of patients discharged twice from the same SNF. Together, with patient-level fixed effects, this allows us to isolate the impact of a change in hospital (and, separately, SNF) performance on the outcomes of patients.

Finally, we performed a variety of sensitivity analyses. First, to account for whether patients were choosing to go to a higher-performing hospital or SNF for the second hospitalisation, we repeated our analyses including a variable to control for whether the first hospitalisation or SNF stay was of higher quality than the second. Second, we re-estimated our models including covariates for SNF characteristics (number of beds, provider type, ownership and whether the SNF provider is located in the facility) and hospital characteristics (size, profit status, geographical region, teaching status, rural designation, membership in a system, presence of a medical or surgical intensive care unit, proportion of Medicaid, and Medicare days). Third, we tested an alternate sample definition where we matched on patients’ diagnosis-related groups (DRGs) and included only patients who had the same DRG for both hospitalisations. This analysis was performed to account for potential bias introduced by patients with certain conditions being systematically referred to certain hospitals or SNFs. Finally, to further minimise bias from patients who are frequently hospitalised and might have systematically biassed SNF discharge patterns, we re-estimated our models including beneficiaries with only two hospital and SNF claims, instead of our primary sample, which included the first two claims among beneficiaries who may have had additional claims thereafter.

All analyses were performed using STATA V.15. Patients or the public were not involved in the design, conduct, reporting or dissemination of our research. We deemed statistical significance at a p value less than 0.05.

RESULTS

We identified 140 583 patients who went to different hospitals but the same SNF, and 183 232 patients who went to the same hospital but different SNFs (see table 1 and online supplementary appendix figure 1 for details on sample creation). The cohort of patients who changed hospitals was more likely to be dually eligible for Medicare and Medicaid (36% among patients who changed hospitals vs 21% among patients who changed SNFs). Clinical comorbidities were similarly distributed between both groups.

Table 1.

Patient characteristics

	Different hospitals, same SNF	Same hospital, different SNFs
	N=140 583	N= 183 232
Age (years), mean (SD)	81.7 (7.8)	82.0 (7.8)
Sex, n (%)
Male	50 885 (36)	66 309 (36)
Female	89 698 (64)	116 923 (64)
Race, n (%)
White	120 663 (86)	160 852 (88)
Black/Hispanic	17 168 (12)	19 366 (11)
Other	2752 (2)	3014 (2)
Dual eligible, n (%)	50 539 (36)	39 236 (21)
Comorbidities, n (%)
Diabetes	54 373 (39)	68 834 (38)
CHF	53 342 (38)	72 100 (39)
CAD/CVD	87 793 (62)	111 747 (61)
COPD	42 371 (30)	55 984 (31)
Infection	62 825 (45)	80 708 (44)
Severe or metastatic cancer	9451 (7)	13 639 (7)
Malnutrition	27 393 (19)	34 645 (19)
Liver disease	2507 (2)	3552 (2)
Hemodialysis	2743 (2)	3234 (2)

Percentages may not add to 100 due to rounding.

CAD, coronary artery disease; CHF, congestive heart failure; COPD, chronic obstructive pulmonary disease; CVD, cerebrovascular disease; SNF, skilled nursing facility.

Patients who changed SNFs were less likely to receive care at small hospitals (7% vs 15%) and public hospitals (8% vs 11%) compared with patients who changed hospitals (online supplementary appendix table 1). There were no large differences in SNF characteristics between the two groups (online supplementary appendix table 2). In switching hospitals, 44 241 patients went from a worse to a better hospital, while 42 736 patients went from a better to a worse hospital. In switching SNFs, 61 307 patients went from a worse to a better SNF, while 53 953 patients went from a better to a worse SNF (online supplementary appendix table 3).

The mean hospital-level RSRR in the sample was 15.9% (range 10.3%–31.3%, table 2). Across hospital quartiles, the mean RSRRs were 14.4%, 15.4%, 16.1% and 17.5% for quartiles 1–4, respectively. The mean SNF-level RSRR was 20.6% (range 9.1%–40.9%,). Across SNF quartiles, the mean RSRRs were 16.6%, 18.7%, 20.5% and 24.0% for quartiles 1–4, respectively.

Table 2.

Mean and range of risk-standardised readmission rates across quartiles of hospitals and SNFs

	Hospital quality	SNF quality
Quartile 1 (highest quality)	14.4% (10.3–15.0)	16.6% (9.1–17.9)
Quartile 2	15.4% (15.0–15.7)	18.7% (17.9–19.6)
Quartile 3	16.1% (15.7–16.5)	20.5% (19.6–21.6)
Quartile 4 (lowest quality)	17.5% (16.5–31.3)	24.0% (21.6–40.9)

SNF, skilled nursing facility.

In multivariable regression, among the cohort of patients who changed hospitals (but not SNFs), being discharged from a low-performing versus a high-performing hospital was associated with a significantly higher likelihood of readmission. Based on the regression model estimates, we calculate and present the mean of the patient-level risk-adjusted readmission rate within each of the quality quartiles in graphical form. Being discharged from the highest-performing hospital (quartile 1, lowest RSRR) was associated with a readmission rate of 12.4%, while discharge from a hospital in quartiles 2, 3 or 4 was associated with slightly higher readmission rates of 12.9%, 13.0% and 13.3%, respectively (figure 1). We also present the regression coefficients for each of the indicator variables for performance quartiles between the two groups and associated p values (table 3). These differences were statistically significant for being discharged from the lowest-performing quartile of hospitals in reference to quartile 1 (quartile 4, table 3). However, differences between other quartiles in reference to quartile 1 were not statistically significant.

Figure 1.

Risk-standardised readmission rates for patients going to same hospitals, different SNFs versus different hospitals, same SNFs. SNF, skilled nursing facility.

Table 3.

Percentage point differences in risk-standardised readmission rates by quality quartiles

	Different hospitals, same SNF		Same hospitals, different SNFs
	N=140 583	P value	N=183 232	P value
Quartile 1 (highest quality)	Reference	–	Reference	–
Quartile 2	0.5	0.074	1.2	<0.001
Quartile 3	0.5	0.051	1.4	<0.001
Quartile 4 (lowest quality)	0.9	0.005	2.0	<0.001

SNF, skilled nursing facility.

Among patients who changed SNFs (but not hospitals), the effect of SNF performance on the likelihood of readmission rate was larger than that of hospital performance. Being discharged to the highest-performing SNF (quartile 1, lowest RSRR) was associated with a readmission rate of 13.7%. Discharge to an SNF in quartiles 2, 3 or 4 was associated with progressively higher readmission rates of 14.9%, 15.1% and 15.8%, respectively (figure 1). The differences between the readmission probabilities in each quartile were statistically significant, with the largest one-quartile-change effect being a 0.6 percentage point difference in likelihood of readmission if discharged to quartile 3 versus quartile 4 (table 3), but all comparisons in reference to quartile 1 were statistically significant.

Our sensitivity analyses confirmed these findings. First, when we included an ordinal variable for the order of the two claims to account for the potential role of patient choice in seeking a higher-performing hospital or SNF during the second hospitalisation, we found no differences in our estimates (online supplementary appendix tables 4A and 4B). Including SNF and hospital characteristics resulted in similarly consistent findings (online supplementary appendix tables 5A and 5B), as did matching patients by admission DRGs, though the overall RSRRs are slightly smaller for this group of patients (online supplementary appendix table 6A and 6B). Finally, when limiting our sample to beneficiaries who had only two claims over the study period, we again found no substantive differences from primary analyses, though the overall readmission rates were lower than those of our primary analyses (online supplementary appendix tables 7A and 7B). This likely reflects differences in patient demographics.

DISCUSSION

In this US-based national study of Medicare beneficiaries discharged to SNFs after hospitalisation, we found that the RSRR performance of the SNF had a larger impact on the likelihood of readmission than that of the hospital. When patients changed SNFs across hospital discharges, they were more likely to be readmitted to the hospital if they went to an SNF with higher baseline readmission rates. However, when patients changed hospitals, the impact of changing from a high-performing to low-performing hospital on hospital readmission was much smaller; the magnitude of the effect of hospital performance was less than half of the effect of SNF performance.

Prior work has found that hospital quality, as measured using the single dimension of CMS’ hospital-wide readmission measure, is associated with the likelihood that a patient is readmitted within 30 days across all patients.⁶ We also found that among the subset of patients being discharged to SNFs, patient readmission rates differ between the highest-quality and lowest-quality hospitals, similar to differences found in prior work.⁶ More striking, however, is that the difference in the likelihood of readmission based on hospital quality is dwarfed by the difference in readmission likelihood based on SNF quality, where we find that the difference in readmission rates between the highest-performing and lowest-performing quartiles of SNFs was over twice as large as it was for hospitals.

Our study builds on prior work¹² examining the relationship between hospital or SNF quality and patient outcomes. First, our estimates use the most recent data available in Medicare claims. Second, while prior studies have found similar results, their estimates of differences in patient readmissions may be prone to selection bias in terms of the types of patients who receive care at high-quality or low-quality hospitals or SNFs. Patients who go to hospitals and SNFs with better readmission rates may be different from those who go to hospitals and SNFs with worse readmission rates in ways that are unobservable in the data. Our analysis overcomes this limitation with the use of patient-level fixed effects, allowing us to control for unobservable factors that could potentially confound the relationship between provider quality and patient outcomes.

SNFs provide care in the period during which patients are most vulnerable for rehospitalisation: right after hospital discharge. Thus, it may not be surprising that high-quality SNFs can have an important effect on the likelihood of readmission. Care and discharge planning in the hospital can impact these outcomes; however, the effect of the hospital is likely to be less direct than the effect of SNFs. While SNFs have competing incentives in that they may reap Medicare reimbursement for a repeat SNF stay following a hospital readmission, they also are in a unique position to improve patient outcomes. SNFs often focus on easing the transition between the hospital setting and the posthospital period.¹⁸ When they are effective, they can likely anticipate and manage difficult transitions and complications that could lead to readmission in a way that hospitals cannot. Furthermore, patients spend more time in SNFs than in hospitals^19,20: whereas the average length of hospital stay for patients going to SNF was 7.3 days in 2015, these patients spent an average of 25.1 days in the SNF.²¹ This difference may also result in SNFs having a greater influence on readmission rates when compared with the hospital.

These results have important implications for practice and policy. First, our work, along with prior studies,^11,12,22 supports the idea that SNFs may be able to play a unique role in preventing readmissions among a particularly comorbid and costly patient population. Hospitals are increasingly working to develop preferred networks through informal partnerships with nearby SNFs.²² These findings suggest that hospitals may benefit from preferentially partnering with SNFs that have low RSRRs. Second, as alternative payment models have pushed providers to be more financially accountable for the quality, hospitals have sought ways to constrain cost. Recent studies have found that the use of accountable care organisations and bundled payment is associated with lower rates of institutional postacute care, such as SNFs.^23,24 However, bypassing high-quality postacute care in an attempt to reduce expenditures may not be of greatest benefit to patients.

Our study has limitations. First, there may be residual unmeasured differences in patients, hospitals or SNFs across hospital–SNF stays that could lead to bias. We also cannot make claims toward a causal mechanism due to our cross-sectional study design. Nonetheless, our approach using patient-level fixed effects accounts for unobserved time-invariant patient confounders, and thus produces a higher level of evidence for the relationship between hospital and SNF quality and patient readmission rates than prior studies. Second, we categorised hospital and SNF quality based only on RSRRs, which may not capture all aspects of quality, and thus can only speak to differences in patient outcomes that are associated with aspects of quality captured in these measures. Third, we were unable to characterise the relationships between individual hospitals and SNFs in terms of their degree of integration or fragmentation, which may have contributed to some of the differences in quality we observed.

Fourth, we cannot determine the reasons behind why patients or families might have switched hospitals or SNFs, or how local hospital practices might drive SNF discharge patterns. The act of switching facilities may cause bias in the results if switching is always in one direction (ie, always to a facility of higher or lower quality). To address this concern, we tested models to account for the order of hospitalisations, and these models resulted in consistent estimates as those in our primary analyses. Patients, families and physicians may decide to switch facilities for a variety of reasons, such as location, access to specialty care, quality of facility or prior experiences. SNFs also have the option to not accept patients for care at their facilities. Unfortunately, claims data do not permit the exploration of such details, though they are important and of policy relevance with regard to issues of patient sorting. While the reasons for switching are important and may directly affect policies or practices aimed at directing patients to higher quality facilities, we did not explore the reasons for switching due to the scope of this study.

Finally, given the inclusion criteria for our sample, our study maximised internal validity at the expense of some external validity. The focused sample and our application of patient-level fixed effects help build on prior literature and more precisely estimate the relationship between hospital and SNF quality and patient outcomes. However, they limit our ability to generalise these findings to other populations. In particular, this study examines only patients discharged to SNFs and does not address the degree to which hospital quality broadly influences outcomes for patients who do not require postacute care. Our study also focused only on readmission rates, which represent a single dimension of quality.

In conclusion, we found that discharging patients to higher-quality SNFs had a larger impact on patient outcomes than admitting patients to higher-quality hospitals. Future work is needed to understand how to best capture the elements that drive high quality at SNFs and to scale them appropriately.

Data availability statement

No data are available. No data are available.