Association of Skilled Nursing Facility Participation in a Bundled Payment Model With Institutional Spending for Joint Replacement Surgery

Key Points

Question

What are the outcomes associated with bundled payment models for joint replacement surgery, in which skilled nursing facilities (SNFs) bear financial risk for all care across a 90-day episode beginning with facility admission?

Findings

In this difference-in-difference analysis of 80 648 Medicare patients undergoing lower extremity joint replacement from 2013-2017, the Medicare Bundled Payments for Care Improvement model 3 was associated with a statistically significant decrease in mean institutional spending of $1008 (a combination of postacute care and hospital Medicare-allowed payments) on patients of SNFs receiving bundled payment compared with control SNFs.

Meaning

SNF-based bundled payments may have the potential to lower spending for lower extremity joint replacement.

Abstract

Importance

Medicare recently concluded a national voluntary payment demonstration, Bundled Payments for Care Improvement (BPCI) model 3, in which skilled nursing facilities (SNFs) assumed accountability for patients’ Medicare spending for 90 days from initial SNF admission. There is little evidence on outcomes associated with this novel payment model.

Objective

To evaluate the association of BPCI model 3 with spending, health care utilization, and patient outcomes for Medicare beneficiaries undergoing lower extremity joint replacement (LEJR).

Design, Setting, and Participants

Observational difference-in-difference analysis using Medicare claims from 2013-2017 to evaluate the association of BPCI model 3 with outcomes for 80 648 patients undergoing LEJR. The preintervention period was from January 2013 through September 2013, which was 9 months prior to enrollment of the first BPCI cohort. The postintervention period extended from 3 months post-BPCI enrollment for each SNF through December 2017. BPCI SNFs were matched with control SNFs using propensity score matching on 2013 SNF characteristics.

Exposures

Admission to a BPCI model 3–participating SNF.

Main Outcomes and Measures

The primary outcome was institutional spending, a combination of postacute care and hospital Medicare-allowed payments. Additional outcomes included other categories of spending, changes in case mix, admission volume, home health use, length of stay, and hospital use within 90 days of SNF admission.

Results

There were 448 BPCI SNFs with 18 870 LEJR episodes among 16 837 patients (mean [SD] age, 77.5 [9.4] years; 12 173 [72.3%] women) matched with 1958 control SNFs with 72 005 LEJR episodes among 63 811 patients (mean [SD] age, 77.6 [9.4] years; 46 072 [72.2%] women) in the preintervention and postintervention periods. Seventy-nine percent of matched BPCI SNFs were for-profit facilities, 85% were located in an urban area, and 85% were part of a larger corporate chain. There were no systematic changes in patient case mix or episode volume between BPCI-participating SNFs and controls during the program. Institutional spending decreased from $17 956 to $15 746 in BPCI SNFs and from $17 765 to $16 563 in matched controls, a differential decrease of 5.6% (−$1008 [95% CI, −$1603 to −$414]; P < .001). This decrease was related to a decline in SNF days per beneficiary (from 26.2 to 21.3 days in BPCI SNFs and from 26.3 to 23.4 days in matched controls; differential change, −2.0 days [95% CI, −2.9 to −1.1]). There was no significant change in mortality or 90-day readmissions.

Conclusions and Relevance

Among Medicare patients undergoing lower extremity joint replacement from 2013-2017, the BPCI model 3 was significantly associated with a decrease in mean institutional spending on episodes initiated by admission to SNFs. Further research is needed to assess bundled payments in other clinical contexts.

This study compares mean levels of institutional spending for Medicare beneficiaries after lower extremity joint replacement at skilled nursing facilities (SNFs) participating vs not participating in Medicare’s bundled payment program, before vs after implementation of the program.

Introduction

Use of institutional postacute care, such as skilled nursing facility (SNF) utilization, has grown rapidly since 2000,^1,2 making it a target for spending reduction in new payment models, such as episode-based or bundled payments.³ To date, most bundled payment models have centered on hospitals, which bear financial risk for care during an initial hospital admission and all other care received during the subsequent 30 to 90 days. The majority of savings in these programs has come from reductions in use of institutional postacute care,^3,4 raising the question of whether bundled payments could be effective if postacute care facilities were to assume fiscal responsibility.

Addressing this question, Medicare conducted a national voluntary payment demonstration called Bundled Payments for Care Improvement (BPCI) model 3 from 2013 through 2018.⁵ In this model, postacute care organizations such as SNFs or other rehabilitation facilities assume accountability for patients’ Medicare spending across an episode encompassing their initial SNF admission after an acute hospital stay and including all care (with minor exceptions, such as hospice and a narrowly defined set of medically unrelated readmissions⁶) in the 30, 60, or 90 days following SNF admission. The program was largely composed of SNFs, which comprised 76% of the 1143 postacute care participants in BPCI model 3 in 2017. If episode costs exceeded a bundled payment target price, based on hospitals’ prior 3 years of historical episode spending, the SNF was responsible for part of the excess; if costs were less, the SNF could keep part of the savings.

To date, little evidence exists on the effectiveness of this innovative SNF-based payment model. This analysis evaluated the association of BPCI model 3 enrollment with patient outcomes in SNFs participating in the lower extremity joint replacement (LEJR) bundle, the most common inpatient surgical procedure in Medicare,^{7,8,9,10,11,12} compared with a matched population of control SNFs across the United States.

Methods

Overview of BPCI

BPCI model 3, the focus of this evaluation, was one of 4 participation models in BPCI; the 3 other models were focused on hospitals or physician group practices. SNFs were paid typical fee-for-service rates with quarterly retrospective reconciliation of their episode payments against a target price based on facilities’ own historical spending, minus a discount factor (typically 3%). The study was approved by the Office of Human Research Administration at the Harvard T.H. Chan School of Public Health. The requirement for informed consent was waived because the data were deidentified.

BPCI Participants

BPCI model 3 participants and their dates of participation were identified from publicly available lists of enrolled facilities published by Medicare. All facilities participating in LEJR episodes within BPCI model 3, from initiation of the program in October 2013 through September 2017, were included. These participants were linked to data from Nursing Home Compare and the Medicare CASPER database to obtain facility characteristics of SNFs such as location, nonprofit or for-profit ownership, and size; they were linked to the Area Health Resource File to obtain county-level characteristics based on the physical location of the SNF, such as the median income of the population in the county. SNF claims were used to determine market share, which was calculated as the proportion of all LEJR SNF admissions in the county at an individual facility. To capture the degree of market competition between SNFs, the Herfindahl-Hirschman index for LEJR admissions in each county was used, defined as the sum of squares of the market shares in the county.¹³ BPCI participants could receive patients undergoing LEJR potentially participating in another bundled payment model for LEJR, such as BPCI model 2 (24% of BPCI model 3 episodes during intervention period) or the Comprehensive Care for Joint Replacement model (27% of episodes), which we address in the sensitivity analyses (see Statistical Analysis).

Control Participants

To construct a set of matched control SNFs, SNF-level propensity scores for participation in BPCI were estimated among all nonparticipating SNFs nationally, based on 25 SNF (eg, size, location) and market (eg, concentration, median income, managed care penetration) characteristics and SNF-level trends in patient outcomes, measured as of the first 9 months of 2013 prior to BPCI initiation (eFigure 1 in the Supplement). To avoid potential bias from regression to the mean, SNFs were matched based on their baseline period trends rather than absolute levels of patient-level outcomes such as cost and readmissions.¹⁴ Each SNF in BPCI for LEJR episodes was matched without replacement with up to 5 non-BPCI–participating SNFs in the same state with propensity scores within a 0.05 log-odds “caliper” of BPCI participants. To enable the definition of a post-BPCI period for control SNFs, control SNFs were assigned a BPCI start date using the initiation date of matched BPCI SNFs.

Identifying LEJR Episodes

Consistent with the BPCI definition of an LEJR episode, Medicare inpatient claims from 2013 through 2017 were used to identify all SNF admissions who had been discharged within the prior 30 days from an acute care hospital with a diagnosis-related group code of 469 or 470. The episode began with SNF admission and included all care in the subsequent 90 days. All episodes for Medicare beneficiaries aged 18 years and older, who were continuously enrolled in Medicare Parts A and B during their entire episode of care, were included.

Study Period

Because SNFs could join BPCI on a rolling basis, there were 3 separate time periods for each index case admitted to a SNF enrolled in BPCI (and its matched controls): (1) a common 9-month preintervention period from January 1, 2013-September 30, 2013; (2) a washout period spanning 3 months before enrollment through 3 months after enrollment to allow a quarter before and after BPCI start for SNFs to adapt their behavior; and (3) the postintervention period extending from 3 months post-BPCI enrollment through September 30, 2016, (eFigure 1 in the Supplement). The postintervention period was fixed for BPCI-enrolled SNFs regardless of whether or when they dropped out of the program. An alternative analysis used a definition of the preintervention period, defined as a 9-month period extending from 12 months to 3 months before BPCI entry (or assigned entry) for treatment and control SNFs (eFigure 2 in the Supplement).

Outcomes

The primary outcome was institutional payments per LEJR episode, which included Medicare payments to a hospital (inpatient or outpatient), postacute care facility, and home health agency, as well as spending on durable medical equipment. Institutional spending and all payment outcomes reflect Medicare-allowed payments, which combine Medicare spending and beneficiary cost sharing. Outpatient payments were not included because these data were only available for a 20% sample of beneficiaries, and outpatient payments have not played a significant role in spending differences in prior evaluations.^8,10,11 Total payments were winsorized at the 95th percentile of national episode payments and adjusted for inflation to prices in 2017.

Secondary outcomes included measures of discharge volume, patient case mix (based on patient characteristics including age, race, sex, Medicaid eligibility, and mean Chronic Conditions Warehouse score [a Medicare-provided comorbidity index that ranges from 0 to 27, with higher scores indicating more coexisting conditions]),¹⁵ individual components of institutional episode payment, per-SNF case volume, and initial SNF length of stay. Race, which was self-reported in fixed categories by Medicare beneficiaries, was included due to prior research documenting differences in the receipt of LEJR care by race^16,17,18 and to assess for the potential that SNFs changed their patterns of admission by patients’ race. Outcomes also included the number of days patients had at home during an LEJR episode (days at home, including all days not in a hospital, SNF, inpatient rehabilitation facility, or emergency department, but allowing for home health agency care),¹⁹ the proportion of LEJR episodes discharged to home, the proportion discharged with any home health agency use after initial SNF admission, rates of inpatient readmission, emergency department visits, and mortality within 30 and 90 days after SNF admission.

Statistical Analysis

The characteristics of BPCI and non-BPCI SNFs were compared before and after matching procedures and the standardized mean differences between them. For other analyses, a difference-in-differences approach was used to estimate the change in outcomes associated with BPCI participation among BPCI vs matched control SNFs in the postimplementation vs preimplementation periods. For each primary and secondary outcome, separate patient-level clustered linear regression models were estimated for binary (ie, linear probability models) and continuous outcomes, adjusting for correlation between patients within a SNF using the GENMOD procedure in the SAS statistical software package. The key quantity of interest was the regression coefficient on an interaction between BPCI enrollment status and time, coded as a binary indicator for post- vs pre-BPCI intervention period. This quantity estimates the average differential change in an outcome for BPCI-participating SNFs relative to control matched SNFs. All models adjusted for fixed effects for each match group of BPCI treatment and control SNFs to control for all time-invariant observable and unobservable characteristics of these groups. Beyond indicators for patients’ chronic conditions from the Chronic Condition Warehouse,¹⁵ other controls were not included in the models because they were accounted for in the matching process and by the fixed effects for match groups.

The validity of the difference-in-differences approach was assessed by testing the parallel trends assumption, comparing the slope over time in the pre-BPCI period for BPCI participants and nonparticipants. None of the outcomes had violations of this assumption (eTable 1 in the Supplement).

To assess the degree to which results were influenced by BPCI SNFs that disenrolled from the program, a secondary analysis of the treatment on the treated effect was performed, including only SNFs actively enrolled in BPCI in any quarter. The sensitivity of these results to several different specifications were examined, including an alternative definition of the preintervention period (eFigure 2 in the Supplement), excluding patients admitted to SNFs from hospitals participating in other bundled payment programs, and repeating the main analysis without winsorizing costs. In a post hoc analysis, we also assessed whether there was any change in the difference-in-differences estimate for institutional spending (the primary outcome) between an early (April 2014-September 2016) and late (October 2016-September 2017) BPCI period.

There were no missing data; the beneficiaries with an unknown race (<1%) were classified as other or unknown. Analyses were performed in SAS version 9.4. A P value of less than .05 was considered to be statistically significant for the single primary outcome. For other results, 95% CIs around other reported estimates reflect 0.025 in each tail. CIs were not adjusted for multiple hypothesis testing and should be considered exploratory.

Results

Patient and Facility Characteristics

There were 467 BPCI SNFs with 18 870 LEJR episodes matched to 1958 controls with 72 005 LEJR episodes in the preintervention and postintervention periods, a total of 80 648 beneficiaries with LEJR (Table 1). Key characteristics were that 79% of matched BPCI SNFs were for-profit facilities, 85% were located in an urban area, and 85% were part of a chain. In the matched BPCI SNF group, 16 837 beneficiaries with LEJR at a BPCI SNF had mean (SD) age of 77.5 (9.4) years and 12 173 (72.3%) were women; 63 811 beneficiaries at matched control SNFs had mean (SD) age of 77.6 (9.4) years and 46 072 (72.2%) were women. Propensity score matching improved the balance on observable SNF, patient, and market characteristics with all standardized mean differences between BPCI SNFs and control SNFs decreasing to 0.1 or less.

Table 1. SNF Characteristics Before and After Matching.

	SNFs		Standardized mean difference	Matched SNFs		Standardized mean difference
	BPCI (n=467)	Non-BPCI (n=12 670)		BPCI (n=448)	Non-BPCI (control) (n=1958)
SNF characteristics
No. of LEJR admissions, preintervention period to 9 mo	14.57	7.7	0.19	10.8	9.9	0.06
For profit, %	79.2	69.2	0.21	79.0	75.0	0.09
Nonprofit, %	18.4	24.8	0.15	18.5	20.6	0.05
Public, %	2.4	5.9	0.15	2.5	4.3	0.10
Urban, %a	85.4	71.6	0.30	85.0	85.1	0.00
No. of federally certified beds	115.1	110	0.09	115	117	0.04
Part of a chain, %	85.7	56.6	0.58	85	82	0.08
Hospital-based, %	1.1	4.1	0.17	0.9	1.0	0.01
Region, %
Northeast	25.1	17.2	0.22	25	22	0.08
Midwest	31.5	32.7	0.02	32	34	0.05
South	26.8	35.8	0.20	26	28	0.04
West	16.1	14.3	0.06	16	15	0.03
Overall Medicare star ratingb	3.5	3.3	0.11	3.5	3.4	0.04
Monthly trends in patient-level outcomesc
SNF days	−0.023	−0.01	0.07	−0.02	−0.02	0.02
Home health agency visits	0.0003	−0.002	0.05	−0.0004	−0.0003	0.00
Hospital readmission rate, %	−0.001	−0.13	0.02	−0.13	−0.13	0.00
Mortality, %	−0.0002	−0.03	0.03	−0.02	−0.02	0.02
Institutional spending, $	−8.64	1.35	0.08	−7.61	−3.80	0.03
County-level characteristicsd
Population aged 65 and older	105 778	95 806	0.05	104 589	94 629	0.07
Median household income, $	53 814	51 539	0.17	53 805	53 226	0.04
% Medicare Advantage penetration	29.22	25.9	0.24	29.3	30.2	0.07
SNF beds per 10 000	416.6	467.6	0.21	420.1	428.1	0.04
No. of inpatient rehabilitation facilitiese	0.53	0.46	0.07	0.54	0.46	0.07
Mean SNF market share in countyf	0.16	0.21	0.20	0.15	0.14	0.02
Herfindahl-Hirshman Index for SNFsg	0.08	0.08	0.06	0.08	0.07	0.06

Abbreviations: BPCI, Bundled Payment for Care Improvement; LEJR, lower extremity joint replacement; SNF, skilled nursing facility.

^aUrban location defined using the Health Resources and Service Administration rural-urban commuting area code database (http://depts.washington.edu/uwruca/index.php). Urban was defined as a patient residing in a metropolitan ZIP code.²⁰

^bMedicare star ratings range from 1 (much below average) to 5 (much above average). This score is a composite ranking of individual SNFs that incorporates multiple measures of SNF quality, staffing, and health inspection performance.²⁵

^cMeasured using claims data from the baseline period (January 2013-September 2013). To improve stability of estimates, trends were estimated based on all Medicare admissions to each SNF. All outcomes were measured in the period of BPCI model 3 bundles (ie, 90-day post SNF admission).

^dAssessed in 2013 using the Area Health Resource File, except for SNF market share (see footnote e).

^eInpatient rehabilitation facilities are a separate type of facility from SNFs that provide postacute care for patients requiring higher-intensity therapy. They can be competitors to SNFs for market share for patients undergoing LEJR.

^fIndicates the proportion of all Medicare SNF admissions in a county that occur at a single SNF. The mean county-level SNF market share is the average market share of all SNFs in that county.

^gThis index is commonly used to measure market concentration, which was defined using the sum of squares of the market shares in the county.¹³

Patient Case Mix and Volume

With the exception of an increasing proportion of female patients and patients with dementia, there were no significant differential changes in the composition of admitted patients undergoing LEJR between BPCI and control SNFs, as measured across 15 separate patient characteristics (Table 2) including markers of social and medical complexity, specifically the proportion of patients who were Medicaid eligible and patients’ average number of chronic conditions. There was no significant differential change in mean quarterly discharges per SNF between BPCI and control SNFs in the post period (0.0; 95% CI, −0.3 to 0.4).

Table 2. Changes in Volume and Case Mix.

	BPCI			Matched controls			Adjusted differential change in case mix, difference in differences (95% CI)a
	Baseline period	Intervention period	Unadjusted difference	Baseline period	Intervention period	Unadjusted difference
No. of SNFs	448	448		1958	1958
No. of beneficiaries	5075	11 762		20 509	43 302
No. of episodes	5556	13 314		22 562	49 443
Quarterly discharges per SNF, mean (SD)	4.1 (5.4)	2.1 (4.4)	−2.0	3.8 (5.8)	1.8 (4.1)	−2.0	0.0 (−0.3 to 0.4)
Patient characteristics, %
Age, y
≤64	6.2	6.2	0.0	6.5	6.0	−0.5	0.5 (−0.5 to 1.4)
65-79	51.1	50.4	−0.7	50.4	49.5	−0.9	0.2 (−1.9 to 2.3)
≥80	42.7	43.4	0.7	43.1	44.5	1.3	−0.7 (−2.7 to 1.4)
Men	28.3	27.6	−0.7	26.8	28.0	1.2	−1.9 (−3.5 to −0.3)
Women	71.7	72.4	0.7	73.2	72.0	−1.2	1.9 (0.3 to 3.5)
Medicaid dual eligible	16.9	16.6	−0.3	17.5	17.3	−0.3	0.0 (−1.8 to 1.8)
Disabled without end-stage kidney disease	16.2	16.8	0.6	15.7	16.4	0.7	−0.1 (−1.5 to 1.4)
Frailb	9.4	10.4	1.0	9.7	10.8	1.1	−0.06 (−1.41 to 1.29)
Dementiac	4.3	4.4	0.0	5.2	4.4	−0.8	0.8 (0.0 to 1.62)
No. of chronic conditionsd	5.0	5.1	0.1	5.0	5.1	0.1	0.00 (−0.12 to 0.12)
Race/ethnicity
White	89.0	88.8	−0.2	90.0	89.6	−0.4	0.2 (−1.3 to 1.6)
Black	7.3	7.2	−0.2	5.8	6.1	0.2	−0.4 (−1.6 to 0.8)
Hispanic	1.0	0.9	0.0	1.4	1.3	−0.1	0.1 (−0.4 to 0.5)
Unknown or othere	2.7	3.2	0.5	2.8	3.1	0.3	0.2 (−0.4 to 0.8)
Level of complexity, %f
DRG without complications or comorbiditiesf	91.7	90.0	−1.7	91.6	89.7	−1.9	0.21 (−0.9 to 1.3)
DRG with major complications or comorbiditiesg	8.3	10.0	1.7	8.4	10.3	1.9	−0.21 (−1.3 to 0.9)

Abbreviations: BPCI, Bundled Payment for Care Improvement; DRG, diagnosis-related group; LEJR, lower extremity joint replacement; SNF, skilled nursing facility.

^aAdjusted estimates show the differential change in the proportion of patient characteristics between treatment and control groups after vs before BPCI implementation. All estimates use standard errors clustered at the SNF level. The differential change estimates can be interpreted as the average within-SNF change attributable to BPCI implementation.

^bFrailty was defined using a previously validated claims-based frailty index from Kim et al.^21,22 Patients in the top 10% of the frailty index over the entire population of SNF admissions were classified as frail.

^cDementia was defined using International Classification of Diseases-Ninth Revision codes described in Goodman et al.²³ Patients with not otherwise-specified dementia diagnoses were included because routine use of nonspecific diagnostic codes is common, such that excluding them would result in omission of many affected individuals.

^dThe presence of 27 conditions was gathered from the Chronic Condition Data Warehouse, which uses claims since 1999 to describe Medicare beneficiaries’ accumulated chronic disease burden. Chronic conditions were defined as any condition present by the end of the calendar year prior to the LEJR episode. Conditions included Alzheimer disease, Alzheimer disease and related disorders or senile dementia, anemia, asthma, atrial fibrillation, benign prostatic hyperplasia, breast cancer, cataract, chronic kidney disease, chronic obstructive pulmonary disease, colorectal cancer, depression, diabetes, endometrial cancer, glaucoma, heart failure, hip or pelvic fracture, hyperlipidemia, hypertension, hypothyroidism, ischemic heart disease, lung cancer, osteoporosis, prostate cancer, acute myocardial infarction, rheumatoid arthritis, and stroke or transient ischemic attack.

^eThe category of Other includes American Indian or Alaska Native, Asian, Native Hawaiian or Pacific Islander, these are the options for self-identified race/ethnicity in the Medicare enrollment file besides White and Black race or Hispanic/Latino. Less than 1% of beneficiaries are coded as unknown race.

^fWithout complications or comorbidities refers to DRG 470 (LEJR without major complications or coexisting conditions).

^gWith major complications or comorbidities refers to DRG 469 (LEJR with major complications or coexisting comorbidities).

Medicare Spending

Primary Outcome

Institutional spending (the primary outcome) decreased significantly more in participating SNFs compared with controls (from $17 956 to $15 746 in BPCI SNFs and from $17 765 to $16 563 in matched controls; differential change, −$1008 [95% CI, −$1603 to −$414]; P < .001), a −5.6% relative decrease compared with BPCI SNF baseline spending (Table 3). In a post hoc analysis, the differential decline in institutional spending was comparable between the early BPCI period (April 2014-September 2016, −$900 [95% CI, −$1549 to −$252] and late BPCI period (October 2016-September 2017, −$1129 [95% CI, −$1801 to −$457]; eFigure 3 in the Supplement).

Table 3. Differential Change in Total Episode Spending and Components of Spending^a.

	BPCI, $			Matched controls, $			Adjusted difference in differences (95% CI)b	P valuec
	Baseline period	Intervention period	Unadjusted difference	Baseline period	Intervention period	Unadjusted difference
Primary outcome
Total institutional spending over 90-d episoded	17 956	15 746	−2210	17 765	16 563	−1201	−1008 (−1603 to −414)	<.001
Secondary outcomes
Overall SNF spending (initial + subsequent)	13 959	11 848	−2111	13 907	12 776	−1130	−981 (−1446 to −516)
Initial SNF stay	12 150	10 148	−2002	12 012	10 856	−1155	−847 (−1253 to −440)
Subsequent SNF stays	1809	1700	−109	1895	1920	25	−134 (−381 to 113)
Readmission	1974	1952	−22	1976	2060	84	−106 (−315 to 103)
Inpatient rehabilitation facility	219	141	−78	129	138	9	−87 (−257 to 84)
Long-term acute care hospital	129	219	90	150	192	42	48 (−28 to 125)
Home health agency	1628	1559	−68	1556	1371	−185	117 (38 to 196)
Durable medical equipment	48	27	−21	47	26	−20	0 (−3 to 3)

Abbreviations: BPCI, Bundled Payment for Care Improvement; SNF, skilled nursing facility.

^aAll estimates use standard errors clustered at the SNF level.

^bAdjusted estimates show the differential change in spending outcomes during the 90-day postadmission bundle period between treatment and control groups after vs before BPCI implementation. The differential change estimates can be interpreted as the average within-SNF change attributable to BPCI implementation.

^cOnly 1 P value is shown for the primary outcome of institutional spending.

^dInstitutional spending is any Medicare payments to a hospital (inpatient or outpatient), post–acute care facility, home health agency, hospice, as well as spending on durable medical equipment.

Secondary Outcomes

The reduction in institutional spending was largely driven by significantly reduced spending on overall SNF spending (from $13 959 to $11 848 in BPCI SNFs and from $13 907 to $12 776 in matched controls; differential change, −$981 [95% CI, −$1446 to −$516]) due to lower spending during the initial SNF stay (from $12 150 to $10 148 in BPCI SNFs and from $12 012 to $10 856 in matched controls; differential change, −$847 [95% CI, −$1253 to −$440]) with no significant reduction from subsequent SNF stays (from $1809 to $1700 in BPCI SNFs and from $1895 to $1920 in matched controls; differential change, −$134 [95% CI, −$381 to $113]). There were no significant changes in spending on readmissions, inpatient rehabilitation facility, long-term acute care hospitals, or durable medical equipment.

Postacute Care Use and Patient Outcomes

Overall length of stay decreased significantly in episodes at BPCI SNFs across the full bundled period by −2.0 days (95% CI, −2.9 to −1.1; [from 26.2 to 21.3 in BPCI SNFs and from 26.3 to 23.4 in matched controls]; Table 4) postperiod compared with control SNFs, which was almost entirely driven by a significant reduction in the initial SNF stay (from 21.9 to 17.4 in BPCI SNFs and from 22.0 to 19.1 in matched controls; differential change, −1.5 days; 95% CI, −2.4, −0.7). The number of days spent at home during the 90-day episode increased by 2.3 (95% CI, 1.2 to 3.3; [from 59.9 to 65.2 in BPCI SNFs and from 59.9 to 62.9 in matched controls]) and home health agency use on SNF discharge also increased (from 51.5% to 59.1% in BPCI SNFs and from 49.2% to 53.2% in matched controls; differential change, 3.6% [95% CI, 1.4% to 5.8%]) for patients admitted to BPCI SNFs compared with controls, both statistically significant.

Table 4. Differential Changes in Postacute Care Use and Discharge Destination^a.

	BPCI			Matched controls			Adjusted difference in differences (95% CI)b
	Baseline period	Intervention period	Unadjusted difference	Baseline period	Intervention period	Unadjusted difference
Average length of stay, initial SNF admission, d	21.9	17.4	−4.5	22.0	19.1	−3.0	−1.5 (−2.4 to −0.7)
Average No. of SNF days, full bundle period, d	26.2	21.3	−4.9	26.3	23.4	−2.9	−2.0 (−2.9 to −1.1)
Average days at home during bundle period, dc	59.9	65.2	5.3	59.9	62.9	3.0	2.3 (1.2 to 3.3)
Discharged home after initial SNF stay, %d	68.1	63.3	−4.7	68.1	65.5	−2.6	−2.2 (−6.1 to 1.8)
Use of home health agency after SNF discharge, %e	51.5	59.1	7.5	49.2	53.2	3.9	3.6 (1.4 to 5.8)

Abbreviations: BPCI, Bundled Payment for Care Improvement; SNF, skilled nursing facility.

^aAll estimates use standard errors clustered at the SNF level.

^bAdjusted estimates show the differential change in utilization outcomes during the 90-day postadmission bundle period between treatment and control groups after vs before BPCI implementation. The differential change estimates can be interpreted as the average within-SNF change attributable to BPCI implementation.

^cDays at home defined as the number of days during the 90-day post-SNF admission bundle period that were not spent in a SNF, emergency department, or hospital.

^dDischarge home defined using the disposition status in SNF claims on discharge.

^eUse of home health agency defined by the presence of any paid home health claim after SNF discharge.

There was no significant change in patient outcomes in the BPCI vs matched control SNFs, including emergency department visits, inpatient readmissions, and 30-day or 90-day patient mortality (Table 5).

Table 5. Differential Changes in Hospital Use and Mortality.

	BPCI, %			Matched controls, %			Adjusted difference in differences (95% CI)
	Baseline period	Intervention period	Unadjusted difference	Baseline period	Intervention period	Unadjusted difference
ED use without admissiona
30 d	3.9	5.6	1.7	4.4	5.4	1.0	0.7 (−0.9 to 2.2)
90 d	12.4	16.0	3.6	12.7	14.5	1.8	1.8 (−0.7 to 4.3)
Inpatient readmission
30 d	8.8	9.0	0.1	9.0	9.6	0.5	−0.4 (−1.4 to 0.6)
90 d	16.4	16.6	0.2	16.5	17.3	0.7	−0.5 (−1.9 to 0.8)
Mortality
30 d	1.9	1.6	−0.4	1.8	1.7	−0.1	−0.2 (−0.7 to 0.3)
90 d	4.5	3.6	−0.9	4.6	4.1	−0.5	−0.4 (−1.2 to 0.4)

Abbreviations: BPCI, Bundled Payment for Care Improvement; ED, emergency department; SNF, skilled nursing facility.

^aAll estimates use standard errors clustered at the SNF level.

^bAdjusted estimates show the differential change in patient outcomes during the 90-day postadmission bundle period between treatment and control groups after vs before BPCI implementation. The differential change estimates can be interpreted as the average within-SNF change attributable to BPCI implementation.

^cED visits without admission were defined using professional claims in the Medicare Part B carrier file, which was only available for a 20% random sample of patients.

Sensitivity Analyses

In a secondary analysis of spending and utilization outcomes using a “treatment on the treated” approach that excluded BPCI SNFs after disenrolling from the program, there were no qualitative changes in the results (eTable 2 in the Supplement). Results were also similar in sensitivity analyses including using a 9-month preintervention period that preceded each BPCI SNF’s entry month to estimate difference-in-differences models, excluding all episodes from BPCI model 2 or Comprehensive Care for Joint Replacement hospitals, and replicating analyses without winsorizing outlier costs (eTable 3 in the Supplement).

Discussion

In a novel SNF-based bundled payment model for LEJR, program participation was associated with decreased Medicare spending on institutional care without evidence of increased hospitalizations or mortality. The spending decrease was primarily driven by shortening patients’ length of stay during the initial SNF stay, with a resulting decrease in SNF spending. BPCI participation was not associated with systematic evidence of patient selection or increased volume. These findings suggest that bundled payments may be a promising SNF payment model for lower extremity joint replacement.

These results do not account for the cost of reconciliation payments to SNFs, which were estimated to exceed the overall savings by federal evaluators considering the entire BPCI model 3 program including all conditions. Moreover, reconciliation payments in the beginning of the BPCI model 3 program were greater than planned because Medicare eliminated downside risk in the first 5 quarters of the program due to problems with the calculation of target prices.¹⁰

Almost all bundled payment programs to date have focused on hospitals and physician group practices rather than SNFs or other postacute care organizations. Model 3 of BPCI is the notable exception. Part of the reason may be the perception that hospitals or physician groups are more tightly organized, have greater resources for care redesign, and have closer control over decisions about discharge and subsequent ambulatory care. Also, many patients admitted to the hospital are never treated in a postacute care setting. Another reason may be the potential for SNFs to exercise discretion in their admissions to generate savings in a bundled payment model. Acknowledging these key differences between SNFs and hospitals or physician groups, the findings in this study suggest that postacute care organizations may be an important group on which to focus for future payment models.

Hospitals and SNFs may have complementary levers for success in lowering costs under bundled payments. In prior work on hospital-based bundled payments for LEJR, hospitals in both BPCI and the Comprehensive Care for Joint Replacement Program were more likely than control hospitals to discharge patients home without postacute care services, which was the primary driver of savings.^8,24 The savings per episode in those models were similar in magnitude (ranging from $435 to $1116)^6,8,9,11,12 to the overall savings per episode found in this study ($1008). In this study, participating SNFs reduced Medicare spending by reducing payment for SNF stays themselves through shorter length of stay. This came with the tradeoff of somewhat higher use of home health agencies, which could serve as a substitute for patients requiring low-intensity rehab in the final days of what might have been a longer SNF stay. Used in different circumstances, bundled payment models centered on hospitals and SNFs could create synergistic savings.

To date, to our knowledge there are no prior demonstrations of episode-based payment focused on SNFs. Despite some methodological differences, these overall findings are consistent with results from the federally contracted evaluation of BPCI model 3, which also found a reduction in Medicare-allowed costs.⁶ The savings they found were largely from payments to SNFs, although they did not separate out differences in treatment during the initial index SNF stay vs subsequent stays at SNFs during the 90-day episode.

Limitations

This study has several limitations. First, this analysis is focused on LEJR, which is 1 surgical condition among 48 surgical and medical conditions in BPCI. These conclusions may not apply in other clinical contexts, although as the most common surgical procedure in Medicare beneficiaries, LEJR has policy relevance as a stand-alone procedure. Second, BPCI is also a voluntary program. As such, unobserved differences likely exist between volunteering SNFs and control SNFs that could bias these results. The results may also not generalize to SNFs that differ from the profile of BPCI participants in this study. Propensity score matching within the same state, and match-group fixed effects within the regression models controlled for as many of these factors as possible, but unobserved confounding could remain. Third, outcomes such as functional status, time to recovery, burden on family caregivers, or patient satisfaction that may be particularly relevant for a surgical procedure associated with mobility and function like LEJR were not captured. The outcomes reported reflect serious health exacerbations but may not reflect on the quality of surgical care or rehabilitation itself. Fourth, as more recent administrative data become available, it will be important to examine the longer-term outcomes of this program beyond 2017.

Conclusions

The BPCI model 3 was significantly associated with a decrease in mean institutional spending on LEJR episodes initiated by admission to SNFs. Further research is needed to assess bundled payments in other clinical contexts.

Supplement.

eFigure 1. Schematic of Pre- and Post-intervention Period Definition for Main Analysis, 2013 Baseline

eFigure 2. Schematic of Pre- and Post-intervention Study Periods, Sensitivity Analysis

eFigure 3. Trends in Institutional Payment and Difference-in-Differences Estimates in BPCI Model 3 During the Study Period

eTable 1. Pre-period Outcome Trends, Intervention and Matched Control BPCI SNFs

eTable 2. Treatment-on-Treated Sensitivity Analyses

eTable 3. Other Sensitivity Analyses