Clean Up City of St. Augustine, Florida: DOES PRIVATE EQUITY INVESTMENT IN HEALTHCARE BENEFIT PATIENTS? EVIDENCE FROM NURSING HOMES (Atul Gupta, Sabrina T. Howell, Constantine Yannelis & Abhinav Gupta, National Bureau of Economic Research, Working Paper 28474, February 2021)

https://www.jeffdowney.com/wp-content/uploads/2021/03/w28474.pdfxxxxx

NBER WORKING PAPER SERIES

DOES PRIVATE EQUITY INVESTMENT IN HEALTHCARE BENEFIT PATIENTS?

EVIDENCE FROM NURSING HOMES

Atul Gupta

Sabrina T. Howell

Constantine Yannelis

Abhinav Gupta

Working Paper 28474

http://www.nber.org/papers/w28474

NATIONAL BUREAU OF ECONOMIC RESEARCH

1050 Massachusetts Avenue

Cambridge, MA 02138

February 2021

We are grateful to Abby Alpert, Pierre Azoulay, Zack Cooper, Liran Einav, Matthew Grennan,

Arpit Gupta, Jarrad Harford, Steve Kaplan, Holger Mueller, Aviv Nevo, Adam Sacarny, Albert

Sheen, Arthur Robin Williams, and participants at the NBER Aging Summer Institute, American

Economics Association, UBC Winter Finance Conference, Whistler Junior Health Economics

Summit, Virtual Corporate Finance Seminar, ESSEC, Norwegian School of Economics,

Oklahoma Price, PE Research Consortium (PERC) Annual Symposium, Tulane, NYU Stern, U

Penn Wharton, Yale Health Policy, Oregon Lundquist, and Johns Hopkins Carey for their

comments and suggestions. Jun Wong, Mei-Lynn Hua, and Sarah Schutz provided excellent

research assistance. We thank Christianna Williams of Abt Associates for insights into CMS

quality measures, Liz Liberman for industry information, and Mohan Ramanujan at NBER for

assistance with the CMS data. Funding from the Wharton Dean’s Research Fund and the Mack

Institute (Gupta), the Kauffman Foundation (Howell), and the Fama Miller Center at the

University of Chicago (Yannelis) is greatly appreciated. We gratefully acknowledge funding

through National Institute of Aging pilot grant P01AG005842-31. All remaining errors are our

own. The views expressed herein are those of the authors and do not necessarily reflect the views

of the National Bureau of Economic Research.

NBER working papers are circulated for discussion and comment purposes. They have not been

peer-reviewed or been subject to the review by the NBER Board of Directors that accompanies

official NBER publications.

reserved. Short sections of text, not to exceed two paragraphs, may be quoted without explicit

permission provided that full credit, including © notice, is given to the source.Does Private Equity Investment in Healthcare Benefit Patients? Evidence from Nursing Homes

Atul Gupta, Sabrina T. Howell, Constantine Yannelis, and Abhinav Gupta

NBER Working Paper No. 28474

February 2021

JEL No. G3,G32,G34,G38,I1,I18

ABSTRACT

The past two decades have seen a rapid increase in Private Equity (PE) investment in healthcare,

a sector in which intensive government subsidy and market frictions could lead high-powered for-

profit incentives to be misaligned with the social goal of affordable, quality care. This paper

studies the effects of PE ownership on patient welfare at nursing homes. With administrative

patient-level data, we use a within-facility differences-in-differences design to address non-

random targeting of facilities. We use an instrumental variables strategy to control for the

selection of patients into nursing homes. Our estimates show that PE ownership increases the

short-term mortality of Medicare patients by 10%, implying 20,150 lives lost due to PE

ownership over our twelve-year sample period. This is accompanied by declines in other

measures of patient well-being, such as lower mobility, while taxpayer spending per patient

episode increases by 11%. We observe operational changes that help to explain these effects,

including declines in nursing staff and compliance with standards. Finally, we document a

systematic shift in operating costs post-acquisition toward non-patient care items such as

monitoring fees, interest, and lease payments.

Atul Gupta

Wharton Health Care Management

3641 Locust Walk, CPC 306

Philadelphia, PA 19104

atulgup@wharton.upenn.edu

Sabrina T. Howell

NYU Stern School of Business

KMC 9-93

44 West 4th Street

New York, NY 10012

and NBER

showell@stern.nyu.edu

Constantine Yannelis

Booth School of Business

University of Chicago

5807 S. Woodlawn Avenue

Chicago, IL 60637

and NBER

constantine.yannelis@chicagobooth.edu

Abhinav Gupta

New York University

agupta4@stern.nyu.edu1 Introduction

The U.S. spends more than other developed countries on healthcare, yet has worse health

outcomes (Garber and Skinner, 2008). In light of evidence from other sectors, private equity

(PE) ownership of healthcare providers could improve productivity (Kaplan, 1989; Davis

et al., 2014; Bloom et al., 2015b; Bernstein and Sheen, 2016). PE ownership can aﬀect

firm performance because it confers distinct incentives to quickly and substantially increase

firm value.1 However, PE’s success in other sectors may not be relevant to healthcare,

which suﬀers from unique market frictions. For example, patients cannot accurately assess

provider quality, they typically do not pay for services directly, and a web of government

agencies act as both payers and regulators (Cutler, 2011; Skinner, 2011). These features

weaken the natural ability of a market to align firm incentives with consumer welfare and

could mean that high-powered incentives to maximize profits have detrimental implications

for consumer welfare (Hansmann, 1980; Hart et al., 1997; Chandra et al., 2016).

Policymakers appear increasingly concerned about this possibility given the rapid

growth of PE in healthcare. For example, in 2019 U.S. Senators asked about “the role of

PE firms in the nursing home care industry, and the extent to which these firms’ emphasis

on profits and short-term return is responsible for declines in quality of care," while a

member of the British Parliament argued that PE-owned nursing homes in the UK pursue

"profiteering, cost and corner cutting, all the while their owners are loading them up on

debt with high interest rates and expecting the taxpayer to pay when it fails" (Brown et al.,

2019; Hodgson, 2020).2 Meanwhile, voices from the private sector often paint a diﬀerent

picture; for example, a 2019 report from consulting firm EY concluded that “Not only is

PE perceived to have a beneficial overall impact on health care businesses, it is also

considered to positively influence the focus on quality and clinical services" (Saenz, 2019).

This debate has come to the fore in part because of rising PE activity in U.S. healthcare

1See Section 2.2 for an explanation of PE and its incentive structure.

2Policymaker concern has focused not only on nursing home quality (Doggett et al., 2020), but also surprise

medical bills (Cooper et al., 2020) and predatory acquisitions of physician practices (King, 2020), among other

areas. Also see Flood (2019) and Lewis (2019).

1over the last two decades, with total investment increasing from less than $5 billion in 2000

to more than $100 billion in 2018 (Appelbaum and Batt, 2020). PE-owned firms provide the

staﬃng for more than one-third of emergency rooms, own large hospital and nursing home

chains, and are rapidly expanding ownership of physician practices.3 Thus far, evidence to

inform the animated policy discussion is limited and inconclusive.

We focus on nursing homes in the U.S., which represent a large sector with spending at

$166 billion in 2017 projected to grow to $240 billion by 2025 (Martin et al., 2018).

Nursing homes oﬀer an attractive setting in which to examine the impact of PE in

healthcare. First, they have historically had a high rate of for-profit ownership (about 70%),

allowing us to study the eﬀects of PE ownership relative to for-profit ownership more

generally. Second, PE firms have acquired both large chains and independent facilities,

enabling us to make progress in isolating the eﬀects of PE ownership from the related

phenomenon of corporatization in medical care (Eliason et al., 2020). We use patient- and

facility-level administrative data from the Centers for Medicare & Medicaid Services

(CMS), which we match to PE deal data. The data include 18,485 unique nursing homes

between 2000 and 2017. Of these, 1,674 were acquired by PE firms in 128 unique deals.

We observe about 7.4 million unique Medicare patients.

We overcome two empirical challenges to estimating causal eﬀects of PE ownership.

The first is non-random selection of acquisition targets. To address this we include facility

fixed eﬀects, which also eliminate cross-sectional diﬀerences in the types of locations where

PE firms tend to buy nursing homes. The second challenge is unobserved changes in patient

composition following PE ownership, perhaps reflecting new advertising, hospital ties, or

patient reactions to quality. We control for the patient-facility match with a diﬀerential

distance instrumental variables (IV) strategy (McClellan et al., 1994; Grabowski et al., 2013;

Card et al., 2019), exploiting patient preference for a nursing facility close to their home (the

median distance is 4.6 miles). To our knowledge, no national study on PE or on ownership

in healthcare has simultaneously addressed both challenges.

A key measure of patient welfare is short-term survival. We find that going to a PE-

3See Bruch et al. (2020); Brown et al. (2020); Casalino (2020).

2owned nursing home increases the probability of death during the stay and the following

90 days by 1.7 percentage points, about 10% of the mean. This estimate implies about

20,150 Medicare lives lost due to PE ownership of nursing homes during our sample period.

We use the observed age and gender distribution of Medicare decedents to estimate the

corresponding implied loss in life-years – 160,000. Using a conventional value of a life-

year from the literature, this estimate implies a mortality cost of about $21 billion in 2016

dollars. To put this in perspective, this is about twice the total payments made by Medicare

to PE facilities during our sample period, about $ 9 billion.

The mortality eﬀect is concentrated among older patients, especially those with

relatively low disease burdens. Exploring treatment eﬀect heterogeneity more formally

using Marginal Treatment Eﬀect (MTE) analysis, we find evidence of reverse selection on

treatment gains, i.e., patients with the lowest unobserved resistance to going to a PE-owned

facility experience the highest increase in mortality (nearly 4 pp). We estimate an

unconditional Average Treatment Eﬀect (ATE) of about 1.3 pp. Hence, the conclusion that

patients are worse oﬀ at PE-owned facilities can be generalized beyond compliers to the

average Medicare patient. However, we also estimate negative MTE values for patients

with the highest unobserved resistance, implying that a small fraction of patients are better

oﬀ receiving care at such facilities.

The eﬀect on mortality is robust to a battery of specification checks, and does not

appear in a placebo analysis testing for pre-buyout eﬀects. It also remains intact when we

restrict our attention to PE acquisitions of the largest chains, in which chain size remained

constant over the sample period, implying that the eﬀect reflects the nature of ownership

rather than consolidation or corporatization more broadly. To ensure the eﬀect is not

spurious, we study other measures of patient well-being using the same IV approach. We

find that going to a PE-owned nursing home increases the probability of taking

antipsychotic medications – discouraged in the elderly due to their association with greater

mortality – by 50%. Similarly, patient mobility declines and pain intensity increases

post-acquisition. Finally, the amount billed per 90-day episode increases by 11%. Taken

together, these results suggest that PE ownership decreases nursing home productivity, as

3measured by our proxies for quality output per dollar spent.

To explore mechanisms for the eﬀect on mortality, we assess operational changes using

facility-level data. Here we are limited to using a diﬀerences-in-diﬀerences research design,

which has been standard in the literature on PE eﬀects. We find negative eﬀects on facility

Five Star ratings, which are constructed by CMS to provide summary information on quality

of care. We next consider nurse availability, which is the most important determinant of

quality of care (Zhang and Grabowski, 2004; Lin, 2014). We find that PE ownership leads

to a 3% decline in hours per patient-day supplied by the frontline nursing assistants who

provide the vast majority of caregiving hours and perform crucial well-being services such

as mobility assistance, personal interaction, and cleaning to minimize infection risk and

ensure sanitary conditions. Overall staﬃng declines by 1.4%.

The loss of front-line staﬀ is most problematic for older but relatively less sick patients,

who drive the mortality result. There may be less scope to reduce the costs of care for

the sicker patients, as they have explicit medical needs. Elevated antipsychotic use could

also be partly explained as a substitution response to lower nurse availability (Cawley et

al., 2006). We can explain about a third of the mortality eﬀect using previously published

mortality eﬀects of antipsychotics and lower nurse availability and assuming these factors

are additive (Schneider et al., 2005; Tong, 2011). However, this may be an understatement

if these factors are more harmful when they interact.

Finally, we assess facility finances to shed light on how the financial strategies particular

to the PE industry aﬀect operations. A puzzle is why nursing homes are attractive targets

given their low and regulated profit margins, often cited at just 1-2%. Using CMS cost

reports, we find that there is no eﬀect of buyouts on net income, raising the question of how

PE firms create value. There are three types of expenditures that are particularly associated

with PE profits and tax strategies: “monitoring fees" charged to portfolio companies, lease

payments after real estate is sold to generate cash flows, and interest payments reflecting the

importance of leverage in the PE business model (Metrick and Yasuda, 2010; Phalippou et

al., 2018). We find that all three types of expenditures increase after buyouts, with interest

payments rising by over 300%. These results, along with the decline in nurse availability,

4suggest a systematic shift in operating costs away from patient care.

This paper contributes to three strands of the literature. First, we provide new evidence

on the eﬀects of PE ownership on target firm operations (Boucly et al., 2011), product quality

(Lerner et al., 2011; Fracassi et al., 2020), and value (Gupta and Van Nieuwerburgh, 2019;

Bernstein et al., 2019; Biesinger et al., 2020). We overcome most limitations of previous

studies on PE in healthcare, such as limited geographies, a short sample period, a lack of

patient-level data, or a small number of deals (Stevenson and Grabowski, 2008; Harrington

et al., 2012; Pradhan et al., 2013, 2014; Cadigan et al., 2015; Huang and Bowblis, 2019;

Gondi and Song, 2019; Casalino, 2020; Gandhi et al., 2020).

In contrast with much of the existing literature, and likely reflecting the considerable

market frictions in healthcare, our results suggest that PE owners may breach implicit

contracts with stakeholders to maximize profits (Shleifer and Summers, 1988). Eaton et al.

(2019) come to the same conclusion in their study of PE ownership of colleges. It is

noteworthy that nursing homes operate under much more intense regulatory scrutiny of

their daily operations than do colleges. Hence, our results raise concerns about the

eﬀectiveness of the elaborate state and federal oversight infrastructure in place to ensure

nursing home quality.

Second, this paper adds to the literature on how provider ownership interacts with price

incentives and regulation in healthcare Duggan (2000); Grabowski and Hirth (2003);

Grabowski et al. (2013); Clemens and Gottlieb (2014); Adelino et al. (2015); Hill et al.

(2019); Curto et al. (2019).4 Some work points to non-pecuniary objectives of nonprofits as

one reason nonprofit providers can outperform for-profits. Our results appear consistent

with this theme, potentially raising questions about whether antitrust regulators should

prospectively review PE deals in healthcare. While the large deals in our sample did not

soften competition, they may have hurt consumers.5

4There is also a related literature on competition in healthcare markets Bloom et al. (2015a), Curto et al.

(2021); Grabowski and Hirth (2003); Dafny et al. (2012); Cooper et al. (2018); Pelech (2018); Ho and Lee

(2019).

5The largest deals in our sample involved purchases of nursing home chains owning hundreds

of facilities already and which remained stable in size. Both the US House and Senate are

considering expanding the scope of the prevailing anti-trust laws. As example, see https:

//www.cnbc.com/2021/02/04/klobuchar-unveils-sweeping-antitrust-bill-laying-out-

5Third, this paper contributes to the emerging literature on the industrial organization

of the nursing home sector, which has received less attention than hospitals in economics

(Lin, 2015; Hackmann and Pohl, 2018; Hackmann, 2019). Previous work has focused on

the role of competition and payment rates in determining quality. Our results imply that

owner incentives are of first-order importance, which may be helpful for policymakers as

they consider regulatory actions to improve transparency and accountability. For example,

in light of prior work showing how PE increases performance when incentives between

investors and consumers are well-aligned, government reimbursements targeting outcomes

could potentially improve patient welfare. These issues have become more urgent as the

COVID-19 pandemic has exposed flaws in the regulation and financing of long-term care

facilities, which have accounted for nearly 40% of U.S. deaths from the virus.6

The paper proceeds as follows. Section 2 provides institutional background. Section 3

describes the data. The strategy for patient-level analysis is explained in Section 4, and the

results are in Section 5. The facility-level estimation is in Section 6. Section 7 concludes.

2 Institutional Background

2.1 The Economics of Nursing Homes

Nursing homes provide both short-term rehabilitative stays – usually following a hospital

procedure – as well as long-term custodial stays for patients unable to live independently.

There are two unique features of the long-term care market in the U.S. relative to other

healthcare subsectors. First, government payers (Medicaid and Medicare) account for 75%

of revenue, while private insurance plays a much larger role in other subsectors (Johnson,

2016).7 Second, about 70% of nursing homes are for-profit, which is a much larger share

than other subsectors. For example, fewer than one-third of hospitals are for-profit.

her-vision-as-new-subcommittee-chair-.html.

6Source: The New York Times Coronavirus Tracker, as of October 2020.

7Medicaid is a means-tested insurance program targeted at low income and disabled non-elderly

individuals. It is the primary payer for custodial care and accounts for about 60% of nursing home patient-

days in our data. Medicare is an entitlement health insurance program for individuals older than age 65, and it

covers limited short-term rehab care following hospital inpatient care, accounting for 15% of patient-days.

6Policymakers have long been concerned about low-quality care at nursing homes in the

U.S. and for-profit ownership has often been proposed as a causal factor (Institute of

Medicine, 1986; Grabowski et al., 2013).8

As with any business, the economics of nursing homes are shaped by the nature of

demand, the cost structure, and the regulatory environment. On the demand side, nursing

homes serve elderly patients but are paid by third-party, largely government payers. Over

95% of facilities treat both Medicare and Medicaid patients (Harrington et al., 2018). Both

programs pay a prospectively set amount per day of care for each covered patient (‘per

diem’), which does not incorporate quality of care, reputation, or other determinants that

would be considered by a well-functioning market. These rates are non-negotiable, and

facilities simply choose whether they will accept beneficiaries of these programs. Medicare

fee-for-service pays much more, at roughly $515 per patient day relative to $209 per patient

day from Medicaid.9 Overall profit margins are in the low single digits (MedPAC, 2017),

a topic we return to at the end of the paper. Payments are adjusted for patient complexity,

so there is an incentive to overstate their severity—a practice known as ‘up-coding.’ This

makes it diﬃcult to use risk as an outcome.

Nursing homes provide institutional care and so have high fixed costs, making the

occupancy rate an important driver of profitability. Nursing staﬀ represent the largest

component of operating cost, at about 50% (Dummit, 2002). Broadly speaking, there are

three types of nurses. Low-skill Certified Nurse Assistants (CNAs) account for about 60%

of total staﬀ hours and provide most of the direct patient care. Licensed Practical Nurses

(LPNs) have more training and experience than CNAs but cannot manage patients

independently. Registered Nurses (RNs) have the highest skill and experience levels, and

can independently determine care plans for patients. LPNs and RNs each account for about

8This concern is frequently reflected in the popular media, including as a reason for high death rates

from Covid-19 in nursing homes. For example, a New York Times article in December, 2020 asserted that:

“Long-term care continues to be understaﬀed, poorly regulated and vulnerable to predation by for-profit

conglomerates and private-equity firms. The nursing aides who provide the bulk of bedside assistance still

earn poverty wages, and lockdown policies have forced patients into dangerous solitude" (Kim, 2020).

9https://skillednursingnews.com/2019/03/medicare-advantage-eats-into-margin-

gains-for-skilled-nursing-facilities/. Medicaid still pays more than the marginal cost of

treatment per day. Hackmann (2019) calculates that the marginal cost of treatment per-day is about $160 on

average.

720% of nurse hours. Nurse availability is crucial to the quality of care and there is a

consensus that low ratios of nursing staﬀ to residents are associated with higher patient

mortality and other adverse clinical outcomes (Tong, 2011; Lin, 2014). Staﬃng ratios are

therefore standard metrics to examine nursing home quality (Hackmann, 2019).

Economists have long emphasized the importance of information asymmetry between

patients and healthcare providers (McGuire, 2000). It is diﬃcult for patients to assess

nursing home quality and compare it to available alternatives, and as discussed above, price

is not a helpful quality signal in this setting. Reputation is therefore likely to have an

unusually large influence on demand for nursing homes (Arrow, 1963). Profit maximizing

facilities should optimally invest in building and sustaining their reputation. This creates a

dynamic incentive problem where they could generate higher profits in the short-term by

cutting patient care costs (nursing staﬀ, for example), but they may optimally sacrifice

these short-term profits in order to maintain their reputation and safeguard patient demand

for the long-term. It remains unclear which factor inputs aﬀect nursing home reputation,

but evidence from prior studies suggests that patient demand does not respond to poor

quality scores on government mandated report cards (Grabowski and Town, 2011; Werner

et al., 2012).

2.2 The Economics of Private Equity Control

PE firms conduct leveraged buyouts (LBOs), in which a target firm is acquired primarily

with debt financing – which is placed on the target firm’s balance sheet – and a small

portion of equity.10 PE is associated with particularly high-powered incentives to maximize

profits because fund managers are compensated through a call option-like share of the

profits, employ large amounts of leverage collateralized with target firm assets, aim to

liquidate investments within a short time frame, and do not have existing relationships with

target firm stakeholders (Kaplan and Stromberg, 2009). Specifically, the compensation of

the General Partners (GPs) who manage PE funds stems primarily from increasing

10Kaplan and Stromberg (2009) provide a detailed discussion of the PE business model and review the

academic evidence on their eﬀects. In the interest of brevity, we limit our discussion.

8portfolio company value between the time of the buyout and an exit, when the company is

sold to another firm or taken public. GPs typically receive 20% of profits after a hurdle

rate, which the fund’s investors (Limited Partners) are guaranteed before GPs receive any

profits. GPs also receive transaction and monitoring fees, which are not tied to

performance. Overall, however, PE managers typically do not earn returns if the business

continues as-is, motivating aggressive value-creation strategies. In contrast, a traditional

business owner running the firm as a long-term going concern with a lower debt burden

may prefer lower but more stable profits.

A large literature in finance beginning with Kaplan (1989) and Kaplan and Schoar

(2005) has shown that in part due to the call option-like nature of GP compensation, PE

buyouts increase productivity, operational eﬃciency, and generate higher returns. Kaplan

and Stromberg (2009) argue that PE owners increase firm value through three channels,

which they call financial, governance, and operations engineering. The first channel

includes the alleviation of credit constraints (Boucly et al., 2011), which may lead to

increases in investment and improved operations, and exploiting the favorable tax code

treatment of debt financing (Spaenjers and Steiner, 2020), which may increase financial

stress and the chances of default.

Governance engineering includes a number of changes to the compensation, benefits,

and composition of the management team at the target firm to align their incentives with

those of the PE owners. For example, in addition to frequently changing key managers, PE

owners also increase the equity stake of the management team and introduce performance-

based compensation (Gompers et al., 2016). Bloom et al. (2015b) show that PE-owned firms

are better managed than similar firms that are not PE-owned.

Operations engineering refers to the more recent practice of PE firms applying their

industry expertise to add value to their investments. PE owners identify both strategic and

operational opportunities, such as re-branding, organization restructuring, investing in new

technology, expanding to new markets, and cost-cutting (Gadiesh and MacArthur, 2008;

Acharya et al., 2013; Bernstein and Sheen, 2016). Davis et al. (2014) show that PE buyouts

are linked to greater labor churn, the expansion of eﬃcient operations and the closure of

9ineﬃcient operations.

Considering these changes in the context of nursing homes, the eﬀects on patients are

theoretically ambiguous. On the one hand, better management, stronger incentives, and

access to credit may lead to improvements in care quality. On the other hand, three forces

could adversely aﬀect quality. The first is that cost cutting measures and a focus on capturing

subsidies could come at the expense of quality improvement eﬀorts. The second is that

the nursing home incurs a large debt obligation as part of the buyout, and the resulting

interest payments can reduce the cash available for care. A related additional cost is leasing

property that formerly was owned by the nursing home. PE owners often sell real estate

assets shortly after the buyout, which generates cash that can be returned to investors. Such

cash flows early in the deal’s lifecycle boost ultimate discounted returns. For example, in

one of the largest PE deals in our sample, the Carlyle Group bought HCR Manorcare for

about $6.3 billion in 2007, of which roughly one quarter was equity and three-quarters were

debt. Four years later, Carlyle sold the real estate assets for $6.1 billion, oﬀering investors

a substantial return on equity (Keating and Whoriskey, 2018). Afterward, HCR Manorcare

rented its facilities; the monthly lease payments are essentially another debt obligation, and

a Washington Post investigation found that quality of care deteriorated following the real

estate sale (Keating and Whoriskey, 2018).

Finally, the third force is the relatively short-term time horizon of PE investments,

which could push managers to focus on maximizing short-term profits even if they come at

the expense of long term reputation and performance. In the case of HCR Manorcare, the

nursing home chain was ultimately unable to make its interest and lease payments and

entered bankruptcy proceedings in the spring of 2018. Carlyle sold its stake to the landlord.

3 Data and Descriptive Statistics

In this section we briefly summarize our data sources and provide descriptives about the

sample, including an analysis of PE targeting. In Appendix A, we describe these elements

in comprehensive detail.

103.1 Data

We obtained facility-level annual data between 2000 and 2017 from publicly-available

CMS sources. In each year we observe about 15,000 unique skilled nursing facilities (we

use the term “nursing home" interchangeably), for a total of approximately 280,000

observations. These data include variables such as patient volume, nurse availability, and

various components of the Five Star ratings. The ratings first appear in 2009. Fortunately,

half of the PE deals in our sample occurred after 2009.

Our second data source consists of patient-level data for Medicare beneficiaries from

2004 to 2016. We use the Medicare enrollment and claims files (hospital inpatient,

outpatient, and nursing homes) for the universe of fee-for-service Medicare beneficiaries.

We merge these files with detailed patient assessments recorded in the Minimum Data Set

(MDS) to obtain additional clinical insights. These data are confidential and were accessed

under a data use agreement with CMS. They include patient enrollment details,

demographics, mortality, and information about care in nursing homes and hospitals during

this period.

In patient-level analysis, the unit of observation is a nursing home stay for a Medicare

beneficiary that begins during our sample period, which we begin in 2005 in order to have

at least one look-back year. We consider only the patient’s first nursing home stay in our

entire sample period so that we can unambiguously attribute outcomes to one facility and

make our patient sample more homogeneous. This produces a sample of more than seven

million patients over 12 years. We are most interested in the eﬀect on mortality, which is an

unambiguously bad outcome, has little measurement error, and is diﬃcult to “game” on the

part of a facility or a government agency. For these and other reasons, short-term mortality

(with suitable risk adjustment) has become the gold-standard measure of provider quality

in the health economics and policy literature (McClellan and Staiger, 1999). We use an

indicator for death during the stay or within 90 days following discharge, based on death

dates recorded in the Medicare master beneficiary summary file. There is a high level of

short-term mortality—one in six patients die within three months of discharge—indicating

11the general morbidity of this patient cohort.

We use two measures of spending: the amount billed to Medicare for the patient stay,

and the amount for the stay plus the following 90 days, in case better quality care leads to

lower subsequent spending (both in 2016 dollars). Medicare covers the entire cost until the

21st day of stay, at which point the patient begins paying a coinsurance. Consequently, about

90% of total payments in our data are made by Medicare and patients bear the remainder.

We complement the mortality analysis by examining some ancillary measures of patient

well-being using the clinical assessments. We study four outcomes that CMS uses when

computing the Five Star quality ratings for nursing homes. The first is an indicator for the

patient starting antipsychotic medication during the stay. Antipsychotic use in the elderly

is known to increase mortality, and non-pharmaceutical interventions such as music and

breathing exercises have been shown to be more eﬀective (Taragano et al., 1997; Kuehn,

2005; Sink et al., 2005; Schneider et al., 2005; Banerjee et al., 2011; Press and Alexander,

2013). The second is an indicator for the patient’s self-reported mobility score declining

during the stay. The third is an indicator for developing a pressure ulcer. The fourth is an

indicator for the patient’s self-reported pain intensity score increasing during the stay.

To identify nursing homes acquired in PE deals, we make use of a proprietary list of

transactions in the “elder and disabled care” sector compiled by Pitchbook Inc., a leading

market intelligence firm in this space. The deals span 2004 to 2015, so that we will have

suﬃcient time to evaluate outcomes. We match the target names to individual nursing

facilities using name (facility or corporate owner) and address as recorded in the CMS data.

This process yields 128 deals, which correspond to a change in ownership to PE for 1,674

facilities. The vast majority of deals in Pitchbook are not at hazard of matching, as they

concern assisted living or other elder care companies that are not Medicare-accepting

skilled nursing facilities. (See Appendix A for details.)

Figure B.1 shows the number of deals in each year; the deals are spread over time, and

no part of the business cycle dominates. The deals are also spread across PE firms. In

total, our data contain 136 unique PE firms that acquired nursing homes. Most deals are

syndicated and involve multiple PE firms. Table B.1 presents the top 10 deals by number

12of facilities acquired. On average, we observe PE-owned facilities for eight years post-

acquisition. Hence the results should be interpreted as medium to long-term eﬀects of PE

ownership. It is diﬃcult to ascertain whether we comprehensively capture PE activity in

this sector. While there is no ‘oﬃcial’ tally of PE-owned nursing homes to benchmark

against, our sample size compares favorably against an estimate of 1,876 nursing homes

reportedly acquired by PE firms over a similar duration, 1998–2008 (GAO, 2010). The PE

investors in our sample include very large funds, smaller funds, and specialized healthcare

PE investment funds. The funds which account for the greatest number of deals are Onex,

Fillmore Capital Partners, The Hillview Group, The Carlyle Group, Cammeby International,

Heritage Partners, Lydian Capital, Formation Capital, and Oaktree.

3.2 Descriptive Statistics

Overall, PE investment in healthcare has increased dramatically in recent decades, as shown

using Pitchbook data in Panel A of Figure 1. Panel B focuses on the Elder and Disabled

Care sub-sector, which includes the nursing homes that we study as well as assisted living

and other types of care. The shaded areas in the graphs correspond to years after our sample

ends, and indicate that deal activity continued to accelerate beyond 2015. The bottom two

panels describe the skilled nursing facilities in our CMS data that are PE-owned. As of 2015,

PE-owned facilities represented about 9% of all nursing facilities in the data, corresponding

to an annual flow of about 100,000 patients. Note that the large spike in the mid-2000s seen

in all the graphs reflects an economy-wide PE boom during this period, and is not specific

to healthcare. Similarly, the flat lining in Panels C and D starting in 2010 reflects the lull in

deal activity due to the Great Recession. Given the patterns in Panel B, the share of facilities

that are PE-owned is likely substantially higher today.

Table 1 Panel A presents summary statistics on key variables used in the analysis at the

facility-year level, where a facility is a single nursing home. Panel B presents summary

statistics at the unique patient level on the final Medicare patient sample (recall we focus

on a patient’s first stay). PE targets are slightly larger, have fewer staﬀ hours per resident,

13and a lower Overall Five Star rating. At the sector level, ratings and staﬃng have secularly

increased over time. For staﬃng, this reflects more stringent standards from regulators over

time. As the PE deals occurred primarily later in the sample, it is therefore remarkable that

they have lower measures of quality on average. Panel B shows that demographic measures

are similar across the types of facilities, such as patient age and a high-risk indicator.11

PE-owned facilities bill about 10% more per stay than non-PE facilities.

We describe which characteristics are associated with buyouts in Table A.1. Facilities

in more urban counties and in states with higher elderly population shares are more likely

to be targeted.12 County-level income, race, and home ownership do not predict buyouts.

Chain-owned facilities are more likely to be acquired than independent facilities, likely

reflecting the fixed costs of a PE deal. A higher share of Medicare patients (the omitted

group) is positively associated with being targeted. Finally, the Five Star overall rating has

a negative relationship with buyouts, indicating that PE firms target relatively

low-performing nursing homes. These factors remain statistically significant predictors

when included simultaneously in the same model, shown in column 5. These results

highlight the need to estimate the eﬀects of PE ownership within-facility.

4 Empirical Strategy for Patient-Level Analysis

There are two primary concerns related to measuring the causal eﬀects of PE ownership on

patient-level outcomes. First, PE funds may target facilities that are diﬀerent in ways the

econometrician cannot observe. To address this concern, we include facility fixed eﬀects,

eliminating time invariant diﬀerences across facilities. Second, following a PE buyout, the

composition of patients may change, confounding the analysis. Diﬀerential customer

selection following PE ownership could reflect both supply-side channels such as changes

in advertising and hospital referrals, or patient perceptions about PE ownership.

11We use the Charlson Comorbidity Index, a standard measure of patient mortality risk based on co-

morbidities. We create a high-risk indicator that is equal to one if the previous-year Charlson score is greater

than two.

12The map in Figure B.2 shows that deals are not excessively concentrated in particular areas of the country.

14Recent studies have documented that nursing homes can select patients based on patient

characteristics, only some of which are observable to CMS (Hackmann and Pohl, 2018;

Gandhi, 2020). We see evidence of changes in patient risk following PE ownership in our

data. Table B.5 Panel A presents point estimates from diﬀerences-in-diﬀerences models

that exploit variation in the timing of the PE deals across facilities. We test for changes in

patient risk (assessed at the time of admission) following acquisition. We examine eﬀects

on a mix of acute and chronic conditions to broadly capture changes in patient risk. The

coeﬃcients indicate that patients are less likely to suﬀer from Dementia and Alzheimers or

from acute conditions like Hip Fractures at the time of admission. However, they are also

more likely to have a Urinary Tract Infection (UTI). Figure B.3 presents the corresponding

event study plots, which generally suggest flat or declining trends in patient risk around the

time of the acquisition. We are concerned that if there is a similar decline in unobserved

patient risk following PE ownership, it will bias downward mortality and spending eﬀects

obtained via OLS. Therefore, we develop an instrument for the match between patients and

nursing homes.

4.1 Distance Instrument

We use a diﬀerential distance instrument (McClellan et al., 1994) to control for endogenous

patient selection into nursing homes. The instrument exploits the well-known patient

preference for nearby healthcare providers (Einav et al., 2016; Card et al., 2019; Currie and

Slusky, 2020). This is especially true for nursing homes; for example, Hackmann (2019)

finds that the median distance between a senior’s residence and her nursing home is under

4.3 miles. This is also evident in our data—the median and 90th percentile distances

between a patient and her nursing home are 4.6 and 18 miles, respectively. About 35% of

all patients choose the facility located closest to them (see Figure B.4).13 As a result of

these patterns, this instrument has been useful in the nursing home setting to control for

patient selection (Grabowski et al., 2013; Huang and Bowblis, 2019).

13Distance patterns remain remarkably stable over time in our sample. Mean distance to facility is unaﬀected

by PE buyout, as shown in Figure B.4D.

15We compute the diﬀerence (in miles) between two distances: from a patient’s home

zip code to the closest PE-owned facility zip code; and from the patient’s residence to the

nearest non-PE facility zip code. A positive value indicates the patient is closer to a non-PE

facility. A lower diﬀerential distance value implies the nearest PE-owned facility is closer

to the patient. PE ownership evolves over time as more deals take place (and some PE funds

exit their investments), creating variation across years in diﬀerential distance for individuals

residing in the same zip code. Following convention in the literature, we drop patients with

outlier diﬀerential distance values.14

The first stage is estimated using Equation (1), and the second stage is estimated using

Equation (2). The endogenous regressor of interest PEi,j,r,t is an indicator set to one if patient

i in Hospital Referral Region (HRR) r chooses PE-owned facility j in year t. We instrument

with linear and squared diﬀerential distance, Di applicable to patient i based on her zip code,

z, and when the nursing home stay began.

PEi,j,r,t = αj + αr,t + ζ1Di + ζ2D2

i + X′

i,z ξ+ νi,j,r,t (1)

Yi,j,r,t = αj + αr,t + φPEi,j,r,t + X′

i,z γ+ εi,j,r,t (2)

Our preferred model controls for facility, αj, and patient HRR by year fixed eﬀects, αr,t. The

vector Xi,z denotes patient risk controls including age, indicators for gender, marital status,

dual eligible, and 17 disease categories.15 We conduct multiple robustness checks, which

include adding time-varying socioeconomic variables at the patient’s zipcode-year level and

omitting all controls.16 Standard errors are clustered by facility to account for unobserved

correlation in outcomes across patients treated at the same nursing home.

The instrument is strongly predictive of choice of nursing home type. The first stage

results are reported in Table 2. Column 2 presents the estimates from our preferred

14Specifically, we drop patients with a diﬀerential distance value beyond 70 miles, which is approximately

the 95th percentile (i.e., the nearest PE facility is 70 miles further than the nearest non-PE facility). The

concern here is that these patients are plausibly located in a diﬀerent market which PE facilities do not operate

in, and hence could diﬀer in unobserved ways correlated with health or spending outcomes. To be symmetric,

we also drop (the very few) patients who have a diﬀerential distance value below -70.

15To construct these indicators, we use diagnoses codes recorded in claims billed over the three months prior

to the index nursing home stay (hospital stays, ED visits, and outpatient visits).

16The socioeconomic variables, from the American Community Survey, are annual median household

income, the share of the population that are white, that are renters rather than home-owners, and that are

below the Federal poverty line.

16specification. A five mile decrease in diﬀerential distance (0.3 s.d.) increases the

probability of going to a PE-owned nursing home by 2.3 percentage points (pp), a quarter

of the mean level. The F-statistic exceeds 200, well above conventional rule-of-thumb

thresholds for weak instruments.

4.2 Instrument Assumptions and Validation Tests

IV estimation diﬀers from randomized controlled trials because the randomization of

patients to treatment is indirect rather than deliberate. As in all such analyses, we must rely

on two untestable identification assumptions. The first is conditional random assignment,

which requires that after conditioning on covariates, unobserved characteristics correlated

with the outcomes of interest are not correlated with diﬀerential distance. This assumption

subsumes the exclusion restriction, that the instrument aﬀects outcomes only through its

eﬀect on the patient’s probability of going to a PE facility. The second assumption is

monotonicity, which assumes that a decrease in diﬀerential distance makes all patients

more likely to choose a PE-owned facility. This is true on average, but the assumption is at

the patient-level which is untestable. Monotonicity is necessary to interpret the IV estimate

as a well-defined local average treatment eﬀect (LATE).

An important test for randomization examines whether diﬀerential distance is

correlated with covariates, particularly those which may aﬀect health outcomes, such as

risk. Comparing the estimates reported in columns 1 and 2 of Table 2, the coeﬃcients on

diﬀerential distance are nearly unaﬀected by including patient-level controls, consistent

with random assignment. Figure 2 Panel A visually presents the relationship between

patient risk and the instrument and indicates little or no correlation.17

Additional evidence for random assignment is that patient characteristics are similar for

high and low values of diﬀerential distance. We document this in Table 3, where we

summarize 21 patient characteristics for above- and below-median diﬀerential distance

17We project the high-risk indicator (see Section A.2) on the controls we use in our main regression, and

collapse the residuals into twenty bins. Similarly, we run a regression of diﬀerential distance on the controls

and collapse the residuals into twenty bins. We plot the means of each bin, with the risk residuals on the Y-axis

and distance residuals on the X-axis. The figure also presents a fitted line and the slope coeﬃcient.

17values. The top two rows of the table show that, consistent with a strong instrument,

below-median diﬀerential distance average is 2.7 miles, while the above-median average is

27 miles. The associated probability of going to a PE-owned facility declines from 17% to

2%. The patient characteristics in the subsequent rows are extremely similar across the two

groups. For example, 64% of each group are women, and about a quarter of the patients in

both groups have diabetes. While diﬀerential distance is highly predictive of going to a

PE-owned facility, it appears to randomize patients with respect to observed covariates.

PE funds may strategically target nursing homes located in places with certain desired

demographic and risk profiles. We account for stable diﬀerences in the patient catchment

of facilities by including facility fixed eﬀects. However, it is possible that PE firms

strategically target geographic markets with desirable trends, for example with increasing

household income. To address this concern, we show robustness to including time-varying

zip code-level socioeconomic controls. We document that these controls do not aﬀect the

first stage in Column 3 of Table 2. The use of HRR-specific year fixed eﬀects further

mitigates the possibility of diﬀerential market trends biasing the eﬀects.

A related concern may be that HRRs are too large and do not suﬃciently control for

unobserved heterogeneity in trends across markets. Hence, we also test robustness to using

the more granular market definition of Health Service Areas (HSA) and counties.18 There

are nearly 800 HSAs and 3,000 counties, respectively, while there are only about 300 HRRs.

Columns 4 and 5 of Table 2 present results using these finer market definitions, respectively,

with slightly smaller estimates.

We provide evidence consistent with the monotonicity assumption in Figure 2 Panel B,

which contains a binscatter plot of the first stage, showing that the likelihood of going to a

PE-owned facility increases nearly linearly with diﬀerential distance. It is estimated in the

same way as Panel A described above, except that the outcome is an indicator for the facility

being PE-owned. The monotonicity assumption also implies that the first stage should be

18HSAs were developed by the National Center for Health Statistics of the Centers for Disease Control

in the mid 1990s. They are designed to identify a single county or contiguous sets of counties where

Medicare patients seek hospital care within the area. We use a slightly modified version developed by the

SEER program of the National Cancer Institute, available for download at https://seer.cancer.gov/

seerstat/variables/countyattribs/hsa.html.

18negative when estimated on sub-samples of patients with diﬀerent characteristics. Table B.2

shows that when we estimate the relationship between below-median diﬀerential distance

and PE ownership (a simplified first stage), we recover coeﬃcients that are very similar to

the full-sample result and all are significant at the .01 level for a variety of sample splits by

age, gender, race, and zipcode income level.

Table B.2 also helps characterize compliers relative to the average patient at a PE

facility. The ratio of the first stage coeﬃcient for a subsample with a specific attribute to

that obtained for the full sample provides the likelihood of compliers having that particular

attribute relative to the average PE patient.19 Compliers appear to have a very similar age

distribution and the probability of being male, married, or white. Intuitively, distance-based

compliers are more likely to be from a low-income zipcode.

5 Patient-Level Eﬀects

This section presents the main results of the paper. We focus on the eﬀects of PE ownership

on short-term mortality and spending per patient, discussing the LATE, heterogeneity in

treatment eﬀects, as well as tests for the mechanism and robustness.

5.1 Main Eﬀects on Mortality and Spending

Table 4 presents the results obtained by estimating Equation (2). These models include

22 patient-level controls (described in Section 4.1), facility fixed eﬀects, and patient HRR-

by-year fixed eﬀects. Column 1 indicates that receiving care at a PE-owned nursing home

increases the probability of death during the stay and the following 90 days by 1.7 pp, about

10% of the mean. In the context of the health economics literature, this is a very large eﬀect.

This estimate remains stable in magnitude at about 10% of the mean regardless of the time

horizon studied (see Table B.3).

19This follows from Bayes rule and the use of a discrete instrument in this model of the first stage. The

coeﬃcient from a subsample with attribute X is P(M|X)= P(X|M)P(M)/P(X) where M denotes a marginal

PE patient. Dividing by the first stage coeﬃcient for the full sample, P(M), gives us P(X|M)/P(X), the relative

likelihood.

19We calculate the implied cost in statistical value of life-years in Table B.4 Panel A.

We translate the IV coeﬃcients into lives and life-years lost based on the number of index

stays by patients of PE-owned nursing homes during our sample period. Accordingly, we

compute about 20,150 additional deaths due to PE ownership over our twelve-year sample

period. To estimate life-years lost, we rely on observed survival rates for Medicare patients

at all nursing homes. This leads to an estimate of about 160,000 lost life-years.20 Applying

a standard estimate of statistical value of a life-year of $100,000 (Cutler and McClellan,

2001), inflated to 2016 dollars, this implies a mortality cost of $20.7 billion.

The next two columns of Table 4 Panel A consider spending per patient. In our data,

more than 90% of the billed amount is paid by taxpayers through Medicare and patients pay

the balance. The amount billed per nursing home stay increases by 19.5% (column 2; note

it is necessary to exponentiate coeﬃcients larger than .1 when the outcome is logged). As

Table 1 shows, on average PE-owned nursing homes bill $14,800 per stay, while non-PE

nursing homes bill $13,500. This does not seem to reflect additional preventive care that is

compensated for by lower subsequent needs, because the total amount billed for both the

stay and the 90 days following the stay (the episode) increases by about 11%.

The most important robustness test we conduct is a placebo analysis, which probes

whether spurious trends rather than the ownership change might explain the results. We use

Medicare patient-level data from 2002–07, a period with little PE ownership of nursing

homes and little overlap with our main sample. We randomly set the PE dummy to turn on

in 2004 or 2005 for facilities that eventually were acquired by PE firms later. Further, we

discard data for any facility starting with the year it actually got acquired. We recompute

diﬀerential distances under these ‘placebo’ assignments and estimate our main IV models.

Table 4 Panel B presents these placebo estimates and reassuringly finds small and

20As life expectancy diﬀers substantially between men and women, we estimate the eﬀect separately by

gender. We calculate the average life expectancy at discharge by gender by observing the actual life span for

each patient in our data. For patients still alive at the end of our sample period, we approximate the year

of death based on patient gender and age using Social Security actuarial tables. We adjust this downward to

account for the fact that decedents tend to be older on average (by about 2 years). We then applied this mean

life expectancy to the number of deaths computed above and obtained the number of life-years lost. This

approach may overstate the true value if the incremental deaths at PE facilities are of older patients. This

approach also understates the true value since we don’t account for the loss in longevity not resulting in death.

20insignificant eﬀects, implying a lack of diﬀerential trends prior to acquisition.

Our IV estimates imply that the reduced form eﬀect on mortality and spending should

decline as diﬀerential distance grows larger (i.e., relative to the nearest alternative, a PE

facility is farther away). Figure 3 visually confirms this pattern. The figure plots

coeﬃcients from regressing each outcome on indicators for quintile of diﬀerential distance,

with the furthest quintile as the reference group using our preferred controls as in the main

specification. By using quantile dummies, this specification is flexible and does not impose

linearity with respect to diﬀerential distance. We find the largest eﬀects among patients in

the bottom two quintiles of diﬀerential distance, i.e., those located nearest to PE-owned

facilities.

Results from OLS models are presented in Table B.5 and the corresponding event studies

are in Figure B.5. They suggest no pre-trends, consistent with the parallel trends assumption

that underlies our empirical model (i.e., target facilities and control facilities would continue

on parallel trends in the absence of the buyout). We observe a statistically significant, but

much smaller increase in mortality in the OLS model (0.3 pp). This is only one-sixth the

size of the IV estimate, consistent with unobservedly lower risk patients matching with PE-

owned facilities. In a similar vein, we also find small, negative eﬀects on spending (1-2%

decrease) and length of stay (not presented).

5.2 Heterogeneity in the Mortality Eﬀect

This section explores heterogeneity both on observed attributes and on unobserved

resistance to treatment, using a Marginal Treatment Eﬀects (MTE) framework.

5.2.1 Observed attributes

To assess heterogeneity in the IV analysis, we split the sample based on observed

characteristics. We first consider four groups based on patient risk and age. We expect that

higher age is associated with a greater need for attentive but not necessarily high-skill or

complex care, for example helping patients to use the toilet and minimizing infection risk.

21Higher risk – a measure constructed from disease burdens – should be associated with

more need for high-skill, medicalized RN care. Older, high risk patients require the most

intensive and high-skill care. Therefore, we split the sample into four groups around the

median age of 80 and around the high-risk indicator (Charlson score above two). The

results, shown in Table 5 Panel A, document that the eﬀect on mortality is driven by

patients who are low risk, with the most robust result among patients who are low risk but

above-median age. This group accounts for nearly half of the sample. The high risk,

above-median age group also has a large, positive coeﬃcient, but it is noisy. In contrast, the

point estimate for high-risk but below-median age patients is negative and marginally

significant. This suggests that PE-owned nursing homes are able to take better care of more

complex patients, especially when they are on the younger side. But lower risk or older

patients suﬀer.

We find positive eﬀects among both men and women, but the eﬀect is larger and much

more robust among female patients, who represent 65% of the sample and are on average

older. The eﬀect is also larger among patients from above-median income zip codes.21 It

is also larger among White patients. Finally, the last set of results divide the sample into

three categories corresponding to the patient’s reason for hospitalization prior to the nursing

home stay; we find the largest eﬀect for patients who were hospitalized due to cardiovascular

disease.

There is evidence that for-profit incentives generally and PE ownership specifically are

associated with lower quality of care in more concentrated markets (Gandhi et al., 2020), so

we examine in Panel B whether the eﬀects vary by market competition, using the Herfindahl-

Hirschman Index (HHI) of the hospital referral region (HRR). We find that the coeﬃcient is

larger among nursing homes in below-median HHI areas, but the coeﬃcient is more precise

among nursing homes in above-median HHI areas. As both coeﬃcients are relatively close

to our main estimate, concentration does not appear to be a driving factor.

21We do not observe beneficiary income directly, so we assign individuals to above-median and below-

median income neighborhoods based on their zip code.

225.2.2 MTE Theory and Estimation Approach

The LATE may mask treatment eﬀect heterogeneity across diﬀerent types of patients. For

example, some patients may benefit from the type of care that is oﬀered by PE-owned

facilities, even though we estimated negative impacts on average for the complier group. It

also ignores the possibility of patient selection on treatment gains. The MTE framework

allows us to examine these dimensions (Heckman and Vytlacil, 2005; Heckman et al.,

2006). It enables us to compute treatment eﬀect parameters of economic interest such as

the Average Treatment Eﬀect (ATE) and Treatment on the Treated (ATT). Unlike the

LATE, these parameters are not specific to the complier group and allow us to make more

general statements regarding the causal eﬀects of PE ownership.

We denote Y0,i and Y1,i as potential outcomes for individual i in the untreated (k= 0) and

treated (k= 1) states, respectively. Treatment in our setting is receiving care at a PE-owned

facility, PEi. We model these potential outcomes Yk,i as a function of observed control

vector Xi and dummies for facility, F j and market-year interactions, Rr,t. Uk,i denotes all

unobserved factors.22

Yk,i = X′

i βk + F j + Rr,t + Uk,i, k= 0,1 (3)

We then propose a latent selection model of how patients choose a PE-owned facility based

on observed and unobserved factors.

PEi = Z′

i δ−Vi,

(4)

PEi = 1 i f PE≥

i 0, PEi = 0 otherwise,

where Z= (X,F,R,D,D2) is a vector including all the controls listed above in Equation (3)

and the diﬀerential distance instruments excluded from the outcome equation, Di and D2

We interpret Vi as the unobserved resistance to going to a PE-owned facility. This selection

model imposes monotonicity by using a constant parameter δfor all individuals. Following

the MTE literature, we transform the selection equation into the quantiles of the distribution

of V rather than its absolute values:

22Following Brinch et al. (2017), we assume that the error term Uk,i is normalized to be conditional mean

zero, i.e., E[U|X= x,F= f,R= r]= 0.

23Z′

i δ−Vi ≥0=⇒ Z′

i δ≥Vi =⇒ Φ(Z′

i δ) ≥Φ(Vi), (5)

where Φ is the cumulative distribution function of Vi. We interpret Φ(Z′

i δ) as the propensity

score, the probability that an individual with observed characteristics Zi chooses a PE

nursing home, and denote it as P(Z). Φ(Vi) represents the quantiles of unobserved

resistance to treatment, and is denoted as UD.

Omitting subscripts for simplicity, the MTE is defined as MT E(X= x,UD = u)= E[Y1−

Y0|X= x,UD = u]. The MTE is the treatment gain for an individual with characteristics X=

x, who is in the uth quantile of the resistance distribution. Such individuals are indiﬀerent to

receiving treatment when their propensity score P(Z) equals u.

We make two untestable assumptions to estimate the MTE. The first, as in Section 4.2,

is random assignment of the instrument, conditional on observables. The second

assumption is of functional form. Following the convention in the recent MTE literature

(Brinch et al., 2017; Cornelissen et al., 2018), we assume that the MTE is additively

separable into an observed and unobserved component. This allows the MTE to be

identified over the unconditional support of P(Z) across all values of X rather than the

support of P(Z) conditional on X= x, easing the burden of identifying variation needed

from the data (Carneiro et al., 2011).

MT E(X= x,UD = u)= E[Y1−Y0|X= x,UD = u]

(6)

= x(β1−β0)

observed

+ E[U1−U0|UD = u]

unobserved

Another implication of this assumption is that treatment eﬀect heterogeneity due to X

aﬀects the MTE curve in u only through the intercept. The slope of the MTE curve in u does

not depend on X, facilitating estimation. The potential outcomes model described above

produces the following outcome equation as a function of observables and P(Z) (Carneiro

et al., 2011).

E[Y|X,F,R,P(Z)= p]= X′β0 + F + R + X′(β1−β0)p + K(p), (7)

where K(p) is a nonlinear function of the propensity score. The derivative of this outcome

equation with respect to p estimates the marginal treatment eﬀect at X= x and UD = p

24(Heckman et al., 2006).

We first estimate the selection model in Equation (4) using a linear probability model and

obtain ˆ p = Z′

δ. Figure 4 Panel A presents the variation in the estimated propensity score.

We collapse the data to percentiles of diﬀerential distance, D and plot a non-parametric

fit of P(Z) values against the corresponding percentile means of D. This shows a similar

pattern first observed in Figure 2—the probability of going to a PE-owned facility declines

nearly monotonically as diﬀerential distance increases. However, this figure masks the full

support of the distribution of P(Z), which extends over the entire unit interval. Figure 4 Panel

B highlights the overlap in distribution of the propensity scores for treated and untreated

patients by plotting histograms for the two groups against P(Z) on the X-axis. We use

log scales on the Y-axes since there are large numbers of observations at the two extremes

of propensity score. The figure confirms that the treated and untreated groups overlap in

distributions over nearly the entire unit interval, enabling the estimation of the unconditional

ATE without the need for extrapolation (Basu et al., 2007). We then estimate the outcome

Equation (8) below, assuming K(p) is a polynomial in p of degree S.

Y= X′β0 + F + R + X′(β1−β0) ˆ p +

ρsK( ˆ p) + ϵ. s=2

(8)

The MTE curve is the derivative of Equation (8) with respect to ˆ p. In our baseline model

we set S= 2, but test robustness to using higher order polynomials. Standard errors are

obtained by block bootstrap, clustering by facility.

5.2.3 MTE Results

We estimate Equation (8) and confirm the presence of selection on unobserved resistance by

testing the joint significance of the coeﬃcients ρs on the higher order terms of the polynomial

in p (Heckman et al., 2006). The coeﬃcient on p2 is highly statistically significant (p value

<0.01), confirming patient selection into PE facilities on unobserved resistance.

Figure 4 Panel C presents the MTE curve along with 90% confidence intervals. Our

primary approach uses a second degree polynomial, so the MTE curve is linear in

unobserved resistance (u). Since it is downward sloping, there is reverse selection on

25treatment gains; that is, individuals with the least resistance to going to a PE facility

experience the worst mortality eﬀects of nearly 4 pp. In contrast, individuals with the

highest resistance experience marginally negative (i.e., beneficial) eﬀects. The MTE values

are not statistically significant for individuals with above median resistance to treatment.

The figure also plots the ATE, which is 1.3 pp (s.e. 1.4 pp). To test sensitivity to the

linearity assumption, we also estimate the MTE curve with 3rd, 4th, and 5th degree

polynomials. Figure B.6 Panel B shows that the curve remains downward sloping

regardless of the polynomial.

We aggregate the marginal treatment eﬀects using the appropriate weights to obtain

various treatment eﬀect parameters such as the unconditional ATE and ATT (Cornelissen et

al., 2016). Given the downward sloping nature of the MTE curve, we expect the average

eﬀect on the treated to be higher than that for the untreated. Figure 4 Panel D presents the

weights to apply to the MTE values to compute the ATT and ATUT. Accordingly, we

estimate an ATT of 3.1 pp (s.e. 0.9 pp) and an ATUT of 1.0 pp (0.9 pp). Only the ATT is

statistically significant among the three treatment eﬀect parameters.

There are two key takeaways from this analysis. First, the ATE implies that a randomly

chosen Medicare patient from our sample would experience an increase of 1.3 pp in the

probability of short-term mortality if she chose a PE-owned nursing home. While about a

third lower than the LATE estimate, it nevertheless implies a large number of deaths in a

counterfactual where all Medicare short-stay patients receive care at a PE-owned facility.

Second, the MTE curve implies reverse selection on gains and that some patients – those

with greater resistance to treatment – experience improvements in mortality if they choose

a PE-owned facility, though the negative MTE values are not statistically significant. This

pattern is consistent with the heterogeneity in treatment eﬀects on observed attributes. For

example, we find a large and statistically significant increase in mortality for individuals

residing in zip codes with greater than median income (see Table 5 row 3). Individuals in

richer neighborhoods are also about 20% more likely to choose PE-owned facilities – their

mean propensity score is 12 pp versus 10 pp for patients from neighborhoods with income

below the median. In contrast, we find a smaller and statistically insignificant eﬀect for

26individuals in lower income neighborhoods.

5.3 Patient Well-Being and Mechanism Tests

If the eﬀect on short-term mortality is related to lower patient welfare, we expect to see

consistent evidence using other wellbeing measures. Therefore, we also use the IV model

to assess eﬀects on the four clinical measures of wellbeing that CMS uses as outcomes for

short-stays when computing Five Star ratings (surprisingly, mortality is not one of them).

The first is whether a patient starts taking antipsychotic drugs. As discussed in Section

3, antipsychotics are discouraged in the elderly due to their association with mortality and

the greater eﬃcacy of behavioral interventions. We find that going to a PE-owned nursing

home increases the chances of starting antipsychotics by 3 pp, or 50% of the mean (Table

6 column 1). Using an estimate from the literature on how antipsychotic medications aﬀect

mortality, this coeﬃcient implies that about 15% of the total eﬀect on mortality is potentially

attributable to starting antipsychotics.23

We also find a positive eﬀect on experiencing worsening mobility, which increases by

4.3 pp, or about 8% of the mean (Table 6 column 2). We do not find a significant eﬀect

on the third measure – developing ulcers – though the coeﬃcient is positive (column 3).

Fourth, there is a positive eﬀect on increasing pain intensity of 2.7 pp, which is 10% of the

mean. Figure B.5 presents the corresponding event studies and indicate no diﬀerential pre-

trends.24 Overall, the evidence of harmful eﬀects on other measures of patient wellbeing are

comfortingly consistent with the estimated eﬀects on mortality.

Thus far we have assumed that ownership type explains any eﬀects of PE buyouts.

Alternatively, PE ownership could bring economies of scale or corporatization, which are

23Several clinical studies have examined the harmful eﬀects of antipsychotic prescribing for the elderly.

The most relevant study for our purposes is by Schneider et al. (2005), who perform a meta-analysis of fifteen

randomized controlled trials (11 from nursing homes) that studied the eﬀects of antipsychotics on mortality for

elderly patients. They report a 50% increase in mortality. The trials evaluated mortality at durations averaging

about 3 months, coincidentally matching our mortality measure. Applying a 50% increase in mortality to our

setting implies an 8 pp increase in 90-day mortality on a mean of 17 pp. We apply this elevated mortality

eﬀect to the 3 percent additional patients at PE facilities who receive antipsychotics. This implies an increase

in mortality of 8 x 0.03= 0.24 pp for PE patients on average.

24Results using OLS models are in Table B.5B. They are typically smaller in magnitude (except ulcers),

consistent with selection leading to downward bias in OLS.

27the explanation that Eliason et al. (2020) propose for negative eﬀects of dialysis center

mergers. To test this hypothesis, we conduct three tests in Table 7. The first adds to our

main model a control for being a chain versus an independent facility. If our eﬀects are

explained by the “rolling-up" of independent facilities into more eﬃcient chains, the

estimates should attenuate. Instead, they are essentially unchanged. The second test

excludes the top two deals—the buyouts of the very large Genesis Healthcare and Golden

Living chains (both have more than 300+ facilities). The coeﬃcient is larger, implying that

our result is not driven by the very largest chains. The most important test is in row 4,

where we use only the top five deals to define PE ownership. In these deals, the target

chains already owned more than 100 facilities and stayed nearly the same size over the

sample period. Therefore, in this model chain size is held constant and we evaluate the

eﬀect of a change in ownership. Again, the eﬀect is larger than in the full sample. In sum, it

does not seem that chain corporate structures or synergies in large firms explain our results.

Another concern is whether the results are spuriously capturing the quality diﬀerence

between for-profit and nonprofit nursing facilities. About 20% of the patients receive

treatment at a non-profit facility. By definition these facilities are part of the control group.

We test the sensitivity of our main estimate to excluding these facilities from the sample

altogether. Row 5 presents the corresponding results and shows that the estimate reduces

about 20% in magnitude but remains statistically significant.

The remaining rows of Table 7 report robustness tests that vary the controls and market

definitions. If the instrument does not randomly assign patient risk, we expect patient

controls to substantially aﬀect the results. Instead, the results are robust to alternative

controls, consistent with random assignment. The first test in this group (row 6) includes

zip-year socioeconomic controls. The coeﬃcients decline only slightly. The next two rows

use the more granular HSAs and counties, respectively, to define patient markets instead of

HRRs. The final row omits all patient controls, estimating larger eﬀects, but well within

two standard errors of the main estimate. Overall, the results are quite stable.

286 Operational Changes

This section uses facility-level data to explore operational changes that could help explain

the adverse patient welfare eﬀects described in the previous section.

6.1 Empirical Strategy

For outcomes available only at the nursing home level, we cannot instrument for patient

selection and the best possible research design therefore is diﬀerences-in-diﬀerences. We

use variants of the following specification:

Yj,t = αj + αt + βPE j,t + P′

j,t γ1 + M′

j,t γ2 + εj,t (9)

PE j,t takes a value of one if facility j is PE-owned in year t. The coeﬃcient of interest is

β, which captures the relationship between PE ownership and the outcome Yj,t. We include

facility (αj) and year fixed eﬀects (αt). We retain all facilities in our preferred specification,

but the results are robust to limiting the sample to for-profit facilities. The vector Pj,t includes

three controls for facility-level patient mix and Mj,t includes five county-level controls for

time-varying market attributes.25 As there may be concern that control variables could be

aﬀected by PE ownership, we also present results without any controls.

The identifying assumption is that PE targets and control facilities would continue on

parallel trends in the absence of the acquisition. We assess whether there are diﬀerential

pre-trends using event study figures, which plot the coeﬃcients βs from estimating Equation

(10) below.

Yj,t = αj + αt +

βsDeal Yearj,s + P′

j,t γ1 + M′

j,t γ2 + εj,t (10)

s 0

25Patient mix controls: Case Mix Index (CMI) is a composite measure of patient risk based on medical

history of diagnosis or treatment for a large number of conditions. Second, Acuity index is a measure of

patient risk computed using the patient’s assessed Activities of Daily Living (ADL) scores. In both cases, a

greater value indicates a riskier patient cohort for the nursing home. We winsorize both the CMI and Acuity

Index at the 1% and 99% level in each year. The third control is the share of the facility’s patients who

are Black. County-level controls: Herfindahl Hirschman Index (HHI) based on shares of beds, number of

for-profits, number of chain-owned, number of hospital-based, and number of overall facilities. These are

calculated using a leave-one-out procedure from the facility-level data.

29Deal Yearj,s is an indicator that is one in year s relative to the buyout year for facility j, and

zero otherwise. The remaining terms are as defined above for Equation (9).

6.2 Results

We consider three types of operational channels. The first two explicitly concern facility

quality, while the last pertains to financial strategies particular to the PE industry. All the

results are presented in Table 8. For each outcome, the top row of coeﬃcients are from

specifications with only facility and year fixed eﬀects, while the bottom row adds the full set

of patient and market controls. Event studies are in Figures 5 and 6.

6.2.1 Compliance With Standards and Staﬀ Availability

First, we consider compliance with care protocols in Panel A of Table 8. Our outcome of

interest is the facility-level Five Star rating, which varies from one (worst) to five (best).

After PE buyouts, the Deficiency rating declines by 0.08 points (column 1), which is about

3% of the mean and 7% of the standard deviation (the most relevant measure given how this

variable is constructed). This rating reflects whether the facility is satisfying care protocols

such as storing and labeling drugs properly, disinfecting surfaces, as well as other aspects

of care such as ensuring resident rights and avoiding patient abuse. The Overall rating

similarly declines (column 2). Figure 5 presents event studies for each outcome. There are

no pre-trends, consistent with the identifying assumption, and the negative eﬀects appear

immediately after the change in ownership and persist for at least five years.26

In Panel B, we assess eﬀects on nursing staﬀ hours per patient-day, a well-established

measure of nursing home quality that accounts for changes in patient volume. Column 1

shows a modest decline of 0.05 hours in aggregate staﬀ hours (1.4% of the mean). This

aggregate eﬀect masks larger changes for diﬀerent types of nurses that oﬀset each other.

There is a decrease in ‘front line’ caregivers (CNAs and LPNs), shown in columns 2 and

26The Overall rating has three components: the Deficiency rating, a Quality rating based on metrics

computed using claims data and clinical assessments, and a Staﬃng rating, which is based on staﬃng measures

evaluated in Panel B. Since we assess quality and staﬃng changes more granularly, we do not present the eﬀects

on these components, but we find negative, significant eﬀects of equal or larger magnitudes there as well.

303, respectively. Together there is a decline of around 0.09 hours for these two groups (3%

of the mean). In contrast, there is an increase in use of Registered Nurses (RNs) by about

0.04 hours (8%). The event studies in Figure 5 again reveal no pre-trends and indicate more

immediate declines after the deal in front-line staﬃng, while the increase in RN staﬃng

appears starting in the third year after the buyout.27 The increase in RN staﬀ hours does not

compensate for the decline in lower skilled nurse hours because RNs account for a small

fraction of all staﬀ hours. Medicare cost reports indicate that CNAs and LPNs receive an

hourly wage that is about 40% and 70% respectively of the wage paid to RNs, which is

around $35 per hour. Unfortunately, we cannot observe whether facilities are taking cost

reduction steps such as using more part-time labor and reducing individual shifts.

The existing literature helps to connect the eﬀects on nurse availability with the

estimated eﬀect on mortality. Tong (2011) exploits an increase in minimum nurse staﬃng

regulation in California and finds a decline in on-site patient mortality due to greater

availability of frontline nurses. Applying her estimates in our setting, the estimated decline

in frontline nurse staﬃng predicts an increase in mortality of 0.25 pp.28 The findings on

increased use of antipsychotics and lower nurse availability may be related. Grabowski et

al. (2011) note that antipsychotics are believed to substitute for nurse care and show that

when nursing homes increase wages, inappropriate use of antipsychotics decreases.

Therefore, it is intuitive that lower staﬃng – in particular low-skill staﬃng – would be

associated with increases in adverse conditions related to lack of attention, such as more

use of antipsychotics and lower mobility. The two channels additively predict an increase

in mortality of 0.5 pp in our setting (about 30% of our mortality eﬀect). However, this may

be an underestimate if they produce larger eﬀects when they interact.

The increase in RN availability is consistent with the negative eﬀects on mortality being

driven by older rather than more complex patients. RN staﬀ are most relevant for the more

27We report the results of robustness tests in Table B.6, which include controls for chains, excluding the top

two deals, and including only for-profit nursing homes.

28Tong (2011) reports a 15% decline in mortality due to an increase in nurse availability of one hour per

resident-day. Since we estimate a decline of 0.09 hours, this predicts an increase of 0.09 x 15= 1.4% of the

mean, or 0.24 pp. More recently, Ruﬃni (2020) exploits variation in minimum wage requirements to isolate

the eﬀects of nurse staﬃng changes on quality and also finds mortality eﬀects.

31medicalized aspects of care, while front line nurses support daily living activities such as

preventing infections and turning patients in bed. One possibility is that managers may

have looked for ways to cut overall labor costs while changing the mix of nursing staﬀ

capability to maintain quality and patient experience, as RNs are crucial to nursing home

quality (Zhang and Grabowski, 2004; Lin, 2014). An alternative explanation is the

regulatory focus on RNs. For example, CMS uses the availability of RNs to determine

eligibility for Medicare reimbursement.29 Given the tight regulatory scrutiny of RN

availability, it is diﬃcult to reduce staﬃng levels in this category.

To explore whether the declines in staﬀ availability and quality are related, we compare

changes in staﬀ availability and Five Star ratings within target facilities around the PE buyout

event. This analysis recovers correlations and does not imply causality, so we present the

raw data in bin-scatter plots. Figure B.7 shows the change in Five Star rating over the

three years around PE acquisition on the Y-axis against the change in aggregate staﬀ hours

per patient day during the same period on the X-axis. The plots show that facilities which

experienced larger declines in staﬀ availability also experienced greater declines in ratings.

The patterns are consistent across rating types and suggest that cuts to nursing staﬀ may be

an important channel to explain the quality declines.

6.2.2 Finances and Operations

Our final analysis uses CMS cost reports to analyze key sources of expenditure related to

the PE business model. We begin by noting that nursing homes are widely known to have

relatively low and regulated profit margins, often cited at just 1-2%.30 Our data on nursing

home cost reports submitted to CMS indicate that nursing homes report negative operating

margins on average, and PE-owned nursing homes are not on average more profitable. In

unreported analysis, we see no eﬀect of buyouts on net income or overall revenue or costs.

29Specifically, such facilities are defined by having “an RN for 8 consecutive hours a day, 7 days

a week (more than 40 hours a week), and that there be an RN designated as Director of Nursing

on a full time basis.” See https://www.cms.gov/Regulations-and-Guidance/Guidance/Manuals/

Downloads/som107c07pdf.pdf.

30http://www.medpac.gov/docs/default-source/data-book/jun17_databookentirereport_

sec.pdf

32This raises the question of how PE firms create value from nursing home investments.

There are three types of firm expenditures that the academic literature and popular press

particularly associates with profits for PE owners. The first is what are often termed

“monitoring fees" charged to portfolio companies. In the CMS cost reports, these are listed

as “management fees”—charges to the nursing home for being owned and managed by a

PE firm.31 Metrick and Yasuda (2010) note that these are thought to be between 1-5% of

EBITDA. Our data suggest that they increase over time after buyouts, as shown in Figure 6

Panel A, where the fees are flat before the buyout, and then rise dramatically afterwards.

Table 8 Panel C column 1 indicates that on average, management fees increase by 7.7%

after acquisition (we exponentiate coeﬃcients in this panel as the outcomes are in logs).

The second type of expenditure is lease payments. The value of underlying real estate

is frequently cited as a reason that nursing homes and other typically low-margin assets can

be profitable investments, because the investor can sell the real estate to a related company

or to a third party (Dixon, 2007; Keating and Whoriskey, 2018; Brown, 2019). Cash from

the real estate sale can be disbursed as profits to the PE fund. A cash inflow early in the

life of the investment is particularly beneficial to the fund’s Internal Rate of Return, a key

performance metric. The nursing home assumes the obligation of future rent payments. As

an example, a New York Times report on the nursing home industry notes that:

“[PE] investors created new companies to hold the real estate assets because the

buildings were more valuable than the businesses themselves, especially with

fewer nursing homes being built. Sometimes, investors would buy a nursing

home from an operator only to lease back the building and charge the operator

hefty management and consulting fees” (Goldstein et al., 2020).32

31 In their summary of buyout fund economics, Metrick and Yasuda (2010) write that “we think of

monitoring fees as just another way for BO funds to earn a revenue stream.” These fees should not be confused

with the usual 2% of fund value that General Partners earn each year for managing Limited Partners’ capital,

before profits from investments.

32Two examples further illuminate these types of transaction. First, the HCR Manorcare deal discussed in

Section 2.2, where the chain’s real estate assets were spun oﬀ and sold shortly after the acquisition by the

Carlyle Group. Second, at a Congressional hearing the executive director of the Long-Term Care Community

Coalition said “more and more with entities buying up nursing homes, they have no experience in the business,

they sell out the underlying property” (Brown, 2019).

33Consistent with this strategy, column 2 shows that facility building lease payments increase

dramatically by about 75% after PE acquisitions. Figure 6 Panel B confirms the lack of

pre-trends and the increase post-buyout.

The third type of expenditure is interest on debt. While not a direct source of PE profits,

debt is tightly related to the overall PE model for creating value. Metrick and Yasuda

(2010) note that the ratio of debt to equity in a buyout deal is typically around 5:1. The

interest payments become a cost to the portfolio company. In Figure 6 Panel C, we see that

like the previous two outcomes, interest payments are flat before the buyout and then rise

dramatically afterwards. Column 3 indicates that the increase is about 325%.

Finally, we find that cash on hand declines after the buyout by 38%. Unlike the other

outcomes, the event study in Figure 6 Panel D indicates that cash on hand initially

increases after the buyout as profits increase and cash is injected, perhaps to invest in

eﬃciency improvements. However, as the strategies for returning profit to the investors are

implemented, such as selling the real estate and thus requiring the operator to take on lease

payments, the cash on hand turns negative. This could make the nursing home less

well-equipped to manage sudden shocks such as, for example, needing to buy personal

protective equipment following an infectious disease breakout.

Taking the results on nurse availability together with the estimated eﬀects on interest,

lease, and management fees payments, we infer that PE ownership shifts operating costs

away from staﬃng towards costs that are profit drivers for the PE fund. To our knowledge,

this paper oﬀers the first instance in the literature on PE in which these three profit drivers

have been documented systematically.

The final outcome we explore is patient capacity and volume. Table B.7 column 1 finds

no measurable change in the number of beds, which may partly reflect state regulations

restricting expansions. Admissions increase by 3.5%, or 6.5 patients per year for the average

facility (column 2). However, we interpret this eﬀect with caution since the corresponding

event study suggests a pre-trend (Figure B.8B). The apparent disconnect between demand

and quality of care may reflect information frictions in observing nursing home quality, as

discussed earlier (Arrow, 1963; Grabowski and Town, 2011; Werner et al., 2012).

34Higher admissions raise the question of whether PE ownership increases overall access

to nursing home care, providing care for individuals who would not otherwise have gone

to a nursing home. To test whether this is the case, we assess the eﬀects of PE entry into

a nursing home market, using the HRR definition. Table B.7 column 3 shows that there is

no eﬀect of initial PE entry on admissions at the market level, corroborated by flat patterns

in the event study (Figure B.8C). Hence, the data are more consistent with the facility-level

admissions increase reflecting business stealing.

7 Conclusion

This paper studies PE buyouts in healthcare, an important sector where PE activity has

increased dramatically, generating policy debate. Nursing homes are a useful setting

because they have particularly high levels of for-profit ownership and subsidy and have

experienced extensive PE investments. In an instrumental variables design incorporating

facility fixed eﬀects, we address both targeting and patient selection challenges to

identification. We find that going to a PE-owned facility increases 90-day mortality by

about 10% for short-stay Medicare patients, while taxpayer spending over the 90 days

increases by 11%. Furthermore, we document declines in nurse availability per patient and

in measures of compliance with Medicare’s standards of care. We also find a corresponding

increase in operating costs that tend to drive profits for PE funds.

There are many channels for future work. Although our results imply PE ownership

reduces productivity of nursing homes, it may have more positive eﬀects in other sectors

of healthcare with better functioning markets. Beyond healthcare, there has been significant

PE investment in sectors such as education, defense and infrastructure, which like healthcare

rely on high levels of government subsidy but are characterized by opaque product quality.

Further work is needed to determine how government programs can be redesigned to align

the interests of PE-owned firms with those of taxpayers and consumers.

35References

Acharya, Viral V, Oliver F Gottschalg, Moritz Hahn, and Conor Kehoe, “Corporate

governance and value creation: Evidence from private equity,” The Review of Financial

Studies, 2013, 26 (2), 368–402.

Adelino, Manuel, Katharina Lewellen, and Anant Sundaram, “Investment decisions of

nonprofit firms: Evidence from hospitals,” The Journal of Finance, 2015, 70 (4), 1583–

1628.

Appelbaum, Eileen and Rosemary Batt, “Private equity buyouts in healthcare: Who wins,

who loses?,” Institute for New Economic Thinking Working Paper Series, 2020, (118).

Arrow, Kenneth J, “Uncertainty and the welfare economics of medical care,” The American

Economic Review, 1963, 53 (1/2), 941–973.

Banerjee, Sube, Jennifer Hellier, Michael Dewey, Renee Romeo, Clive Ballard,

Robert Baldwin, Peter Bentham, Chris Fox, Clive Holmes, Cornelius Katona et al.,

“Sertraline or mirtazapine for depression in dementia (HTA-SADD): a randomised,

multicentre, double-blind, placebo-controlled trial,” The Lancet, 2011, 378 (9789), 403–

411.

Basu, Anirban, James J Heckman, Salvador Navarro-Lozano, and Sergio Urzua,

“Use of instrumental variables in the presence of heterogeneity and self-selection: an

application to treatments of breast cancer patients,” Health Economics, 2007, 16 (11),

1133–1157.

Bernstein, Shai and Albert Sheen, “The operational consequences of private equity

buyouts: Evidence from the restaurant industry,” Review of Financial Studies, 2016, 29

(9), 2387–2418.

, Josh Lerner, and Filippo Mezzanotti, “Private equity and financial fragility during the

crisis,” Review of Financial Studies, 2019, 32 (4), 1309–1374.

Biesinger, Markus, Cagatay Bircan, and Alexander Ljungqvist, “Value creation in

private equity,” EBRD Working Paper, 2020.

Bloom, Nicholas, Carol Propper, Stephan Seiler, and John Van Reenen, “The Impact

of Competition on Management Quality: Evidence from Public Hospitals,” Review of

Economic Studies, 2015, (82), 457–489.

, Raﬀaella Sadun, and John Van Reenen, “Do private equity owned firms have better

management practices?,” The American Economic Review, 2015, 105 (5), 442–446.

Boucly, Quentin, David Sraer, and David Thesmar, “Growth LBOs,” Journal of Financial

Economics, 2011, 102 (2), 432–453.

Brinch, Christian N, Magne Mogstad, and Matthew Wiswall, “Beyond LATE with a

discrete instrument,” Journal of Political Economy, 2017, 125 (4), 985–1039.

Brown, Benjamin, Eloise O’Donnell, and Lawrence P Casalino, “Private equity

investment in behavioral health treatment centers,” JAMA Psychiatry, 2020, 77 (3), 229–

230.

Brown, Danielle, “Lawmakers pepper private equity firms on nursing home investments,

quality issues,” McKnight’s Long-Term Care News, 2019.

36Brown, Sherrod, Mark Pocan, and Elizabeth Warren, “Letter to The Carlyle Group

(KewsongLee and Glenn A. Youngkin),” November 2019.

Bruch, Joseph D, Suhas Gondi, and Zirui Song, “Changes in hospital income, use, and

quality associated with private equity acquisition,” JAMA Internal Medicine, 2020.

Cadigan, Rebecca Orfaly, David G Stevenson, Daryl J Caudry, and David C

Grabowski, “Private investment purchase and nursing home financial health,” Health

Services Research, 2015, 50 (1), 180–196.

Card, David, Alessandra Fenizia, and David Silver, “The health impacts of hospital

delivery practices,” Technical Report, National Bureau of Economic Research 2019.

Carneiro, Pedro, James J Heckman, and Edward J Vytlacil, “Estimating marginal

returns to education,” American Economic Review, 2011, 101 (6), 2754–81.

Casalino, Lawrence P, “Private equity, women’s health, and the corporate transformation

of American medicine,” JAMA Internal Medicine, 2020.

Cawley, John, David C Grabowski, and Richard A Hirth, “Factor substitution in nursing

homes,” Journal of Health Economics, 2006, 25 (2), 234–247.

Chandra, Amitabh, Amy Finkelstein, Adam Sacarny, and Chad Syverson, “Health care

exceptionalism? Performance and allocation in the US health care sector,” American

Economic Review, 2016, 106 (8), 2110–44.

Charlson, Mary, Ted P Szatrowski, Janey Peterson, and Jeﬀrey Gold, “Validation of a

combined comorbidity index,” Journal of Clinical Epidemiology, 1994, 47 (11), 1245–

1251.

Clemens, Jeﬀrey and Joshua D Gottlieb, “Do physicians’ financial incentives aﬀect

medical treatment and patient health?,” American Economic Review, 2014, 104 (4), 1320–

49.

Cooper, Zack, Fiona Scott Morton, and Nathan Shekita, “Surprise! out-of-network

billing for emergency care in the United States,” Journal of Political Economy, 2020,

128 (9), 3626–3677.

, Stuart V Craig, Martin Gaynor, and John Van Reenen, “The price ain’t right?

Hospital prices and health spending on the privately insured,” The Quarterly Journal of

Economics, 2018, 134 (1), 51–107.

Cornelissen, Thomas, Christian Dustmann, Anna Raute, and Uta Schönberg, “From

LATE to MTE: Alternative methods for the evaluation of policy interventions,” Labour

Economics, 2016, 41, 47–60.

, , , and , “Who benefits from universal child care? Estimating marginal returns to

early child care attendance,” Journal of Political Economy, 2018, 126 (6), 2356–2409.

Currie, Janet and David Slusky, “Does the Marginal Hospitalization Save Lives? The

Case of Respiratory Admissions for the Elderly,” Technical Report, National Bureau of

Economic Research 2020.

Curto, Vilsa, Liran Einav, Amy Finkelstein, Jonathan Levin, and Jay Bhattacharya,

“Health care spending and utilization in public and private Medicare,” American

Economic Journal: Applied Economics, 2019, 11 (2), 302–32.

37, , Jonathan Levin, and Jay Bhattacharya, “Can health insurance competition work?

evidence from medicare advantage,” Journal of Political Economy, 2021, 129 (2), 570–

606.

Cutler, David M, “Where are the health care entrepreneurs? The failure of organizational

innovation in health care,” Innovation Policy and the Economy, 2011, 11 (1), 1–28.

and Mark McClellan, “Is technological change in medicine worth it?,” Health Aﬀairs,

2001, 20 (5), 11–29.

Dafny, Leemore, Mark Duggan, and Subramaniam Ramanarayanan, “Paying a

premium on your premium? Consolidation in the US health insurance industry,”

American Economic Review, 2012, 102 (2), 1161–85.

Davis, Steven J, John Haltiwanger, Kyle Handley, Ron Jarmin, Josh Lerner, and Javier

Miranda, “Private equity, jobs, and productivity,” The American Economic Review, 2014,

104 (12), 3956–3990.

Dixon, Kim, “US senators probe nursing homes run by private equity,” Reuters, 2007.

Doggett, Lloyd, Rosa DeLauro, Katie Porter, and Mary Scanlon, “Letter to Leonard

Green and Partners (Jonathan Sokoloﬀ),” June 2020.

Duggan, Mark G, “Hospital ownership and public medical spending,” The Quarterly

Journal of Economics, 2000, 115 (4), 1343–1373.

Dummit, Laura A, Nursing homes: Quality of care more related to staﬃng than spending,

US General Accounting Oﬃce, 2002.

Eaton, Charlie, Sabrina Howell, and Constantine Yannelis, “When investor incentives

and consumer interests diverge: Private equity in higher education,” The Review of

Financial Studies, 2019.

Einav, Liran, Amy Finkelstein, and Heidi Williams, “Paying on the margin for medical

care: Evidence from breast cancer treatments,” American Economic Journal: Economic

Policy, 2016, 8 (1), 52–79.

Eliason, Paul J, Benjamin Heebsh, Ryan C McDevitt, and James W Roberts, “How

acquisitions aﬀect firm behavior and performance: Evidence from the dialysis industry,”

The Quarterly Journal of Economics, 2020, 135 (1), 221–267.

Flood, Chris, “US Congress examines private equity role in surging healthcare costs,” The

Financial Times, 2019.

Fracassi, Cesare, Alessandro Previtero, and Albert W Sheen, “Barbarians at the store?

Private equity, products, and consumers,” Technical Report, National Bureau of Economic

Research WP 27435 2020.

Gadiesh, Orit and Hugh MacArthur, Lessons from private equity any company can use,

Harvard Business Press, 2008.

Gandhi, Ashvin, “Picking your patients: Selective admissions in the nursing home

industry,” Available at SSRN, 2020.

, YoungJun Song, and Prabhava Upadrashta, “Private Equity, Consumers, and

Competition: Evidence from the Nursing Home Industry,” 2020.

GAO, “Complexity of private investment purchases demonstrates need for CMS to improve

38the usability and completeness of ownership data,” Technical Report september 2010.

Garber, Alan M and Jonathan Skinner, “Is American health care uniquely ineﬃcient?,”

Journal of Economic Perspectives, 2008, 22 (4), 27–50.

Goldstein, Matthew, Jessica Silver-Greenberg, and Robert Gebeloﬀ,

“https://www.nytimes.com/2020/05/07/business/coronavirus-nursing-homes.html,”

The New York Times, 2020.

Gompers, Paul, Steven N Kaplan, and Vladimir Mukharlyamov, “What do private

equity firms say they do?,” Journal of Financial Economics, 2016, 121 (3), 449–476.

Gondi, Suhas and Zirui Song, “Potential implications of private equity investments in

health care delivery,” Journal of the American Medical Association, 2019, 321 (11),

1047–1048.

Grabowski, David C and Richard A Hirth, “Competitive spillovers across non-profit and

for-profit nursing homes,” Journal of Health Economics, 2003, 22 (1), 1–22.

Grabowski, David C. and Robert J. Town, “Does information matter? Competition,

quality, and the impact of nursing home report cards,” Health Services Research 46, 2011,

46 (6.1), 1698–1719.

Grabowski, David C, John R Bowblis, Judith A Lucas, and Stephen Crystal, “Labor

prices and the treatment of nursing home residents with dementia,” International Journal

of the Economics of Business, 2011, 18 (2), 273–292.

, Zhanlian Feng, Richard Hirth, Momotazur Rahman, and Vincent Mor, “Eﬀect of

nursing home ownership on the quality of post-acute care: An instrumental variables

approach,” Journal of Health Economics, 2013, 32 (1), 12–21.

Gupta, Arpit and Stijn Van Nieuwerburgh, “Valuing private equity strip by strip,”

National Bureau of Economic Research Working Paper, 2019.

Hackmann, Martin B, “Incentivizing better quality of care: The role of Medicaid and

competition in the nursing home industry,” American Economic Review, 2019, 109 (5),

1684–1716.

and R Vincent Pohl, “Patient vs. provider incentives in long term care,” Technical

Report, National Bureau of Economic Research 2018.

Hansmann, Henry B, “The role of nonprofit enterprise,” The Yale Law Journal, 1980, 89

(5), 835–901.

Harrington, Charlene, Brian Olney, Helen Carrillo, and Taewoon Kang, “Nurse staﬃng

and deficiencies in the largest for-profit nursing home chains and chains owned by private

equity companies,” Health Services Research, 2012, 47 (1pt1), 106–128.

, Helen Carrillo, Rachel Garfield et al., “Nursing facilities, staﬃng, residents and

facility deficiencies, 2009 Through 2016,” Menlo Park: The Henry J. Kaiser Family

Foundation, 2018.

Hart, Oliver, Andrei Shleifer, and Robert W Vishny, “The proper scope of government:

theory and an application to prisons,” The Quarterly Journal of Economics, 1997, 112

(4), 1127–1161.

Heckman, James J and Edward Vytlacil, “Structural equations, treatment eﬀects, and

39econometric policy evaluation 1,” Econometrica, 2005, 73 (3), 669–738.

, Sergio Urzua, and Edward Vytlacil, “Understanding instrumental variables in models

with essential heterogeneity,” The Review of Economics and Statistics, 2006, 88 (3), 389–

432.

Hill, Elaine L, David JG Slusky, and Donna K Ginther, “Reproductive health care in

Catholic-owned hospitals,” Journal of Health Economics, 2019, 65, 48–62.

Ho, Kate and Robin S Lee, “Equilibrium provider networks: Bargaining and exclusion in

health care markets,” American Economic Review, 2019, 109 (2), 473–522.

Hodgson, Leah, “Pandemic deepens woes of UK’s private equity-backed nursing homes,”

Pitchbook, 2020, September 21.

Huang, Sean Shenghsiu and John R Bowblis, “Private equity ownership and nursing home

quality: an instrumental variables approach,” International Journal of Health Economics

and Management, 2019, 19 (3-4), 273–299.

Institute of Medicine, Improving the quality of care in nursing homes, Vol. 85, Natl

Academy Pr, 1986.

Johnson, Richard W, “Who is covered by private long-term care insurance,” Urban

Institute. tinyurl. com/yxqklvdz. Retrieved December, 2016, 28, 2018.

Kaplan, Steven, “The eﬀects of management buyouts on operating performance and value,”

Journal of Financial Economics, 1989, 24 (2), 217–254.

Kaplan, Steven N and Antoinette Schoar, “Private equity performance: Returns,

persistence, and capital flows,” The Journal of Finance, 2005, 60 (4), 1791–1823.

and Per Stromberg, “Leveraged buyouts and private equity,” Journal of Economic

Perspectives, 2009, 23 (1), 121–46.

Keating, Dan and Peter Whoriskey, “Overdoses, bedsores, broken bones: What happened

when a private-equity firm sought to care for society’s most vulnerable,” The Washington

Post, 2018.

Kim, Tammy, “This is why nursing homes failed so badly,” New York Times, 2020,

December 31.

King, Robert, “Private equity firms in Congress’ crosshairs with legislation calling for

transparency,” Fierce Healthcare, 2020, February 12.

Kuehn, Bridget M, “FDA warns antipsychotic drugs may be risky for elderly,” Journal of

the American Medical Association, 2005, 293 (20), 2462–2462.

Lerner, Josh, Morten Sorensen, and Per Strömberg, “Private equity and long-run

investment: The case of innovation,” The Journal of Finance, 2011, 66 (2), 445–477.

Lewis, Adam, “Blackstone, KKR draw ire from Congress over surprise medical bills,”

2019.

Lin, Haizhen, “Revisiting the relationship between nurse staﬃng and quality of care in

nursing homes: An instrumental variables approach,” Journal of Health Economics, 2014,

37, 13–24.

, “Quality choice and market structure: A dynamic analysis of nursing home oligopolies,”

International Economic Review, 2015, 56 (4), 1261–1290.

40Martin, Anne B, Micah Hartman, Benjamin Washington, Aaron Catlin, National

Health Expenditure Accounts Team et al., “National health care spending in 2017:

growth slows to post–Great Recession rates; share of GDP stabilizes,” Health Aﬀairs,

2018, pp. 10–1377.

McClellan, Mark and Douglas Staiger, “The quality of health care providers,” Technical

Report 7327, National Bureau of Economic Research Working Paper 1999.

, Barbara J McNeil, and Joseph P Newhouse, “Does more intensive treatment of

acute myocardial infarction in the elderly reduce mortality?: analysis using instrumental

variables,” Journal of the American Medical Association, 1994, 272 (11), 859–866.

McGuire, Thomas G, “Physician agency,” Handbook of health economics, 2000, 1, 461–

536.

MedPAC, “Report to the Congress: Medicare payment policy,” March 2017.

Metrick, Andrew and Ayako Yasuda, “The economics of private equity funds,” The

Review of Financial Studies, 2010, 23 (6), 2303–2341.

Pelech, Daria, “Paying more for less? Insurer competition and health plan generosity in the

Medicare Advantage program,” Journal of Health Economics, 2018, 61, 77–92.

Phalippou, Ludovic, Christian Rauch, and Marc Umber, “Private equity portfolio

company fees,” Journal of Financial Economics, 2018, 129 (3), 559–585.

Pradhan, Rohit, Robert Weech-Maldonado, Jeﬀrey S Harman, Alex Laberge, and

Kathryn Hyer, “Private equity ownership and nursing home financial performance,”

Health Care Management Review, 2013, 38 (3), 224–233.

, , , and Kathryn Hyer, “Private equity ownership of nursing homes: implications

for quality,” Journal of Health Care Finance, 2014, 42 (2).

Press, Daniel and Michael Alexander, “Treatment of behavioral symptoms related to

dementia,” UpToDate. Available at: http://www. uptodate. com/contents/treatment-of-

behavioral-symptoms-related-todementia, 2013.

Ruﬃni, Krista, “Worker earnings, service quality, and firm profitability: Evidence from

nursing homes and minimum wage reforms,” Unpublished Mimeo, 2020.

Saenz, Andres, “How private equity can improve the health of healthcare,” EY Global

Private Equity Report, 2019.

Schneider, Lon S, Karen S Dagerman, and Philip Insel, “Risk of death with atypical

antipsychotic drug treatment for dementia: meta-analysis of randomized placebo-

controlled trials,” Journal of the American Medical Association, 2005, 294 (15), 1934–

1943.

Shleifer, Andrei and Lawrence H Summers, “Breach of trust in hostile takeovers,” in

“Corporate Takeovers: Causes and Consequences,” University of Chicago Press, 1988,

pp. 33–68.

Sink, Kaycee M, Karen F Holden, and Kristine Yaﬀe, “Pharmacological treatment

of neuropsychiatric symptoms of dementia: a review of the evidence,” Journal of the

American Medical Association, 2005, 293 (5), 596–608.

Skinner, Jonathan, “Causes and consequences of regional variations in health care,” in

41“Handbook of Health Economics,” Vol. 2, Elsevier, 2011, pp. 45–93.

Spaenjers, Christophe and Eva Steiner, “Do private equity investors create value?

Evidence from the hotel industry,” Available at SSRN, 2020.

Stevenson, David G and David C Grabowski, “Private equity investment and nursing

home care: is it a big deal?,” Health Aﬀairs, 2008, 27 (5), 1399–1408.

Taragano, Fernando E, Constantine G Lyketsos, Carlos A Mangone, Ricardo F

Allegri, and Enrique Comesaña-Diaz, “A double-blind, randomized, fixed-dose trial of

fluoxetine vs. amitriptyline in the treatment of major depression complicating Alzheimer’s

disease,” Psychosomatics, 1997, 38 (3), 246–252.

Tong, Patricia K, “The eﬀects of California minimum nurse staﬃng laws on nurse labor and

patient mortality in skilled nursing facilities,” Health Economics, 2011, 20 (7), 802–816.

Werner, Rachel M., Edward C. Norton, R. Tamara Konetzka, and Daniel Polsky, “Do

consumers respond to publicly reported quality information? Evidence from nursing

homes,” Journal of Health Economics, 2012, 31 (1), 50–61.

Zhang, Xinzhi and David C Grabowski, “Nursing home staﬃng and quality under the

nursing home reform act,” The Gerontologist, 2004, 44 (1), 13–23.

42Figures and Tables

Figure 1: Private Equity Ownership in Healthcare

A: PE Deals in Healthcare B: PE Deals in Eldercare

C: Number of Facilities and Patients Acquired D: Percentage of Facilities and Patients Acquired

Note: This figure shows PE deals in health care over time. Panels A and B present the total capital invested

(left axis) and number of transactions (right axis) by PE firms in healthcare and eldercare, by year. Panels C

and D focus on the number of active nursing homes owned by PE firms in each year. Panel B presents the

number of PE-owned facilities (left axis) and patients admitted at these facilities (right axis). Note that the total

number of facilities ever bought by PE firms is larger (1,674) than what is plotted here since some of these

facilities closed or went back to non PE ownership over time. Panel D presents these trends as a percentage of

total number of facilities and patients admitted, respectively.

43Figure 2: Patient Characteristics with Diﬀerential Distance

A: High Risk Patients

B: Stay at PE Nursing Home

Note: This figure presents scatter plots of patient characteristics against diﬀerential distance to the nearest PE

facility. The independent variable is the diﬀerence in distance (in miles) of the nearest PE nursing home to the

nearest non-PE nursing home for the patient. The dependent variable in Panel A is an indicator for the patient to

have a Charlson Co-morbidity Index (based on diagnoses recorded in hospital inpatient and outpatient claims

over the 3 months before admission to nursing home) greater than 2, and in Panel B is an indicator for the

nursing home being PE-owned. The data was collapsed into 20 equal sized bins and we plot the means of

residuals in each bin that were obtained from models including facility and patient HRR x Year fixed eﬀects,

and patient demographics: age, race, gender, maritial status, and an indicator if patient is dual eligible. The

figures also present quadratic fitted lines for these plots. Each plot also presents the slope coeﬃcient (per 10

miles of diﬀerential distance) with the corresponding standard error. Standard errors are clustered by facility.

44Figure 3: Patient Outcomes and Diﬀerential Distance

A: Mortality (Stay + 90 Days) B: Log Amount Billed per Patient Stay

C: Log Amount Billed per Patient Stay + 90 Days

Note: This figure presents results from reduced form regressions for patient-level outcomes on the instrument,

diﬀerential distance. Each blue point in the figure represents a coeﬃcient βs, obtained by estimating the

equation Yi = αm,t + αj + 5

s=2 βs 1(QDD = s)i + γ1 Xi + +ϵi, where 1(QDD = s)i is an indicator for the qth

quintile of diﬀerential distance. The highest quintile group, i.e., individuals relatively furthest away from a PE

facility, is the reference group. Log total payment in Panel B refers to the total payment for the index nursing

home stay. Standard errors are clustered by facility.

45Figure 4: Marginal Treatment Eﬀects

P(z)

0.6

0.5

0.4

0.3

0.2

0.1

0.0

-20 -10 0 10 20 30 40 50 60 70

Differential distance (miles)

A: First stage B: Common Support

0.050

MTE

0.08

0.07

0.06

0.05

0.04

0.03

0.02

0.01

0.00

-0.01

-0.02

-0.03

-0.04

0.050

0.040

Weights

0.030

0.020

0.010

0.000

0.030

0.020

Treatment effect

0.010

0.000

0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0

Unobserved resistance to treatment

0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0

Unobserved resistance to treatment

MTE 90% CI ATE (.013)

TT weights TUT weights

ATT (.031) ATUT (.01)

C: MTE curve

D: ATT and ATUT

Note: This figure presents results pertaining to Marginal Treatment Eﬀects (MTE) analysis using the Medicare

patient-level data. Panel A presents the ‘first stage’ fit of predicted probability of treatment or propensity score,

w.r.t the instrument. Panel B presents the overlap in distributions of PE and non-PE groups by propensity

score. This plot uses a log scale due to the large number of non-PE patients with low propensity. Appendix

figure B.6A presents the corresponding plot using a linear scale. Panel C presents the MTE curve with 90%

confidence intervals obtained using block bootstrap and the Average Treatment Eﬀect (ATE) estimate. Panel D

presents the weights for the Average Treatment on the Treated (ATT) and Average Treatment on the Untreated

(ATUT) and the corresponding estimates. Section 5.2.2 presents details of the MTE estimation.

46Figure 5: Aggregate Quality and Staﬃng Outcomes

A: Deficiency Survey Rating B: Overall Rating

C: Staﬀ Hour per Patient Day D: Nurse Asst. per Patient Day

E: Lic. Nurse per Patient Day F: Registered Nurse per Patient Day

Note: This figure presents event studies on quality of care measures (Five Star ratings) and Staﬃng around the

time a nursing home experiences a PE buyout. Each point in the figures represents the coeﬃcient βs obtained

by estimating Equation (10) as discussed in Section 6. Year= -2 is the omitted point. In Panels A and B, we

present eﬀects on the Five-star ratings awarded by CMS - deficiencies identified by independent contractors in

audits and overall rating, respectively. A negative eﬀect on ratings implies a decline in quality. Panels C to F

present results on nurse staﬃng per-patient for all staﬀ, nurse assistants, licensed nurses, and registered nurses

respectively. All models include facility and year fixed eﬀects, patient mix and market controls, as described

in Section 6.1. All dependent variables are winsorized at 1 and 99% level. Standard errors are clustered by

facility.

47Figure 6: Facility Finances

A: Log Management Fee Cost B: Log Building Lease Cost

C: Log Interest Cost D: Log Cash on Hand

Note: This figure presents event studies on facility finances around the time a nursing home experiences a

PE buyout. Each point in the figures represents the coeﬃcient βs obtained by estimating Equation (10) as

discussed in Section 6. Year= -2 is the omitted point. Panels A to D present results on the log of management

fee cost, building lease cost, interest cost, and cash on hand, respectively. All models include facility and

year fixed eﬀects, patient mix and market controls, as described in Section 6.1. All dependent variables are

winsorized at 1% and 99% levels. Standard errors are clustered by facility.

48Table 1: Descriptive Statistics

All Not PE-owned PE-owned

Mean SD Count Mean Count Mean Count

A. Facility Level Attributes

Overall Five-Star Rating 3.17 1.30 138,204 3.20 127,441 2.83 10,763

Deficiency Five-Star Rating 2.84 1.25 138,204 2.86 127,441 2.62 10,763

Staﬀ Hours per Pat. Day 3.59 1.49 284,108 3.60 271,118 3.38 12,990

Nurse Assistant Hours per Pat. Day 2.28 0.79 284,108 2.29 271,118 2.06 12,990

Licensed Nurse Hours per Pat. Day 0.82 0.46 284,108 0.82 271,118 0.82 12,990

Registered Nurse Hours per Pat. Day 0.46 0.57 284,108 0.46 271,118 0.49 12,990

Number of Beds 104.48 56.60 284,108 104.11 271,118 112.34 12,990

Admissions 184.16 166.97 284,108 180.40 271,118 262.47 12,990

Ratio Black 0.10 0.17 284,108 0.10 271,118 0.12 12,990

Ratio Medicaid 0.60 0.24 284,104 0.60 271,114 0.60 12,990

Ratio Medicare 0.15 0.17 284,104 0.15 271,114 0.18 12,990

Ratio Private 0.25 0.19 284,104 0.25 271,114 0.22 12,990

Management Fees (2016$) 7,076 120,673 231,795 6,001 219,231 25,833 12,564

Building Lease (2016$) 5,860 80,223 231,826 4,825 219,262 23,919 12,564

Interest Expense (2016$) 12,911 163,562 231,855 5,588 219,291 140,733 12,564

Cash on Hand (2016$) 1,110,000 10,600,000 231,811 1,150,000 219,257 516,772 12,554

B. Medicare Patient Attributes

Age 81.41 8.10 7,365,953 81.46 6,668,539 80.92 697,414

Female 0.64 0.48 7,365,953 0.64 6,668,539 0.62 697,414

Black 0.08 0.27 7,365,953 0.08 6,668,539 0.09 697,414

White 0.88 0.32 7,365,953 0.88 6,668,539 0.88 697,414

Married 0.34 0.47 7,365,953 0.34 6,668,539 0.35 697,414

Charlson Score (Previous) > 2 0.27 0.44 7,365,953 0.27 6,668,539 0.29 697,414

Cardio-Vascular Disease 0.18 0.39 7,365,953 0.18 6,668,539 0.18 697,414

Injury 0.19 0.39 7,365,953 0.19 6,668,539 0.19 697,414

Other 0.63 0.48 7,365,953 0.63 6,668,539 0.63 697,414

Dual Eligible 0.18 0.38 7,365,953 0.18 6,668,539 0.17 697,414

Diﬀerential Distance (Miles) 14.87 16.70 7,365,953 16.21 6,668,539 2.11 697,414

Mortality (Stay + 90 Days) 0.17 0.38 7,365,953 0.17 6,668,539 0.18 697,414

Starts Anti-Pyschotics 0.06 0.23 7,365,953 0.06 6,668,539 0.06 697,414

Mobility Reduces 0.54 0.50 7,365,953 0.53 6,668,539 0.62 697,414

Develops Ulcers 0.09 0.28 7,365,953 0.09 6,668,539 0.09 697,414

Pain Intensity Increases 0.27 0.45 7,365,953 0.27 6,668,539 0.30 697,414

Amount Billed per Patient Stay (2016$) 13,600 12,200 7,365,953 13,500 6,668,539 14,800 697,414

Amount Billed per Patient Stay + 90 Days (2016$) 21,100 20,100 7,365,953 20,900 6,668,539 22,600 697,414

Note: This table presents descriptive statistics for key variables used in the analysis. Panel A presents

descriptives on facility-level data for all nursing homes over the years 2000–17 while Panel B presents patient-

level data for Medicare patients with index stays over the years 2005–16. A unit of observation is a facility-year

in Panel A and a unique patient in Panel B (since we retain only the first stay per patient). Columns 1, 2 and 3

present means, standard deviations and number of observations for the full sample. We categorize facilities into

two groups. Columns 4 and 5 present means and number of observations at facilities that never experienced

a PE acquisition or before PE acquisition during our sample period. Columns 6 and 7 present corresponding

values for facilities in the post-buyout period. For most variables, about 10% of the observations pertain to

facilities that experienced a PE acquisition. Sample sizes diﬀer across variables in Panel A since they were

sourced from multiple sources or in some cases were reported only for more recent years. In Panel A, all

continuously varying variables were winsorized at the 1% and 99% levels. We compute the Charlson Co-

morbidity Index using co-morbidities diagnosed in hospital inpatient and outpatient claims (first 10 dx codes)

over the 3 months prior to, but not including, the index stay. Spending values in Panel B are winsorized at the

99% level and deflated to be in 2016 dollars. ‘Total’ billing includes hospital inpatient, outpatient including

emergency department, and nursing home stay spending over the 90 days following discharge from the index

stay and includes the index stay. The following patient-level variables were sourced from the Minimum Data

Set (MDS): marriage, antipsychotics, mobility, and pressure ulcers. Medicare patients that could not be merged

into the MDS (94% match rate) were dropped from the sample. Facilities with less than 100 Medicare patients

over the entire period were omitted from the patient-level sample. If any of the MDS variables was missing,

then we set the respective indicator to zero. We exclude patients facing a diﬀerential distance of greater than

70 miles, approximately the 95th percentile value, or below -70 miles.

49Table 2: Patient-Level Analysis: First Stage

(1) (2) (3) (4) (5)

1(PE) 1(PE) 1(PE) 1(PE) 1(PE)

Diﬀerential Distance -0.0480*** -0.0480*** -0.0479*** -0.0454*** -0.0419***

(In 10 Miles) (0.002) (0.002) (0.002) (0.002) (0.002)

(Diﬀerential Distance)2 0.0062*** 0.0063*** 0.0062*** 0.0059*** 0.0055***

(In 10 Miles) (0.000) (0.000) (0.000) (0.000) (0.000)

Market Controls Y

Patient Controls Y Y Y Y

Facility FEs Y Y Y Y Y

Patient FEs Level HRR x Year HRR x Year HRR x Year HSA x Year County x Year

Observations 7,365,934 7,365,934 7,358,129 7,365,752 7,365,246

Y-Mean 0.09 0.09 0.09 0.09 0.09

F-Stat 224 224 222 220 203

Note: This table presents estimates of the relationship between PE ownership of the nursing home and the

patient’s diﬀerential distance. Each cell presents the coeﬃcient β obtained by estimating Equation (1). The

independent variable is the diﬀerence in distance (both linear and quadratic, in 10 miles) to the nearest PE

nursing home and the nearest non-PE nursing home for the patient. This is calculated based on distances

between the respective zip code centroids. The outcome variable is an indicator for whether the nursing home

serving the patient is PE-owned (=1 if PE-owned, 0 otherwise). Column 1 controls for facility and patient

market (Hospital Referral Region) x Year fixed eﬀects. Column 2 (our preferred specification) adds controls

for patient risk controls (indicators for 17 pre-existing conditions used to define the Charlson Co-morbidity

Index inferred from claims over the three months prior to admission, and sex, age, race, marital status, and

an indicator if patients are dual eligible). Column 3 adds controls for patient zip-year characteristics: median

household income, the shares of the population that are white, that are renters rather than home-owners, that

are below the Federal poverty line, and that are enrolled in the medicare advantage program. Column 4 uses

the same controls as in col. 2 but defines patient market using a narrower market definition: Health Service

Area (HSA) instead of HRR. Column 5 uses the same controls as in col. 2 but defines patient market using a

narrower market definition: County instead of HRR. Standard errors are clustered by facility.

50Table 3: Balance of Patient Characteristics

(1) (2)

Patient Attribute DD < Median DD > Median

Diﬀerential Distance 2.70 27.04

PE-owned Nursing Home 0.17 0.02

Age 81.40 81.42

Female 0.64 0.64

Black 0.09 0.07

Married 0.35 0.34

Dual Eligible 0.16 0.19

AMI 0.08 0.08

Congestive Heart Failure 0.22 0.24

PVD 0.05 0.05

CEVD 0.13 0.14

Dementia 0.04 0.05

COPD 0.21 0.23

Rheumatoid Arthritis 0.03 0.03

Peptic Ulcer 0.02 0.02

Mild Liver Disease 0.01 0.01

Diabetes 0.21 0.22

Diabetes + Complication 0.04 0.04

Paraplegia 0.03 0.03

Renal Disease 0.14 0.13

Cancer 0.09 0.08

Severe Liver Disease 0.01 0.01

Metastatic Cancer 0.04 0.04

AIDS 0.00 0.00

Number Of Patients 3,683,135 3,682,818

Note: This table presents the balance in patient attributes with respect to the

instrument: diﬀerential distance. We divide patients into two groups based

on whether their diﬀerential distance is below or above the median value

(8.9 miles). Recall that diﬀerential distance (DD) is the diﬀerence between

distance to the nearest PE nursing home and the nearest non-PE nursing

home for the patient. Column 1 presents the means of patient characteristics

for patients with DD below the median value, while Column 2 presents

the means for patients with DD greater than the median. Characteristics

include four demographics and 17 pre-existing co-morbidity indicators used

to compute the Charlson Co-morbidity Index. Paraplegia includes both partial

and complete paralysis. We generated indicators for the 17 disease groups

using the ‘charlson’ command in Stata, available at http://fmwww.bc.

edu/RePEc/bocode/c/charlson.html. We considered diagnosis codes

on hospital inpatient and outpatient claims over the 3 months prior to, but not

including, the index nursing home stay.

51Table 4: Patient-Level Analysis: IV Results

A: Main Results

(1) (2) (3)

Mortality Log Amount Billed Log Amount Billed

(Stay + 90 Days) Per Patient Stay Per Patient Stay + 90 Days

1(PE) 0.0168** 0.1777*** 0.1054***

(0.007) (0.028) (0.024)

Observations 7,365,934 7,365,934 7,365,934

Y-Mean 0.17 9.07 9.57

F-Stat 224 224 224

B: Placebo Analysis

(1) (2) (3)

Mortality Log Amount Billed Log Amount Billed

(Stay + 90 Days) Per Patient Stay Per Patient Stay + 90 Days

1(PE) 0.006 -0.015 -0.016

(0.004) (0.018) (0.016)

Observations 7,159,535 7,159,535 7,159,535

Y-Mean 0.18 9.01 9.51

F-Stat 441 441 441

Note: This table presents estimates of the relationship between PE ownership and

patient health and spending. In Panel A, each cell presents the coeﬃcient β obtained

by estimating Equation (2) by 2SLS. The independent variable is an indicator for the

patient being admitted to a PE nursing home, instrumented by diﬀerences in distance to

the nearest PE and non-PE facility. Panel B presents results from a placebo analysis of

the relationship between private equity ownership and patient health and spending. For

this analysis, we use data over 2002–07, a period with very little actual PE ownership and

which has little overlap with the main analysis sample. We assign placebo PE acquisition

in 2004 to facilities that were eventually acquired before 2008 and 2005 to facilities

acquired in and post 2008 by PE firms. Accordingly we re-compute diﬀerential distance

values taking into account these placebo acquisitions. We present eﬀects for claims-

based patient quality outcomes - patient death within 90 days of discharge from the

index stay, and total amount billed (2016$). All regressions include facility and patient

HRR x Year fixed eﬀects, and patient risk controls. Patient risk controls include age,

race, gender, marital status, indicators for 17 pre-existing conditions used to compute

the Charlson Index, and an indicator if patients are dual eligible. Standard errors are

clustered by facility.

52Table 5: Heterogeneity in Patient Mortality

(1) (2) (3) (4)

Observations Mean Coeﬃcient (Std. Errors)

A: Patient Level

1. Age & Risk

Low Risk, 65-80 2,052,655 0.08 0.0186* (0.011)

High Risk, 65-80 881,854 0.24 -0.0346* (0.021)

Low Risk, 80+ 3,326,940 0.16 0.0319*** (0.011)

High Risk, 80+ 1,104,387 0.29 0.023 (0.020)

2. Gender

Male 2,640,611 0.21 0.0105 (0.012)

Female 4,725,295 0.14 0.0210** (0.008)

3. Beneficiary Zip Income

Income < Median 3,681,687 0.18 0.0122 (0.010)

Income > Median 3,684,035 0.16 0.0262** (0.011)

4. Race

White 6,483,451 0.17 0.0206*** (0.008)

Other 881,923 0.16 -0.0219 (0.023)

5. Reason for hospitalization

Cardio-Vascular 1,340,956 0.20 0.0298* (0.016)

Injury 1,409,910 0.11 0.0236* (0.014)

Other 4,615,012 0.18 0.0096 (0.009)

B: Market Level

1. Hirfindahl Hirschman Index

HHI < Median 3,706,810 0.16 0.0223 (0.020)

HHI > Median 3,659,035 0.18 0.0144* (0.008)

Note: This table presents heterogeneity in the eﬀects of PE ownership on patient mortality. Column

1 presents the sample size and Column 2 presents the mean. Columns 3 and 4 present the

corresponding coeﬃcient β and its standard error obtained by estimating Equation (2) by 2SLS.

The independent variable is the indicator for a patient being admitted to a PE nursing home,

instrumented by diﬀerences in distance to the nearest non-PE and PE nursing home. The outcome

variable is an indicator for patient death within 90 days of discharge from the index stay. Panel A

explores heterogeneity on patient level factors - by dividing patients into 4 groups based on severity

of pre-existing co-morbidities (high risk= Charlson Index greater than 2) and age (greater than 80)

in row 1, gender in row 2, median income in the patient’s zipcode in row 3, race in 4, and the reason

for hospitalization prior to the nursing home stay in row 5. Panel B explores heterogeneity based on

market factors - dividing markets below and above the median Hirfindahl Hirschman Index (HHI).

We computed the HHI using market shares in terms of beds as observed in 2003-04, where the

HRR in which the nursing home is located is considered its market. All models include facility and

patient HRR x year fixed eﬀects. We additionally control for the usual patient risk controls as in the

main regression. Standard errors are clustered by facility.

53Table 6: Patient wellbeing

(1) (2) (3) (4)

1(Starts Anti- 1(Mobility 1(Develops 1(Pain Intensity

Psychotics) Decreases) Ulcers) Increases)

1(PE) 0.0297*** 0.0425*** 0.0065 0.0273*

(0.006) (0.011) (0.008) (0.016)

Observations 7,365,934 7,365,934 7,365,934 7,365,934

Y-Mean 0.06 0.53 0.09 0.27

Note: This table presents estimates of the relationship between PE ownership and measures of patient

wellbeing obtained from clinical assessments. Each cell in the first row presents the coeﬃcient β obtained

by estimating Equation (2). The independent variable is an indicator for the patient being admitted to a PE

nursing home, instrumented by diﬀerences in distance to the nearest PE and non-PE facility. All models

include facility and patient HRR x Year fixed eﬀects. We additionally control for the usual patient risk controls

as in the main regression. The independent variable is an indicator for whether a nursing home is private

equity-owned (=1 if PE-owned, 0 otherwise) starting in the next year from the deal announcement date. We

present results for patient level outcomes - an indicator for patient starting anti- psychotics, decrease in patient

mobility, developing/worsening pressure ulcers, and increase in pain intensity. These variables take value one

if this condition is not flagged for the patient in the initial assessment, but is flagged at some point during the

stay. Standard errors are clustered by facility.

54Table 7: Patient-Level Analysis: Robustness

(1) (2) (3)

Mortality Log Amount Billed Log Amount Billed

(Stay + 90 Days) Per Patient Stay Per Patient Stay + 90 Days

1. Base Specification

1(PE) 0.0168** 0.1777*** 0.1054***

(0.007) (0.028) (0.024)

2. Chain Controls

1(PE) 0.0169** 0.1777*** 0.1055***

(0.007) (0.028) (0.024)

3. W/O Top 2 Deals

1(PE) 0.0309*** 0.2309*** 0.1429***

(0.011) (0.045) (0.037)

4. Top 5 Deals Only

1(PE) 0.0349*** 0.2469*** 0.1510***

(0.012) (0.046) (0.039)

5. Only For Profits

1(PE) 0.0138** 0.1474*** 0.0836***

(0.007) (0.026) (0.021)

6. Zip-Year Controls

1(PE) 0.0150** 0.1760*** 0.1029***

(0.007) (0.028) (0.024)

7. HSA-Year FEs

1(PE) 0.0211*** 0.1800*** 0.1130***

(0.008) (0.030) (0.025)

8. County-Year FEs

1(PE) 0.0221** 0.1430*** 0.0832***

(0.010) (0.034) (0.029)

9. No Controls

1(PE) 0.0296*** 0.2391*** 0.1131***

(0.008) (0.030) (0.024)

Observations 7,365,934 7,365,934 7,365,934

Y-Mean 0.17 9.07 9.57

Note: This table presents results from specification checks on the relationship between PE ownership and

patient health and spending. Each cell presents the coeﬃcient βobtained by estimating Equation (2) by 2SLS.

The independent variable is an indicator for the patient being admitted to a PE nursing home, instrumented by

diﬀerences in distance to the nearest PE and non-PE facility. We present eﬀects for patient death within 90

days of discharge from the index stay, the log of the total amount billed for the stay and the log of the amount

billed for the stay and across hospital inpatient, outpatient and nursing home over the 90 days following the

stay (2016$). All models include facility fixed eﬀects. The first six rows include HRR x year fixed eﬀects,

the seventh row uses Health Service Areas (HSA), and the eighth row uses county to define patient market

instead of HRR. The second row controls for facility being part of a chain. The third row calculates the results

excluding all data for chains involved in the 2 largest PE deals. The fourth row limits the PE group to only

the facilities bought in the 5 largest PE deals. The fifth row limits the sample only to for-profit facilities. The

sixth row includes patient zip controls: median household income, the shares of the population that are white,

that are renters rather than home-owners, that are below the federal poverty level, and that are enrolled in

Medicare Advantage program. The first eight rows includes patient risk controls: age, race, gender, marital

status, indicators for 17 pre-existing conditions used to compute the Charlson score, and an indicator if patients

are dual eligible. The ninth row presents coeﬃcients from a model with fixed eﬀects only. Standard errors are

clustered by facility.

55Table 8: Mechanisms and Operational Changes

A: Five Star Rating

(1) (2)

Deficiency Overall

Rating Rating

1(PE) -0.075** -0.079**

(No Control) (0.037) (0.036)

1(PE) -0.077** -0.082**

(With Control) (0.037) (0.036)

Observations 138,051 138,051

Y-Mean 2.9 3.2

B: Staﬀ Per Patient Day

(1) (2) (3) (4)

All Staﬀ Nurse Assistant Licensed Nurse Registered Nurse

1(PE) -0.050*** -0.068*** -0.019*** 0.037***

(No Control) (0.017) (0.010) (0.006) (0.005)

1(PE) -0.048*** -0.066*** -0.019*** 0.037***

(With Control) (0.016) (0.010) (0.006) (0.005)

Observations 283,767 283,767 283,767 283,767

Y-Mean 3.6 2.3 0.8 0.5

C: Log Financials

(1) (2) (3) (4)

Management Fee Building Lease Interest Expense Cash on Hand

1(PE) 0.074** 0.564*** 1.181*** -0.322***

(No Control) (0.032) (0.061) (0.096) (0.042)

1(PE) 0.074** 0.560*** 1.175*** -0.318***

(With Control) (0.032) (0.061) (0.096) (0.042)

Observations 231,556 231,584 231,613 231,569

Y-Mean 0.2 0.4 0.3 11.2

Note: This table presents estimates of the relationship between PE ownership and nursing home outcomes.

Each cell presents the coeﬃcient β obtained by estimating equation 9 with a diﬀerent outcome. The

independent variable is an indicator for whether a nursing home is PE-owned (=1 if PE-owned, 0 otherwise)

starting in the next year from the deal announcement date. Panel A presents results for quality outcomes

as measured by Five-star rating awarded by CMS - overall rating and deficiencies identified by independent

contractors in audits, respectively. A negative eﬀect on ratings implies a decline in quality. Panel B presents

results on per patient nurse availability for all nurses, nurse assistants, licensed nurses, and registered nurses.

Panel C presents results on the log of management fees, building lease cost, interest expenses, and cash on

hand. The top row presents results with no controls. The bottom row presents the results including controls,

which consist of market-level and patient mix controls, as described in Section 6.1. All models include facility

and year fixed eﬀects. All variables are winsorized at 1% and 99% levels. Standard errors are clustered by

facility.

56Appendix: For Online Publication

A Data appendix

This paper uses three primary data sources. We use (1) publicly available nursing home-

level data, (2) patient-level administrative claims data, both obtained from CMS, and (3)

Pitchbook data on PE deals. This section provides a detailed explanation of these data

sources and how we arrived at our analysis samples.

A.1 Nursing Home Data

Our data source on nursing home-level operations and performance is a compilation of

information obtained during annual surprise CMS inspector audits and data on nursing

home attributes and patient characteristics reported by the facilities themselves.33 The data

span 2000 through 2017. In each year we observe about 15,000 unique nursing homes, for

a total of approximately 280,000 observations. Of these, about 29,000 observations

represent facilities acquired by PE firms. The aggregate files provide annual data on basic

facility attributes, patient volume and case mix, nurse availability, and various components

of the Five Star ratings.34 These ratings started in 2009, so we cannot observe ratings

pre-buyout for deals before 2010. Fortunately, half of the PE deals in our sample,

accounting for 365 nursing homes, occurred post-2009.

Table 1 Panel A presents summary statistics on the Overall Five Star rating as well as the

other key nursing home-level variables used in the analysis. We first present the mean and

standard deviation for the whole sample (columns 1-2), then divide observations into two

groups—for facilities that are not PE-owned (columns 4-5) and for those that are (columns

6-7). We observe clear diﬀerences between PE-owned facilities and those not owned (all

statistically significant at the 1% level except where noted). PE targets are slightly larger,

have fewer staﬀ hours per resident, and a lower Overall Five Star rating. There have been

secular increases for the whole sector in both ratings and staﬃng over time. For staﬃng, this

reflects more stringent standards from regulators over time. Average staﬀ hours per patient

day increased from 3.5 in 2000 to 3.7 in 2017. Similarly, overall average Five Star ratings

increased from 2.9 in 2009 to 3.25 in 2017. As the PE deals occurred primarily later in the

sample, it is therefore remarkable that they have lower measures of quality on average.

A.2 Patient Data

Our second data source consists of patient-level billing claims and assessment data for

Medicare fee-for-service beneficiaries from 2004 to 2016. We observe the universe of

billing data for hospital care (inpatient and outpatient) and nursing homes for these

beneficiaries, as well as detailed patient assessments recorded in the Minimum Data Set

(MDS).35 We use these files to track beneficiaries’ demographics, spending, and health

33These files were organized and made available for research by the Long Term Care Focus research center

at Brown University. See www.ltcfocus.org for more details.

34For more details on how the ratings are produced, see Rating Guide.

35Specifically, we use 100% samples of the following: Medicare Beneficiary Summary File (MBSF),

Hospital inpatient and outpatient, and Skilled Nursing Facility claims files. These were obtained through a

reuse DUA with CMS and accessed through the NBER.

57outcomes such as mortality. The MDS helps observe clinical assessments such as mobility

and the use of antipsychotic drugs.

The unit of observation is a nursing home stay for a Medicare beneficiary that begins

during our sample period, which we begin in 2005 in order to have at least one look-back

year. Our main sample restriction is to identify index nursing home stays for patients,

defined as stays that begin at least a year after discharge from a previous nursing home stay.

This helps avoid mis-attributing adverse eﬀects to the wrong nursing home. To further

avoid attribution error, we consider only the patient’s first index stay in our entire sample

period. Hence, each patient appears only once in our sample. Using this approach, we settle

on a sample of more than seven million patients over 12 years. For each of these patients,

we also observe clinical assessments from the MDS, which we successfully match to the

claims files. Following the prior literature (Grabowski, Feng, Hirth, Rahman and Mor,

2013), we use some other restrictions to arrive at our sample. We restrict to patients over 65

years of age who are enrolled in Medicare parts A and B for at least 12 months before the

start of the nursing home stay. This restriction ensures that we observe prior medical care

history and pre-existing conditions. We also restrict to stays associated with a hospital visit

in the previous month, so that all patients are admitted after a hospital-based procedure and

are relatively homogeneous. We drop patients who went to a nursing home in a state other

than their state of residence as recorded in the Medicare master beneficiary summary file.

This drops a small fraction of patients (less than 5%) and is meant to exclude patients who

may be traveling when admitted to a nursing home. We match the index nursing home

stays to the MDS sample on beneficiary ID, facility ID, and admission date. We achieve a

match rate of 94% and drop unmatched patients. We drop facilities with fewer than 100

patients over the entire sample period to avoid special facilities and mitigate noise.

Table 1 Panel B presents summary statistics on the final patient-level sample. We use an

indicator for death within 90 days following discharge (including during the stay), based on

death dates recorded in the Medicare master beneficiary summary file. We use two measures

of spending. The first is the total amount that the nursing home bills to Medicare and the

patient for the index stay in 2016 dollars. Medicare covers the entire cost until the 21st day

of stay, at which point the patient begins paying a coinsurance, which has risen somewhat

over time and is now $170.5 per day.36 In our data, about 90% of total payments are by

Medicare. PE-owned facilities charge about 10% more than other facilities. The second

measure is the total amount paid for the stay and the 90 days following discharge. This

captures any subsequent hospital inpatient or outpatient care, and it provides a more holistic

picture of patient care.

Demographic measures associated with risk are quite similar across the types of

facilities, including patient age, the share of patients who are black and married, and the

Charlson Comorbidity Index, a standard measure of patient mortality risk based on

co-morbidities (Charlson, Szatrowski, Peterson and Gold, 1994).37 We create a high-risk

indicator that is one if the previous-quarter Charlson score is greater than two. According

to this definition, about 30% of patients are high-risk. The diﬀerence between facility types

is not significant.

Finally, we examine four measures of patient well-being which comprise inputs to the

36See https://www.resdac.org/cms-data/files/ip-ﬀs/data-documentation and

https://www.medicare.gov/Pubs/pdf/10153-Medicare-Skilled-Nursing-Facility-Care.pdf.

37The “Charlson score" assigns a point score to each of 17 disease categories recorded during the 3 months

before the index stay and sums them to create an overall disease burden score.

58quality portion of CMS’ Five Star ratings. The first is an indicator for the patient starting

antipsychotic medication during the stay. The second is an indicator for the patient’s self-

reported mobility score declining during the stay. The third is an indicator for developing a

pressure ulcer. The fourth is an indicator for the patient’s self-reported pain intensity score

increasing during the stay.

A.3 PE Deal Data

Our primary source of data on PE transactions is a proprietary list of deals in the “Elder

and disabled care” sector compiled by Pitchbook Inc., a leading market intelligence firm in

this space. The deals span 2004 to 2015. We match the target names to individual nursing

facilities using name (facility or corporate owner) and address as recorded in CMS data.38

Target names in these deals typically refer to holding companies, which often do not reflect

the names of individual facilities. The matching process required manual Internet searches

to confirm chain aﬃliations. We supplement the Pitchbook data in two ways. First, we

conduct additional Internet searches that yielded a small number of PE deals not reported by

Pitchbook. Second, we obtain a list of merger and acquisition deals from 2005 to 2016 from

Levin Associates, a market intelligence firm that tracks the healthcare sector. This helps us

to identify facilities that did not experience a new PE deal, but were acquired by an existing

PE-owned chain.39

This process yielded 128 deals, which correspond to a change in ownership to PE for

1,674 facilities. The deals are spread over time (no particular year or part of the business

cycle dominates) and across PE firms. Figure B.1 shows the number of deals in each year.

In total, our data contain 136 unique PE firms that acquired nursing homes. Most deals are

syndicated and involve multiple PE firms. Table B.1 presents the top 10 deals by number

of facilities acquired. Deal sizes are skewed, with the top 10 deals accounting for about

80% of all facilities acquired. On average, we observe PE-owned facilities for eight years

post-acquisition.40

It is diﬃcult to ascertain whether we comprehensively capture PE activity in this sector.

While there is no ‘oﬃcial’ tally of PE-owned nursing homes to benchmark against, our

sample size compares favorably against an estimate of 1,876 nursing homes reportedly

acquired by PE firms over a similar duration, 1998–2008 (GAO, 2010). Nonetheless, our

analysis likely underestimates the extent of PE activity in nursing homes, as matching

between Pitchbook deals and individual facilities is very challenging.

To understand whether deals are concentrated in particular regions, we plot the location

of PE-owned facilities across the U.S. in Figure B.2. PE firms appear to be more active in

large metropolitan markets, and in certain states such as Florida, Texas, New York,

Pennsylvania and Massachusetts. However, there is no obvious concentration, and we do

not find systematic variation with local measures of income, age, elder population, or share

38We obtain data on nursing home attributes (name, address, city, owner name and type, number of beds) and

quality measures (deficiencies) from Nursing Home Compare. See https://data.medicare.gov/data/

nursing-home-compare for more details.

39We matched approximately 290 additional facilities using information from the Levin files to the CMS

data. Of these, about 40 were PE-owned.

40A likely source of measurement error is not capturing PE disinvestment from facility ownership. For the

top 10 deals (80% of facilities) we verified PE exit via manual internet searches and incorporated it in the

analysis. The main results are robust to dropping observations of facilities that have been owned by PEs for 10

years or more. As expected, the coeﬃcients modestly increase in magnitude when we do so.

59of patients eligible for Medicare Advantage.

A.4 Targeting

This paper does not address why nursing homes may or may not be profitable acquisition

targets, and does not assess returns from investing. However, exploring what types of

facilities are targeted can help to interpret the eﬀects of buyouts on patient welfare and is

also useful for identifying the most relevant control variables for our empirical analysis.

We describe which characteristics are robustly associated with buyouts in Table A.1, which

presents estimates of Equation (11):

PE j,t = αs + αt + X′

j,t β+ ϵj,t (11)

Here, PE j,t is set to 100 if the facility j is acquired in a PE deal in year t (we drop all

years post-deal, and multiply by 100 for ease of reading). PEj,t is zero for never-PE and

PE-owned facilities before the deal. We include state and year fixed eﬀects.

We report models including variables known to be central to nursing home quality of care

and economics or that are potentially important and robustly predict buyouts. In column 1,

we find that facilities in more urban counties are more likely to be targeted.41 Urban nursing

homes tend to be closer to hospitals and likely enjoy thicker labor markets. Facilities in a

state with a higher ratio of elderly people are also more likely to be targeted. County-level

income, race, and home ownership do not predict buyouts. Results for these covariates are

not presented.

In column 2, we turn to facility characteristics. Chains are more likely to be acquired

than independent facilities, likely reflecting substantial fixed costs in deal-making. Hospital-

owned facilities are less likely to be targeted. PE firms also tend to target larger and higher-

occupancy facilities. We consider patient-level characteristics in column 3: the share of

the nursing home’s patients covered by Medicaid, the share on private insurance, and the

share who are Black. The first two are strongly negatively associated with buyouts, meaning

that a higher share of Medicare patients (the omitted group) is positively associated with

being targeted. In column 4, we assess two facility-level quality measures we employ in

the analysis: Five Star overall rating and staﬀ hours per patient day. Both are negatively

associated with buyouts, but once we control for rating, staﬃng is not significant. These

results indicate that PE firms target relatively low-performing nursing homes.

Finally, in column 5 we include simultaneously all of the variables from the previous

models that had predictive power. Some, such as admits per bed and hospital ownership,

become small and insignificant after controlling for the other variables. Notably, the state

elder ratio, chain indicator, and Five star rating retain their magnitudes and precision.

41We define urban as being in the top 2 out of 9 county groups classified as urban based on a Department of

Housing and Urban Development 2003 rural-urban classification.

60Table A.1: Targeting

Mean (1) (2) (3) (4) (5)

Urban Indicator 0.56 0.193*** 0.105**

(0.037) (0.041)

State Elder Ratio 0.24 4.340*** 18.819***

(1.328) (3.906)

1(Chain) 0.53 0.835*** 0.367***

(0.033) (0.029)

Hospital-Owned 0.07 -0.221*** -0.003

(0.053) (0.067)

Log(Beds) 4.5 0.287*** 0.086***

(0.030) (0.032)

Admits Per Bed 2.08 0.051*** 0.009

(0.007) (0.015)

Ratio Medicaid 0.60 -0.879*** -0.434*

(0.117) (0.229)

Ratio Private 0.25 -1.441*** -0.422*

(0.144) (0.236)

Ratio Black 0.10 0.002

(0.099)

Overall Rating 3.15 -0.075*** -0.066***

(0.015) (0.015)

Staﬀ Hr per Patient Day 3.55 -0.022

(0.018)

Observations Y-Mean (pp) 235,670 218,592 218,592 103,831 103,831

0.6 0.6 0.6 0.6 0.6

Note: This table shows estimates of the relationship between pre-existing nursing home characteristics and

whether a nursing home is a target of a PE buyout. Column 1 presents market-level attributes: an indicator for

urban and the share of state population which is elderly. Column 2 presents facility-level attributes: indicator

for being member of a chain, indicator for the nursing home being hospital-based, the log number of beds,

and admits per bed. Column 3 presents patient mix controls: share of patients covered by Medicaid, share of

patients who pay privately, and the share of patients who are black. Column 4 presents quality metrics such

as Five-star ratings awarded by CMS and staﬀ hours per patient day. We re-run the regression on all variables

which appear significant in Columns 1 to 4 in Column 5. The dependent variable is 100 if the nursing home

was acquired by PE in that year and 0 otherwise. We remove all observations of private equity-owned facilities

in years following the take-over by PE. We control for state and year FEs. Standard errors are clustered by

facility.

61B Supplementary Figures and Tables

Figure B.1: PE deals for Nursing Homes by Year

Note: This figure presents the number of unique deals for active nursing homes by PE firms for each year over

the period 2004–2015.

62Figure B.2: Location of Private Equity Targets

Note: This figure presents the number of facilities bought by PE firms in each county over the period 2004–

2015. We identified 1,674 such facilities.Figure B.3: Initial Patient Assesments

A: Dementia B: Alzheimers

C: Hip Fracture D: Urinary Tract Infection

Note: This figure presents event studies on initial patient assessments around the time a nursing home

experiences a PE buyout. To match the event study plots presented in the main text, we estimate these models

on collapsed facility-year level data and use the same specification, i.e., facility and year fixed eﬀects, patient

mix, and market controls, as described in Section 6.1. Each point in the figures represents the coeﬃcient

βs obtained by estimating Equation (10) as discussed in Section 6. Year= -2 is the omitted point. Panel A

presents results on the share of patients diagnosed with Dementia, Panel B on Alzheimers, Panel C on Hip

Fractures, and Panel D on Urinary Tract Infections, respectively, at admission to the index nursing home stay.

Standard errors are clustered by facility.

64Figure B.4: Patient distance to Nursing Home

A: CDF: All Patients B: CDF: PE vs. Non-PE Patients

C: Trendline: Share going to Closest Nursing Home D: Event Study: PE Acquisition

Note: This figure provides descriptes on patient zip code distance to index nursing home zip code. Panels A

and B present CDFs of the distance from patient zip code to index nursing home zip code. Panel A presents the

CDF pooling PE and non-PE patients together. It also identifies the median, 75th and 90th percentile values.

Panel B presents the CDFs separately for PE and non-PE patients, and their respective median values. Panel

C presents the annual trendline for the share of patients going to their closest nursing home. Panel D presents

the event study of the mean patient distance around a PE acquisition. Each point in the figure represents the

coeﬃcient βs obtained by estimating Equation (10) as discussed in Section 6. Year= -2 is the omitted point.

The model includes facility and HRR x year fixed eﬀects, patient mix, and market controls. Standard errors

are clustered by facility.

65Figure B.5: Patient Outcome Measures

A: Mortality (Stay + 90 Days) B: Log Amount Billed (Patient Stay + 90 Days)

C: 1(Starts Antipsychotic Medications) D: 1(Mobility Decreases)

E: 1(Develops Ulcers) F: 1(Pain Intensity Increases)

Note: This figure presents event studies on patient outcome measures around the time a nursing home

experiences a PE buyout. To match the event study plots presented in the main text, we estimate these models

on collapsed facility-year level data and use the same specification, i.e., facility and year fixed eﬀects, patient

mix, and market controls, as described in Section 6.1. Each point in the figures represents the coeﬃcient βs

obtained by estimating Equation (10) as discussed in Section 6. Year= -2 is the omitted point. Panels A and

B present results on the share of patients dying within 90 days of discharge from the index stay, and total

amount billed over the 90-day episode including the index stay (2016$). Panels C to F present results for

MDS assessment based outcomes - the facility level mean for indicators for patient starting antipyschotics,

decrease in patient mobility, developing/worsening pressure ulcers, and increase in pain intensity respectively.

Spending is winsorized at the 1% and 99% level. Standard errors are clustered by facility.

66Figure B.6: MTE: Additional figures

A: Common Support

MTE

0.06

0.05

0.04

0.03

0.02

0.01

0.00

-0.01

-0.02

-0.03

-0.04

0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0

Unobserved resistance to treatment

Baseline 3rd order 4th order 5th order

B: Order of polynomial

Note: This figure presents additional plots pertaining to Marginal Treatment Eﬀects (MTE) analysis using

the Medicare patient-level data. Panel A presents the overlap in distributions of PE and non-PE groups by

propensity score, using a linear scale for the Y-axis. Panel B demonstrates robustness of the slope of the MTE

curve to using diﬀerent orders of polynomials. Section 5.2.2 presents details of the MTE estimation.

67Figure B.7: Staﬀ Availability and Five Star Ratings

A: Overall Rating

B: Deficiency Rating

Note: This figure presents scatter plots of changes in total staﬀ hours available per patient day in the three years

post-PE buyout versus three years pre-buyout on the X-axis, against changes in CMS Five-star rating over the

same period on the Y-axis. Panel A presents overall rating, and Panel B presents survey based deficiency

ratings. The data was collapsed into 20 equal sized bins and we plot the means in each bin. The figures also

present fitted lines for these plots obtained using linear regressions on the underlying data. Each plot also

presents the slope coeﬃcient with standard error.

68Figure B.8: Patient Volume

A: Log Beds - Facility B: Log Admissions - Facility

C: Log Admissions - Market (HRR)

Note: This figure presents event studies on facility characteristics around the time a nursing home experiences

a PE buyout. Each point in the figures represents the coeﬃcient βs obtained by estimating Equation (10) as

discussed in Section 6. Year= -2 is the omitted point. Panels A and B present results on the log of beds and

admissions at the facility level, and Panel C on log admissions at the market level (HRR). All models – except

when studying market-level volume – include facility and year fixed eﬀects, patient mix, and market controls,

as described in Section 6.1. All dependent variables are winsorized at 1% and 99% levels. Standard errors are

clustered by facility.

69Table B.1: Top 10 Private Equity Deals

Sr. Target Private Equity Deal Number of

No. Name Firm(s) Year Facilities

1 Genesis Healthcare Formation Capital, JER Partners 2007–15 327

2 Golden Living Fillmore Capital Partners 2006 321

3 Kindred Healthcare Signature Healthcare, Hillview Capital 2014 150

4 HCR Manorcare Stockwell Capital, The Carlyle Group 2007–18 145

5 Mariner Healthcare Fillmore Capital Partners 2004 95

6 Skilled Healthcare Group Onex, Heritage Partners 2005–07 76

7 Trilogy Investors Lydian Capital Partners 2007–15 65

8 Lavie Care Centers Formation Capital, Senior Care Development 2011 61

9 Laurel Health Care Company Formation Capital, Longwing Real Estate Ventures 2006–16 41

10 Harden Healthcare NXT Capital, Oaktree Speciality Lending 2013 35

Note: This table presents some details on the top 10 PE deals in our sample, ordered by the number of unique

nursing home facilities involved in the deal. This represents the number of facilities we were able to identify

and match in our administrative data, the actual number of facilities in the deal may have been diﬀerent. We

set the PE indicator to turn on in the year following the deal year. If a closing year is mentioned, it implies the

PE investors exited or went public in that year. Accordingly, we turn oﬀ the PE indicator in the closing year.

70Table B.2: Complier Characteristics

Observations Coeﬃcient (Std. Errors) Ratio

Full Sample 7,365,934 -0.0445*** (0.003)

A. Age & Risk

Low Risk, 65-80 2,052,655 -0.0405*** (0.002) 0.91

High Risk, 65-80 881,854 -0.0471*** (0.003) 1.06

Low Risk, 80+ 3,326,940 -0.0451*** (0.003) 1.01

High Risk, 80+ 1,104,387 -0.0478*** (0.003) 1.07

B. Gender

Male 2,640,611 -0.0456*** (0.003) 1.02

Female 4,725,295 -0.0439*** (0.003) 0.99

C. Marital Status

Unmarried 4,838,365 -0.0446*** (0.003) 1.00

Married 2,527,548 -0.0439*** (0.003) 0.99

D. Beneficiary Zip Income

Income < Median 3,681,687 -0.0554*** (0.004) 1.24

Income > Median 3,684,035 -0.0353*** (0.003) 0.79

E. Race

White 6,483,451 -0.0451*** (0.003) 1.01

Other 881,923 -0.0380*** (0.003) 0.85

Note: This table presents first stage equivalent estimates of the 2SLS for various patient

subsamples. We present the coeﬃcient β, obtained by estimating the equation PEi =

αj + αm,t + β1(DDi > Median) + ϵi. 1(DDi > Median) is an indicator for patient

i’s diﬀerential distance to the nearest PE-owned facility being greater than the median

value. The model includes facility j and patient HRR x year fixed eﬀects, but no other

controls. We divide the sample by age and risk, gender, marital status, income in patient

zip code, and race. Details are available in Section 4.2. We also present the ratio of the

coeﬃcient obtained for each subsample to that for the full sample. Standard errors are

clustered by facilities.

71Table B.3: Mortality Eﬀects by Duration

(1) (2) (3) (4)

(Stay + 30 Days) (Stay + 60 Days) (Stay + 90 Days) (Stay + 365 Days)

1(PE) 0.009 0.0148** 0.0169** 0.0239***

(0.006) (0.007) (0.007) (0.008)

Observations 7,365,934 7,365,934 7,365,934 7,365,934

Y-Mean 0.12 0.15 0.17 0.24

F-Stat 223.9 223.9 223.9 223.9

Coeﬃcient/ Y-Mean 8% 10% 10% 10%

Note: This table presents estimates of the relationship between PE ownership and patient mortality. Each

cell presents the coeﬃcient β obtained by estimating Equation (2) by 2SLS. The independent variable is an

indicator for the patient being admitted to a PE nursing home, instrumented by diﬀerences in distance to

the nearest PE and non-PE facility. We present eﬀects for mortality at diﬀerent durations - patient death

within 30, 60, 90, and 365 days of discharge from the index stay. All regressions include facility and patient

HRR x year fixed eﬀects, and patient risk controls. Patient risk controls include age, race, gender, marital

status, indicators for 17 pre-existing conditions used to compute the Charlson Index, and an indicator for dual

eligibility. Standard errors are clustered by facility.

72Table B.4: Mortality Costs

(1) (2)

Male Female

A: IV estimates

1(PE) 0.0105 0.0210**

(0.012) (0.008)

Observations 2,640,611 4,725,295

Y-Mean 0.21 0.14

F-Stat 221 221

B: Placebo

1(PE) 0.0091 0.0044

(0.006) (0.005)

Observations 2,497,830 4,661,700

Y-Mean 0.23 0.15

F-Stat 431 440

C: Calculations

Number of Patients in PE Facilities 435,035 741,838

Additional Deaths 4,568 15,579

Total Lives Lost 20,146

Mean Life Expectancy 6.7 8.2

Additional Loss in Person Years 30,814 128,384

Total Person Years Lost 159,198

Value of Life Year (2016 $) 130,000

Total Cost (2016 $) 20.7 Billion

Note: This table presents estimates of additional deaths, life-years lost, and the

associated cost using standard estimates of statistical value of a life-year due to PE

ownership of nursing homes. Panel A presents the coeﬃcient βobtained by estimating

Equation (2) by 2SLS. The independent variable is the indicator for a patient being

admitted to a PE nursing home, instrumented by diﬀerences in distance to the nearest

non-PE and PE nursing home. The outcome variable is an indicator for patient death

within 90 days of discharge from the index stay. Panel B presents a placebo analysis for

this patient subsample using the same approach as for the whole sample, as presented

in Table 4. All models include facility and patient HRR x year fixed eﬀects and the

usual patient risk controls as in the main specification. Standard errors are clustered by

facility. Panel C presents calculations to estimate lives, life-years lost and total cost based

on Panel A coeﬃcients. We calculate average life expectancy at discharge (by gender)

using the observed distribution of lifespans for Medicare patients. For patients still alive

at the end of our sample, we assign a year of death based on patient gender and age

using Social Security actuary tables. We adjust downward the resulting life expectancy

to account for the fact the decedents tend to be older than the average nursing home

patient (about two years).

73Table B.5: Patient-Level Analysis: OLS Results

A: Initial Patient Assesments

(1) (2) (3) (4)

Dementia Alzheimers Hip Fracture Urinary Tract Infec.

At Admission At Admission At Admission At Admission

1(PE) -0.0098*** -0.0040*** -0.0034*** 0.0044**

(0.002) (0.001) (0.001) (0.002)

Observations 7,365,934 7,365,934 7,365,934 7,365,934

Y-Mean 0.16 0.05 0.09 0.16

B: Main Outcomes

(1) (2) (3)

Mortality Log Amount Billed Log Amount Billed

(Stay + 90 Days) Per Patient Stay Per Patient Stay + 90 Days

1(PE) 0.0034*** -0.0221*** -0.0118**

(0.001) (0.006) (0.005)

Observations 7,365,934 7,365,934 7,365,934

Y-Mean 0.17 9.07 9.57

C: Assesment Based Outcomes

(1) (2) (3) (4)

1(Starts Anti- 1(Mobility 1(Develops 1(Pain Intensity

Psychotics) Decreases) Ulcers) Increases)

1(PE) 0.0115*** 0.0349*** 0.0094*** 0.0266***

(0.001) (0.003) (0.003) (0.005)

Observations 7,365,934 7,365,934 7,365,934 7,365,934

Y-Mean 0.06 0.53 0.09 0.27

Note: This table presents OLS estimates of the relationship between PE ownership and patient health and

spending. Each cell presents the coeﬃcient βobtained by estimating Equation (2) by OLS. The independent

variable is an indicator for the patient being admitted to a PE nursing home. In Panel A, we present eﬀects for

initial patient assessments - dementia, alzheimers, hip fracture and urinary tract infection at time of admission.

In Panel B, we present eﬀects on patient death within 90 days of discharge from the index stay and total

amount billed during the stay and during the 90 day episode (2016$). Panel C presents results for assessment

based outcomes recorded in the MDS - an indicator for patient starting antipyschotics, decrease in patient

mobility, developing/worsening pressure ulcers, and increase in pain intensity. All regressions include facility

and patient HRR x Year fixed eﬀects, and patient risk controls. Patient risk controls include age, race, gender,

marital status, indicators for 17 pre-existing conditions used to compute the Charlson Index, and an indicator

for dual eligibility. Standard errors are clustered by facility.

74Table B.6: Robustness: Facility-Level Outcomes

A: Five Star Rating

(1) (2)

Deficiency Overall

Rating Rating

1. Chain Controls

1(PE) -0.074** -0.079**

(0.036) (0.028)

2. W/O Top 2 Deals

1(PE) -0.145*** -0.204***

(0.050) (0.042)

3. Only For Profit

1(PE) -0.077** -0.082**

(0.036) (0.028)

Observations 138,051 138,051

Y-Mean 2.9 3.2

B: Staﬀ Per Patient Day

(1) (2) (3) (4)

All Staﬀ Nurse Assistant Licensed Nurse Registered Nurse

1. Chain Controls

1(PE) -0.050*** -0.068*** -0.019*** 0.037***

(0.016) (0.010) (0.006) (0.005)

2. W/O Top 2 Deals

1(PE) -0.100*** -0.101*** -0.021** 0.030***

(0.026) (0.015) (0.009) (0.008)

3. Only For Profit

1(PE) -0.045*** -0.062*** -0.024*** 0.039***

(0.017) (0.010) (0.006) (0.005)

Observations 283,767 283,767 283,767 283,767

Y-Mean 3.6 2.3 0.8 0.5

C: Log Financials

(1) (2) (3) (4)

Management Fee Building Lease Interest Expense Cash on Hand

1. Chain Controls

1(PE) 0.074** 0.564*** 1.181*** -0.321***

(0.032) (0.061) (0.096) (0.042)

2. W/O Top 2 Deals

1(PE) 0.042 0.809*** 2.048*** -0.366***

(0.050) (0.102) (0.160) (0.068)

3. Only For Profit

1(PE) 0.056* 0.570*** 1.179*** -0.289***

(0.032) (0.061) (0.096) (0.043)

Observations 231,556 231,584 231,613 231,569

Y-Mean 0.2 0.4 0.3 11.2

Note: This table presents robustness tests on the estimates of the relationship between PE buyouts and Five

star ratings, nurse availability, and financials. The corresponding main results are presented Table 8. Each

cell presents the coeﬃcient βobtained by estimating Equation (9) with a diﬀerent outcome. The independent

variable is an indicator for whether a nursing home is PE-owned (=1 if PE-owned, 0 otherwise) starting in the

next year from the deal announcement date. We control for a chain indicator in the first row, remove the top 2

deals by size in the second row, and estimate the results on a sample limited to for-profit facilities in the third

row. We do not present results limiting to the Top 5 deals as Five Star ratings are only available post-2009,

and 4 Top 5 deals occurred before 2009. All models include facility and year fixed eﬀects. All variables are

winsorized at 1% and 99% levels. Standard errors are clustered by facility.

75Table B.7: Patient Volume

Facility Level Market Level

(1) (2) (3)

Log Log Log

Beds Admissions Admissions

1(PE) -0.002 0.036*** 0.014

(No Control) (0.003) (0.009) (0.014)

1(PE) -0.003 0.035*** 0.007

(With Control) (0.003) (0.009) (0.011)

Observations 283,767 283,767 5,364

Y-Mean 4.5 4.8 12.7

Note: This table presents estimates of the relationship between PE ownership and patient volume. Each cell

presents the coeﬃcient β obtained by estimating Equation (9) with a diﬀerent outcome. The independent

variable is an indicator for whether a nursing home is PE-owned (=1 if PE-owned, 0 otherwise) starting in

the next year from the deal announcement date. We present results on the log number of beds, log number

of admissions in facility, and log number of admissions at HRR level. The bottom row presents the results

including controls, which consist of market-level and patient mix controls, as described in Section 6.1. All

models include facility and year fixed eﬀects. All variables are winsorized at 1% and 99% levels. Standard

errors are clustered by facility.

Clean Up City of St. Augustine, Florida

Sunday, September 06, 2026

DOES PRIVATE EQUITY INVESTMENT IN HEALTHCARE BENEFIT PATIENTS? EVIDENCE FROM NURSING HOMES (Atul Gupta, Sabrina T. Howell, Constantine Yannelis & Abhinav Gupta, National Bureau of Economic Research, Working Paper 28474, February 2021)

No comments:

Copyright © 2006-2026, Ed Slavin, All Rights Reserved

Contribute to the cause!

Translate

About Me