Medical research and health care finance: Evidence from Academic Medical Centers

Abstract

Academic Medical Centers (AMCs) — comprising medical schools, teaching hospitals, and research laboratories — play an important role in US biomedical innovation. The Balanced Budget Act of 1997 changed reimbursements for Medicare inpatient claims and subsidies for medical residents. We compare AMCs’ relative exposure to the reform, how these differences affect their researchers’ ability to attract NIH grant funding, and the quantity, impact, and content of their publications. We find that in response to the reform, research activity increased by approximately 6%. Changes in research composition suggest that hospitals responded to Medicare funding cuts by encouraging incumbent investigators to increase their research activities and by redirecting hiring efforts towards individuals attracted to AMCs (e.g., translational researchers). We find little effect on clinical outcomes.

1. Introduction

Research dating back at least to Solow (1956) has highlighted the role that innovation plays in driving economic growth. But the production of novel ideas is ultimately filtered through institutions that provide the incentive structure for knowledge accumulation (Dasgupta and David, 1994; Mokyr, 2002; Rosenberg, 1963). The belief that basic research ultimately generates practical insights is perhaps the central assumption in post-war science policy, and that belief appears particularly well-founded in the health care industry (Gelijns and Rosenberg, 1995; Cutler and McClellan, 2001; Azoulay et al., 2020). However, in that setting and beyond, we have scant systematic understanding of how institutions—both in their formal (i.e., funding models, peer review) and informal (i.e., authorship and collaboration norms, tenure systems) aspects—support the transformation of scientific ideas into technological advances.

In this paper, we study Academic Medical Centers (AMCs). Alongside the NIH and pharmaceutical industry, these organizations play a central role in biomedical innovation. In the United States, 30 percent of health-related research is performed inside these institutions (Commonwealth Fund Task Force on Academic Health Centers, 1999), which bring together in one place a medical school, an owned or closely affiliated hospital, and basic research laboratories with an explicit triple mission of patient care, teaching, and research. Within the innovation system, AMCs are uniquely able to bring together the “ideas sector” of the health care economy (i.e., biomedical research) with its “production sector” (i.e., clinical care). This, in turn, facilitates the bidirectional flow of knowledge between the laboratory bench and the patient bedside (Rosenberg, 2009).

We examine how research within AMCs is shaped by the impact of an external financing shock. Traditionally, financial support for their research mission comes from three different sources: grants from the NIH and private foundations, contracts with the pharmaceutical industry, and importantly, cross-subsidies from patient care activities (Jones and Sanderson, 1996).¹ We examine how a sudden decrease in institutional funding influences the rate, impact, and direction of research within AMCs.

Health care financing cuts have ambiguous effects on the level of subsequent research within AMCs. Cuts in reimbursement levels may encourage hospitals and physicians to substitute effort away from research and towards patient care activities to offset payment reductions, resulting in a net decrease in subsequent research levels. On the other hand, in response to price reductions for patient care activities, hospitals may redirect hiring practices to attract productive researchers, and physicians may increase time spent on research supported by alternative funding sources rather than patient care. Together, this may lead to a net increase in research intensity. The implications for the subsequent quality and direction of research are similarly ambiguous. For example, financing cuts may cause researchers to decrease both high- and low-risk projects or compel them to produce more research projects with less time per project, leading to lower impact outcomes.

We exploit quasi-experimental variation in cuts to clinical care revenues induced by the Balanced Budget Act of 1997 (BBA). The BBA was a major reform that led to considerable reductions in Medicare reimbursements to hospitals (Seshamani et al., 2006a). Following a period of growth after the introduction of Medicare and Medicaid, the growth of clinical revenues slowed in the 1990s (Smith et al., 2005; Chen and Goldman, 2016). This slowdown was partially due to increased federal efforts to contain rising US health care expenditures. Our analysis exploits the fact that the BBA decreased add-on payments made to support graduate medical education, suggesting that the reform disproportionately affected teaching hospitals. Among teaching hospitals, some institutions were harder hit by the reform than others because of differences in their reliance on Medicare patients as well as specific subsidies that increased the price they received from Medicare for the typical discharge (MedPAC 2003).

Our empirical analysis focuses on two samples of hospitals—one that includes all teaching hospitals and one that focuses on a research-intensive set of hospitals. We assemble a rich dataset that includes, for each hospital over the period 1992–2007, grant applications, funded grants, publications, and patient outcomes.

Using a difference-in-differences model that takes advantage of cross-hospital variation in the exposure to the reform, we find that cuts to hospital financing meaningfully increased subsequent research output. We show that hospitals most affected by the BBA experienced a significant increase in subsequent grant applications (6 percent) and funded grants (5 percent). This increase is driven by “seasoned” rather than “novice” investigators. These findings stand in contrast with previous empirical work suggesting that restrictions to the funding environment can dampen subsequent research efforts (Furman et al., 2012; Tabakovic and Wollmann, 2019), as well as with the pronouncements of academic medical leaders at the time of the reform (Iglehart, 1999).

Next, we assess whether the reform has different effects on high- and low-impact publications. Impact is notoriously hard to quantify, so we construct three separate measures. These reflect how many times the publication is cited, whether the publication is cited by a patent, and whether the publication disrupts or consolidates the tenets of a scientific subfield. In each case, we do not find evidence of heterogeneity along these dimensions, with the reform increasing low- and high-impact publication rates in a symmetric fashion.

We also assess whether the effects of the reform vary across the vertical chain. Laboratory-based research, which most closely approximates basic science, is unaffected. Instead, our main effect loads on “translational” research, which bridges the gap from basic science to clinical applications, and clinical research, which is largely supported by industry. These results do not accord with practitioners’ accounts (Meador, 2015) and survey evidence (Weissman et al., 1999) pointing to the greater reliance of clinical investigators on institutional funds to support their research activities. On the contrary, our results are consistent with the view that researchers (or at least their employers) can adapt to negative funding shocks by seeking alternative sources of funding.

As is traditionally the case for difference-in-differences research designs, a threat to identification could arise if the treatment variable is correlated with unobserved trends in the outcome. This concern is salient in our setting because Congress increased the NIH budget from $13.7 billion to $27.2 billion over this five-year period (Congressional Research Service, 2020). It is plausible that more research-intensive hospitals are more likely to benefit from the NIH budget doubling, for example, because they have a research infrastructure in place that better enables them to scale up research activities. To address these concerns, our difference-in-differences specifications include separate trends for research-intensive and less research-intensive institutions. Though the magnitudes of the BBA effects are reduced, we continue to find a robust and sizable effect of the reform.

Finally, to better understand the overall consequences of the financing shock, we explore whether the BBA led physicians to substitute effort towards research, away from improving patient care activities. Looking at 30-day risk-adjusted survival rates across four conditions at the hospital level, we do not find any systematic association between BBA-exposure and subsequent clinical outcomes, suggesting that the positive impact of the BBA on subsequent research was not offset by worsening patient outcomes.

The primary contribution of this paper is to show how research institutions designed to support the transformation of scientific ideas into technological advances respond to changes in funding and the resulting impact on their research activities. Our results suggest they adjust their hiring practices and rely on alternative funding sources when one source is negatively affected, highlighting the resilience of these institutions. Our findings could generalize beyond the biomedical setting to other research institutions where disparate activities are linked and dependent on diverse funding sources. For example, university researchers are influenced by shifts in federal funding (Babina et al., 2023), endowments (Brown et al., 2014), and revenues and donations linked to positive football outcomes (Tabakovic and Wollmann, 2019). Similar to AMCs, universities may partially compensate for declines in federal funding by relying more on private sector funding (Babina et al., 2023).

This paper contributes to three branches of literature. First, this article contributes to the literature on research institutions. There is an extensive literature that examines the connection between the “ideas” and “production” sector of the health care economy (Rosenberg, 1982; Brooks, 1994; Stokes, 1997; Godin, 2006; Balconi et al., 2010; Ahmadpoor and Jones, 2017; Li et al., 2017; Azoulay et al., 2019; Fleming et al., 2019; Poege et al., 2019). However, there has been relatively little work on the institutions that bridge these sectors and facilitate research that integrates them both (i.e., translational research). Prior work on research institutions has primarily focused on the role of universities (Jaffe, 1989; Furman and MacGarvie, 2007; Lach and Schankerman, 2008; Belenzon and Schankerman, 2013; Hvide and Jones, 2018; Tartari and Stern, 2021) and national laboratories (Adams et al., 2003; Jaffe and Lerner, 2001). In this paper, we focus on a unique but important institution in the biomedical ecosystem: AMCs.

The second branch of literature addresses the effect of funding on innovation (Hall and Lerner, 2010; Kerr and Nanda, 2011). Prior work has shown that shifts in funds targeted towards research (e.g., grants, prizes) have important consequences for both research productivity (Jacob and Lefgren, 2011; Howell, 2017) and researcher participation (Tham et al., 2024). We show that in resource-constrained environments (Ganguli, 2017), alternative funding sources not dedicated to research can serve as crucial cross-subsidies for research activities. We highlight that reductions in these non-research funds can lead to shifts in hiring practices and substitution effects across funding resources, resulting in important effects on the rate and direction of subsequent research.

Third, this article contributes to the relationship between health care spending and innovation. Prior research has shown how expansions in health insurance have shifted innovation through shifting expected demand (Blume-Kohout and Sood, 2013). We extend this by showing how changes in insurance can also impact the supply of a key input to research: funding. In response to growing calls to reduce health care spending, our findings suggest that while policy changes aiming to decrease rents captured by health care providers may not lower overall research levels, they may significantly impact the composition of subsequent research.

The remainder of the paper proceeds as follows. Section 2 provides background information about AMCs and health care financing shocks. Section 3 describes the data. Section 4 analyzes the effect of health care financing cuts on the rate, impact, and composition of subsequent research. Finally, Section 5 provides a discussion and concludes.

2. Background and conceptual framework

2.1. Academic medical centers and biomedical research

In the prototypical view of biomedical research, potential treatment discoveries undergo a sequential development process. First, researchers trained in the “basic” life sciences discover a new molecule and show that it inhibits a particular disease pathway in vitro. Then, they develop animal models and gather initial data on the molecule’s safety and efficacy. The new molecule is subsequently turned over to physicians, who clinically test the purported treatment in randomized controlled trials. This stylized view underlies broad-based congressional support for the continuous public funding of biomedical research and drives most of the policy discussion.

However, this linear model of innovation, “however flattering to the scientist and the academic, is economically naïve and simplistic in the extreme” (Rosenberg, 1994: 139). Indeed, a closer examination of major treatment discoveries reveals a significantly more complex picture. In numerous cases, the first biological insight is acquired in a clinical setting, and only subsequently do bench scientists make sense of the mechanisms by which treatment is effective (Gershon, 1998). For example, scientists discovered the first antidepressant drug, iproniazid, because a related compound used to treat tuberculosis made patients so euphoric that they stopped taking it. Subsequent research on iproniazid led to the chemical theories of depression that have generated all later antidepressant agents (Wurtman and Bettiker, 1995).

In general, academic physicians have played an essential role in the development of various new medical technologies that rely on alternative development pathways. In some cases, such as the development of AIDS triple therapies, successful treatments resulted from the ongoing dialog between bench and bedside scientists (Wurtman, 1997). Other times, new clinical uses are discovered for therapies already introduced into clinical practice (Gelijns and Rosenberg, 1995; DeMonaco et al., 2012). Similarly, medical device users have also been instrumental in the invention of new products, from the identification of unmet clinical needs to builders of prototypes or initial field testing—a classic case of user-innovation (Gelijns and Rosenberg, 1995). An extensive set of case studies lend credence to the belief that researchers within AMCs play an essential function in enabling these alternative pathways for treatment discovery, a belief strongly echoed by the academic medical establishment (Crowley and Gusella, 2009).

For the past thirty years, research activities that focus on the bench-to-bedside interface have been labeled “translational”.² Typically, translational research requires expertise in molecular biology, genetics, and a clinical subspecialty. As a result, it is often performed by physician-scientists who split their time between clinical care and research activities. Within that category of researchers, the relationship between patient care and research varies, ranging from individuals who are “interested in a disease mechanism and even occasionally interested in seeing patients”, but “almost never interact in their research with an intact patient” to individuals that “actively search for patients who may enable them to uncover the secrets of complex diseases, care for those patients, and...undertake to explore new diagnoses and therapeutic approaches to treating their disease” (Nathan, 2005).

Rather than sitting midpoint on a continuum stretching from fundamental research in biology all the way to the testing of novel therapeutics in large-scale clinical trials, it is more appropriate to view translational research as belonging to “Pasteur’s Quadrant”, that is, in a class of investigations bringing forth ideas that are simultaneously valuable scientifically and a useful input into the treatment discovery process (Stokes, 1997; Murray, 2002). A well-known example is the work of Joseph Goldstein and Michael Brown, recipients of the 1985 Nobel Prize for Medicine and Physiology. Their initial investigations were inspired by observations of their own patients suffering from familial hypercholesterolemia (Goldstein and Brown, 1997). Through patient-inspired basic investigations performed at the laboratory bench, they identified the underlying root cause of this disease as a lack of low-density lipoprotein receptors. These discoveries, in turn, informed drug development efforts, ultimately leading to the market introduction of statins.

Distinguishing translational research from other types of biomedical research activities is not only more descriptively accurate but also necessary to understand how wider changes in the health care system might affect the rate and direction of research efforts. First, translational research — which often focuses on uncovering the pathophysiologic mechanisms of disease (Ahrens, 1992) — is very hard to appropriate, e.g., through patenting. As a result, its conduct is likely to be underfunded by private biopharmaceutical firms. Second, uniquely among the various types of biomedical research, translational activities are exceedingly difficult to perform outside of the AMC setting. In this respect they differ sharply from bench laboratory research, which often takes place in universities not affiliated with a medical school (e.g., MIT or UC Berkeley), independent research organizations (such as the Salk Institute), or pharmaceutical firms (Henderson, 1994; Flier, 2019). It also differs from clinical trial activities, which over the past three decades, have steadily migrated away from AMCs into a burgeoning ecosystem of for-profit experimental centers (Azoulay and Fishman, 2020).

Consisting of a hospital, research laboratories, and a closely-affiliated medical school, AMCs bring together the wide range of resources, expertise, and personnel necessary to enable treatment discovery.³ They employ laboratory scientists, clinicians, physician-scientists, physicians-in-training, and graduate students. In addition, they attract a large number of patients with diverse phenotypes. This rare confluence entails that these institutions occupy a unique position in the biomedical research ecosystem, with a unique potential to facilitate the flow of knowledge from the laboratory bench to the patient’s bedside (Rosenberg, 2009; Ali and Gittelman, 2016).

2.2. Research funding within academic medical centers

To support their research activities, AMCs primarily rely on three alternative sources of funding: grants (mostly from the National Institutes of Health, but also from other federal agencies and philanthropic foundations), contracts from biotechnology and pharmaceutical firms (including but not limited to the conduct of clinical trials), and “unsponsored” expenditures (i.e., lacking sponsorship from external sources, often a euphemism for cross-subsidies from clinical care).

NIH grants account for the largest share of research funding within AMCs — close to 70%, or $7.4 billion in 1997 (Commonwealth Fund Task Force on Academic Health Centers, 1999) — and AMCs capture more than two-thirds of NIH extramural grant expenditures in a typical year.⁴ However, beginning in the 1980s, the mix of research supported by NIH steadily shifted towards laboratory investigations mostly conducted by scientists holding a PhD.⁵ The perceived lack of balance has led to perennial calls by the academic medical establishment to protect funding for physician-scientists because they constitute an “endangered species” (Wyngaarden, 1979; Rosenberg, 1999; Jain et al., 2019). Over the past 25 years, task forces and blue ribbon commissions have suggested various avenues to reform peer review practices at NIH so as to avert a “crisis in clinical research” (Ahrens, 1992; Nathan and Wilson, 2003).

A second source of funding is contracts from industrial sponsors, typically (but not exclusively) for the conduct of clinical trials.⁶ Historically, AMCs’ infrastructure, proximity to patients, and clinical expertise meant that they were the natural locus for the great bulk of clinical trials sponsored by industry. This began to change in the 1990s as a new crop of for-profit experimental centers started to compete with AMCs to provide biopharmaceutical sponsors with a more streamlined and lower-cost alternative (Azoulay and Fishman, 2020). Private nonacademic clinical trial sites came to partly replace their academic counterparts because the academic incentive system sometimes struggles to reward the conduct of research sponsored by firms. First, fellow academics sometimes view investigators with ties to industry as “tainted” (Prasad, 2020); these perceptions have become heightened following a number of scandals involving human subjects protection and conflicts of interest (Stelfox et al., 1998; Baird et al., 2002; Nathan and Weatherall, 2002). Second, whereas investigator-initiated research makes unique demands on the creative and scientific potential of an academic, clinical trials involve a substantial relinquishing of intellectual autonomy since the investigator must adhere to an agreed-upon plan of research designed by others. As a result, participation in this activity does not produce career benefits commensurate with those generated by NIH-sponsored research—except, perhaps, for the “key opinion leader” recruited by industry to write the trial’s clinical protocol.

The last substantial source of funding for research within AMCs falls under the nebulous umbrella of “unsponsored research”. This funding includes expenditures by institutions or group practices, using rents from clinical care activities, endowment income, or institutional reserves, and by individual faculty members who devote their own resources to conduct research, drawing from discretionary funds or working additional uncompensated hours. These cross-subsidies are not systematically quantifiable, but survey evidence suggests that they play a non-negligible role in supporting AMCs’ research mission.⁷ While much less important quantitatively than sponsored research expenditures, unsponsored expenditures might support the careers of young investigators struggling to establish their own independent program of research, or those of established scientists whose traditional sources of funding have temporarily run out.

In contrast with sponsored research, where proposals undergo stringent peer review (in the case of NIH grants) or must pass a market test (in the case of clinical trials), the welfare impact of unsponsored research activities is ambiguous. First, the existence of cross-subsidies means that AMCs earn rents from their clinical care activities, so that the same quality of care might have been delivered to patients at a lower cost (Nicholson and Song, 2001). Second, the lack of transparency in the disbursement of institutional funds makes it susceptible to capture by entrenched interests, potentially resulting in low-quality or wasteful research by “hobby doctors”. Our evidence will indirectly speak to this question, by examining the quantity, impact, and composition of research by institutions differentially exposed to financial stress.

2.3. Medicare payments and financing cuts

To provide plausibly exogenous variation in the extent of financial slack faced by AMCs and teaching hospitals, we rely on a measure of exposure to the Medicare reimbursement cuts triggered by the Balanced Budget Act of 1997.⁸

Under a system known as the Prospective Payment System, Medicare reimburses most hospitals on a per-admission basis (See Appendix A for a detailed discussion). In turn, each admission payment is a function of three types of adjustments: indirect medical education (IME) subsidies, disproportionate share hospital (DSH) payments, and outliers payments (Nicholson, 2002). IME payments are meant to compensate teaching hospitals for indirect expenses stemming, for example, from use of diagnostic services by clinically inexperienced residents or decreased productivity of nurses and support staff involved in the teaching of residents. DSH payments correspond to payments received by hospitals for the additional cost of treating poor patients. Finally, outlier adjustments are reimbursements to compensate providers for patients with exceptionally costly stays (Keeler et al., 1988).

The BBA reduced the scale of all three adjustments and slowed the growth of payments for all diagnosis-related groups (See Appendix A for details). As a result, Medicare inpatient payments decreased by approximately 5 percent between 1998 and 2000 and many hospitals saw their financial status deteriorate significantly (Iglehart, 1999; Dickler and Shaw, 2000; Seshamani et al., 2006a; Shen and Wu, 2013). Importantly, the reform was not anticipated, as its passage depended on the delicate balance of power between a reduced Republican majority in Congress and the Clinton administration following the November 1996 Federal election (Kahn and Kuttner, 1999).⁹

As 30 percent of inpatient visits are funded by Medicare, cuts to Medicare reimbursements can represent a significant financial strain on hospitals. The effect of financial cuts on hospital activities has been widely studied in the realm of clinical care, with inconclusive results. In theory, hospitals may respond to reduced Medicare payments by cost-shifting — i.e., increasing prices for privately-insured patients—or cost-cutting — i.e., lowering hospital costs, decreasing support for unprofitable services (Cutler, 1998; David et al., 2014). The empirical evidence to date has shown that, in response to adverse financial shocks, hospitals cut costs by lowering the quality of care (Lindrooth et al., 2007). In contrast, Wu (2010) finds that in response to the BBA, hospitals were limited in their ability to shift costs to private insurers (mostly due to the rise of managed care delivery models during this period).

2.4. Conceptual framework

While the BBA clearly led to significant financial pressures on AMCs, the impact on their research missions remains unclear. In this section, we provide a brief conceptual framework for assessing the expected impact of Medicare reimbursement cuts on the rate, impact, and composition of research activities within AMCs.

First, looking to the rate of research, the BBA could have impacted both the number of individuals that participate in research at AMCs (i.e., the extensive margin) and the research activity of existing physician-scientists (i.e., the intensive margin). On the extensive margin, the BBA may have affected AMCs’ hiring practices or exit rates among incumbent researchers. Physicians commonly move across institutions due to professional opportunities and economic pressures (Benham et al., 1968; Steiber, 1982). One study found that approximately 25 percent of physicians move from one institution to another (Ricketts, 2013). In response to Medicare funding cuts, hospitals might have redirected hiring efforts towards individuals attracted to AMCs (e.g., translational researchers) and those who can more easily secure non-Medicare funding (e.g., trialists and others seeking private sector funding); alternatively, they may have slowed faculty hiring.

In addition to influencing research activity along the extensive margin, the BBA could also shift incumbent researchers’ activities. For incumbent scientists, the net effect of financing cuts on research levels theoretically depends on the relative strength of the substitution and income effects (McGuire and Pauly, 1991; Yip, 1998; Jacobson et al., 2010). If substitution effects dominate, scientists may substitute other career-advancing or revenue-generating activities — such as industry-funded clinical trials — for clinical care. In this scenario, financing cuts would result in an increase in more easily-funded research activities (i.e., research activities not funded by Medicare). This is consistent with Ganguli (2017), who finds that in financially constrained environments, alternative funding sources can significantly impact scientists’ research productivity. Alternatively, if income effects dominate, researchers may instead direct more effort toward patient care, leading to a reduction in research activities.

The impact of the BBA on research quality is similarly ambiguous. For example, if the quality of a scientist’s research output is dependent on their expertise and the amount of time they allocate to research (Bobtcheff et al., 2017), and the BBA encouraged AMCs to hire more experienced researchers or led incumbent scientists to devote more time to research, we might expect an increase in the average quality of research outcomes. On the other hand, the BBA may have increased pressure on physician-scientists to secure funding from non-institutional sources. This could result in less time being allocated to each research project and an overall decline in average research quality.

Finally, we consider how the BBA may have impacted the composition of subsequent research activity. Qualitative accounts from physicians contemplating integrating a research component into their clinical practice often highlight the lumpy adjustment costs that need to be incurred in this process, with most expressing skepticism that their institutions could adequately support these efforts using the dwindling rents from clinical care activities. Because of the lack of fungibility between research and clinical care efforts in basic research, our prior is that the BBA would disproportionately affect within-physician substitution activities associated with clinical research. However, the direction of the effects is ambiguous.

To summarize, the conceptual framework presented in this section indicates a range of potential impacts on research productivity with AMCs. Hospitals may respond along both the extensive and intensive margin, with ambiguous effects on the rate, impact, and direction of research. In light of this uncertainty, we now present empirical estimates of how AMC research changed in response to the BBA.

3. Data

3.1. Data construction and sources

Our analysis combines data from several primary sources (i) hospital characteristics from the Healthcare Cost Report Information System (HCRIS) and the Inpatient Prospective Payment System (IPPS) Payment Impact Files; (ii) administrative data on NIH grants from the NIH IMPAC II database; (iii) publication and citation data from PubMed and the Web of Science, respectively; (iv) patent-to-publication citation linkages (from the USPTO and PubMed combined); and (v) hospital-level clinical outcomes data. Appendix Figure B1 describes how these data sources fit together and how we construct the variables used in this analysis.

We leverage information from HCRIS, which contains administrative data covering the universe of Medicare-certified hospitals. We identify hospitals with information from 1992 to 2007. For each hospital record, we observe hospital characteristics data, such as the total number of patient discharges, inpatient days, IME payments, and DSH payments. Medicare pricing information from the 1995 IPPS Payment Impact Files is used to calculate hospitals’ average PPS revenue per discharge.

We supplement this hospital-level dataset with several measures of research activity. First, we link the hospital dataset to grants data from the NIH IMPAC II database.¹⁰ For each grant, we obtain information on investigators, their institutions, and a number of project characteristics. Next, we collect data on publications from PubMed, the public-access database which indexes the scientific literature, and we obtain publication citation data from the Web of Science (up to 2015).

A key challenge for estimating the causal impact of the BBA is that the level of the shock (changes to Medicare funding due to the BBA) and the measures of research outcomes (NIH grants and publications) do not coincide. An analysis that examines how Medicare payments at the individual hospital-level affect NIH grants allocated at the medical-school level is likely to produce estimates that suffer from measurement error. We overcome these challenges by employing an outcome assignment mechanism that uses principal investigator (or author) addresses to allocate each grant (or publication) to the “correct” hospital.

As an example, the University of California, San Francisco Medical Center includes the Parnassus Heights Campus, Mt. Zion Hospital and Medical Center, and San Francisco General Hospital (see Appendix Figure C1). Each of these locations has a unique Medicare provider number and therefore receives their own specific Medicare payments. However, the three campuses share a single, common NIH institutional code. Our strategy for matching hospitals to grants consists of looking at each of the PI addresses affiliated with the UCSF NIH institutional code and allocating each PI (and grant) to one of the three hospitals.

We execute a similar strategy to match hospitals to publications: for each publication, we use the Web of Science to determine its authors’ institutional affiliation and address. This allows us to match a publication to a hospital by matching on both the hospital name and address.¹¹

The final sample consists of all hospitals that show evidence of teaching and research activity. To identify the set of research-active hospitals, we make several restrictions. First, we start with a list of 1,195 unique hospitals (as captured by unique Medicare provider numbers in the HCRIS data between 1992 and 2007).¹² We drop 210 specialty hospitals as they are not paid under Medicare’s Prospective Payment System, leading to a sample of 985 hospitals.¹³ Next, we exclude any hospitals that close during the 4 year period between 1992 and 1995 (the focal pre-BBA year in our analysis) by restricting our hospital sample to those with all observations between 1992 and 1995, resulting in 920 hospitals.¹⁴ Finally, to focus on a subset of hospitals for which the BBA is most likely to induce a meaningful shift in hospital outcomes, we restrict our sample to hospitals that receive (1) at least one indirect medical education payment and (2) produce at least one publication or submit at least one NIH grant application between 1992 and 2007. This results in a final sample consisting of 780 teaching hospitals (hereafter the “teaching hospital sample”).

In addition to our primary hospital sample, we also consider a second, more research-focused subset of the teaching hospital sample (the “AMC sample”). To create this second hospital sample, we follow the definition of a “major teaching hospital” used by Burke et al. (2017). Major teaching hospitals are members of the Council of Teaching Hospitals (COTH) and have a medical school affiliation reported to the American Medical Association.¹⁵ This results in a sample of 274 hospitals.

The primary outcomes of interest are the total number of grant applications, grant awards, and publications that accrued to the researchers affiliated with a particular hospital in a given year. In addition, we characterize grants by grant cycle and principal investigator type. To assess how the BBA shapes research activity across principal investigators with varying levels of “grantsmanship” experience, we designate applications and funded grants as coming from a “novice” or “seasoned” investigator.¹⁶ A novice grant application comes from an investigator whose first NIH grant application occurred within the last 6 years; in contrast, a seasoned grant application comes from an investigator who applied to NIH for the first time more than 6 years ago.¹⁷ We use a similar approach to assess how the BBA shapes hospital hiring practices by designating a grant application as a coming from an “incumbent” or “new hire” based on the timing of the applicant’s first application associated with the focal hospital.

Further, we go beyond these raw counts by investigating how health care financing cuts shape the impact and composition of subsequent research. We rely on three measures of research impact. For our first measure, we assign each article to one of six mutually exclusive bins: those that fall in the bottom quartile of the article-level distribution of long-run citations, those that fall in the second quartile, in the third quartile, in the top quartile but not in the top ventile, in the top ventile but not the top percentile, and finally publications in the top percentile of the citation distribution. The percentile rankings are vintage-adjusted, which allows us to compare the citation impact of publications published in different years.¹⁸ Our second measure of research impact comes from identifying the number of publications that were subsequently cited by a patent (Marx and Fuegi, 2020).

Finally, our third measure of research impact comes from identifying each publications’ “disruptive” impact using an index recently proposed by Funk and Owen-Smith (2017). Their citation-based measure captures the degree to which the ideas embodied in a paper consolidate or destabilize the scientific status quo. An article is considered “consolidating” if it tends to reference the same publications as the articles that will cite it in the future. Conversely, “disruptive” research draws on articles that will not be acknowledged by its own citing papers. For each hospital in the sample, we compute in every year the number of articles published that fall above the 95th percentile of their index.

To investigate how the BBA shifts the allocation of research along the bench-to-bedside continuum, we construct measures of basic laboratory research (hereafter characterized as “laboratory” research), translational research, and clinical research. To partition the research space in this way, we take advantage of Medical Subject Headings (MeSH) terms, a hierarchical controlled vocabulary maintained by the National Library of Medicine. Most publications listed in PubMed are tagged with a set of MeSH terms, which characterize their scientific content. Using the MeSH-based definitions outlined in Azoulay et al. (2020), we designate articles as bench research if they fulfill three criteria: (i) they are not disease-oriented (i.e., contain no disease MeSH terms); (ii) they do not report the results of a clinical trial (which we ascertain by examining the publication type field in PubMed); and (iii) they are tagged by MeSH terms denoting either the use of a molecular biology technique, the use a model organism, the study of cellular structures and macromolecules, or the study of biochemical and cellular processes.

Next, we identify the set of publications that can plausibly be deemed to be translational in nature. Following Azoulay et al. (2020), we generate three MeSH-based measures related to translational research. First, we label a publication translational if it is (i) disease-oriented; (ii) not a clinical trial; and (iii) also tagged by a basic science keyword used to describe bench research. Next, we denote a publication as “inspiring translational research” if it is translational and is cited by at least one clinical trial publication. Finally, we identify work that “builds on translational research”, i.e., those that report the results of a clinical trial and list a translational publication in the reference list.

Our final measure of research outcomes is meant to capture research investments directly relevant for the bedside—i.e., clinical research. We focus on two measures: (i) clinical trial articles (identified using MeSH terms corresponding to clinical trials or the publication type field in PubMed) and (ii) “other” clinical articles, which are disease-oriented, not clinical trials, and not tagged by any bench MeSH keywords.

3.2. Financial pressure from the BBA

To evaluate the impact of the BBA on subsequent research activity, we exploit the fact that some hospitals were more exposed to the reform than others. Following Shen (2003), we construct a “BBA Bite” variable that is the BBA-induced change in PPS revenue per discharge weighted by the share of Medicare patients in a base, pre-BBA year. As discussed in Section 2.3, the BBA changed the formula for Medicare subsidy payments, which are a function of teaching load and the disproportionate share of poor patients. This suggests that cross-hospital variation in the BBA’s impact comes from differences in (i) hospital characteristics (e.g., the number of residents) and (ii) the hospital’s reliance on Medicare.

To identify the effect of the BBA on PPS revenue per discharge, we rely on a simulated revenue approach. In particular, we use pre-BBA data to simulate the revenue that hospitals would have lost under the BBA had the reform been in effect earlier. Using pre-BBA data isolates the mechanical effect of the policy from hospitals’ endogenous responses. We simulate the average PPS revenue per discharge that hospitals would have received from Medicare had the BBA occurred in 1995, two years prior to its actual implementation.¹⁹ The mechanical effect of the BBA is estimated as the difference between the simulated revenue per discharge and actual PPS revenue per discharge in 1995:

$$Sim\Delta Re{v}_{h,1995}=Re{v}_{h,1995}-SimRe{v}_{h,1995}$$

where $Re{v}_{h,1995}$ is the log of true PPS revenue per discharge in 1995 and $SimRe{v}_{h,1995}$ is the log of simulated PPS revenue per discharge in 1995. Both revenue measures are averages across all Medicare discharges. Figure A1 shows that hospitals with a larger $Sim\Delta Re{v}_h$ are more research-active in the pre-BBA era—a finding consistent with the expectation that research-active teaching hospitals are more likely to be disproportionately impacted by the BBA.

Finally, we weigh this measure by the share of a hospital’s discharges that are reimbursed by Medicare in 1995. This follows previous literature which measures the financial impact of Medicare payment changes by using the hospital’s base year Medicare discharge share (Acemoglu and Finkelstein, 2008; Kaestner and Guardado, 2008; Wu and Shen, 2014). The overall BBA Bite is calculated as:

$$BBA\text{\_}Bit{e}_h={\left[\frac{MedicareDischarges}{TotalDischarges}\right]}_{h,1995}\times Sim\Delta Re{v}_{h,1995}$$ (1)

The distribution of BBA Bite among hospitals in the teaching hospital sample is depicted in Fig. 1. The figure shows that there is substantial variation in BBA Bite: the average BBA Bite is 0.0045 with a standard deviation of 0.0035. Looking to the extremes of the distribution provide some illustrative examples: on the far left is South Bay Medical Center. In addition to having a low BBA Bite (0.00003), the California-based hospital has relatively low teaching and research intensity: the institution had three residents and produced one publication in 1997. In contrast, the St. Louis University Hospital — a similarly sized institution with a BBA Bite of 0.0184 — had more than 200 residents and produced 122 publications in the same year.

Fig. 1.

Distribution of BBA Bite.

Notes: This figure shows a histogram of the BBA Bite for the teaching hospital sample, where BBA Bite is the product of (i) the difference between the (log) simulated price per discharge and actual (log) PPS price per discharge in 1995; and (ii) the Medicare share of discharges averaged over 1992–1995. The solid line indicates the mean (0.0045) of this variable and the dotted lines indicates the median (0.0035).

Source: Healthcare provider cost reporting information system; Inpatient prospective payment system payment impact files.

3.3. Summary statistics

Table 1 provides hospital-level summary statistics for the teaching hospital sample. As can be seen in Panel A, the mean hospital has around 16,000 patient discharges, contains 400 beds, has 100 residents and interns, and reported roughly $10 million in combined IME and DSH payments. Approximately a third of patient discharges correspond to Medicare patients. The average Medicare revenue per discharge is $8,000.

Table 1.

Summary statistics.

	Count	Mean	Median	SD	Min	Max
Panel A. Hospital Characteristics
Discharges (1,000s)	780	16.82	14.91	10.39	0.34	62.79
Inpatient Days (1,000s)	780	89.84	76.03	60.25	0.96	439.06
Medicare Teaching Payment ($1,000,000s)	780	5.45	2.26	7.91	0.00	59.81
Medicare Disproportionate Share Payment ($1,000,000s)	780	3.97	2.38	4.54	0.00	36.21
Total Revenue ($1,000,000s)	765	486.94	374.24	410.95	14.43	3,182.46
Beds	780	360.54	319.41	202.23	29.12	1,397.62
Residents and Interns	767	101.92	41.82	139.35	0.06	1,097.72
Pre-1992 Research (Publications)	780	30.34	2.00	93.50	0.00	869.50
BBA Bite (x100)	780	0.45	0.35	0.35	0.00	1.84
Medicare Share of Discharges in 1995	780	0.34	0.34	0.13	0.02	0.71
Medicare Share of Inpatient Days in 1995	780	0.43	0.46	0.15	0.03	0.79
Medicare Price Per Discharge in 1995 ($1,000s)	780	8.30	7.48	2.85	3.67	27.48
Panel B. Grants
Number of Grant Applications
Total	780	8.82	0.00	32.97	0.00	444.00
New	780	7.15	0.00	26.43	0.00	355.75
Competitive Renewal	780	1.67	0.00	6.60	0.00	88.25
Seasoned Investigator	780	3.39	0.00	12.86	0.00	156.94
Novice Investigator	780	4.13	0.00	15.21	0.00	202.62
Incumbent Investigator	780	3.39	0.00	13.23	0.00	165.88
New Hire Investigator	780	3.88	0.00	13.98	0.00	180.88
MD Principal Investigator	780	3.08	0.00	11.87	0.00	158.62
PhD Principal Investigator	780	4.26	0.00	16.22	0.00	193.38
MD-PhD Principal Investigator	780	1.36	0.00	5.81	0.00	87.75
Number of Grants Funded
Total	780	2.55	0.00	10.42	0.00	147.75
New	780	1.82	0.00	7.40	0.00	106.38
Competitive Renewal	780	0.73	0.00	3.05	0.00	41.38
Seasoned Investigator	780	1.07	0.00	4.43	0.00	55.44
Novice Investigator	780	1.13	0.00	4.59	0.00	67.31
Incumbent Investigator	780	1.06	0.00	4.51	0.00	59.75
New Hire Investigator	780	1.06	0.00	4.19	0.00	58.38
MD Principal Investigator	780	0.96	0.00	4.05	0.00	55.00
PhD Principal Investigator	780	1.16	0.00	4.77	0.00	61.00
MD-PhD Principal Investigator	780	0.40	0.00	1.92	0.00	30.88
Panel C. Publications
Total	780	45.40	2.06	148.55	0.00	1,683.62
Citation Ranking: ≤25	780	11.19	0.81	31.02	0.00	306.12
Citation Ranking: 26–50	780	10.41	0.44	32.00	0.00	333.38
Citation Ranking: 51–75	780	11.03	0.38	37.19	0.00	413.19
Citation Ranking: 76–95	780	9.84	0.25	37.25	0.00	456.62
Citation Ranking: 96–99	780	2.27	0.06	9.98	0.00	130.56
Citation Ranking: >99	780	0.67	0.00	3.18	0.00	45.12
Cited In a Patent	780	11.34	0.25	43.40	0.00	547.31
Not Cited In a Patent	780	34.07	1.75	106.05	0.00	1,136.31
Disruptive	780	1.69	0.12	4.86	0.00	51.00
Consolidating	780	40.72	1.75	135.60	0.00	1,551.44
Laboratory Research: Pub. with Bench MeSH	780	12.41	0.06	47.18	0.00	487.50
Translational Research: Pub. with Translational MeSH	780	12.34	0.25	42.86	0.00	516.00
Builds on Translational MeSH	780	3.83	0.25	11.94	0.00	122.25
Inspires Translational MeSH	780	6.40	0.12	23.48	0.00	297.81
Clinical Research: Clinical Trial Pub	780	5.43	0.38	16.94	0.00	179.69
Clinical Research: Pub. with Clinical MeSH	780	11.54	0.94	34.80	0.00	402.31
Panel D. Clinical Outcomes (Risk-Adjusted Survival Rates [30 Days])
Heart Attack	700	0.90	0.91	0.03	0.73	1.00
Heart failure	700	0.96	0.96	0.02	0.89	1.01
Pneumonia	700	0.98	0.98	0.02	0.89	1.04
Hip/knee	700	0.95	0.95	0.02	0.86	1.04

Notes: This table shows hospital characteristics for the hospitals in the teaching hospital sample between 1992–2007. All variables are measured yearly. For example, “Discharges (1000s)” is the average number of patients a hospital receives in a year. “Medicare Teaching Payment,” “Medicare Disproportionate Share Payment,” and “Total Revenue” are adjusted for inflation using the CPI and are measured in 1997 dollars. All hospital characteristics, grant, and publication variables have 780 observations, except for “Total Revenue” and “Residents and Interns,” because of missing data from the Healthcare Provider Cost Reporting Information System. Similarly, clinical outcomes data is only available for the subset of hospitals in the teaching hospital sample that matched to the clinical outcomes data. Clinical outcomes are measured in three-year bins—e.g., hospital-level survival rates in 2005 are estimated over patient claims in 2003, 2004, and 2005. For each hospital, we use three-year bins for four years (1996, 2002, 2005, and 2008). Each publication’s disruptiveness is measured by using the Funk and Owen-Smith (2017) disruptiveness index, denoted by d. “Disruptive” publications refer to publications with d ≤ 95th percentile and “Consolidating” publications refers to publications with d > 95th percentile. Values up to the second decimal place are shown. See Section III for detailed data descriptions. Sources: Azoulay et al. (2020); Chandra et al. (2016); Funk and Owen-Smith (2017); Healthcare Provider Cost Reporting Information System; Inpatient Prospective Payment System Payment Impact Files; NIH IMPAC II; Marx and Fuegi (2020); PubMed; Web of Science.

Looking next to Panel B, the distributions of both the number of grant applications and the number of awarded grants are highly skewed (see Appendix Figure B2). For example, the mean number of grants awarded is roughly three, while one hospital (Massachusetts General Hospital) has nearly 150 grants funded during the period considered in our study. The majority of grant applications and awards correspond to new research proposals, as opposed to competitive renewals of already-funded grants.

Panel C shows that the average hospital had 45 publications in a year, but that this average also masks substantial variation—ranging from a minimum of 0 publications to a maximum of roughly 1,700 (see Appendix Figure B2). Among the publications produced, 4 percent were considered disruptive (as opposed to consolidating), there were similar rates (27 percent) of bench and translational research, and 37 percent were considered clinical.

Across all hospitals, there is a marked increase in the level of grant applications (Appendix Figure B3) and grants funded (Appendix Figure B4) over our study period. The BBA occurred shortly before a sustained increase in the NIH budget, which concluded in 2003 (Korn et al., 2002; Freeman and Van Reenen, 2009). Consequently, most of the teaching hospitals in our sample experienced an absolute increase in the grant funding awarded to their investigators, at least in the immediate aftermath of the shock. Our regression specifications will flexibly control for this secular increase through the inclusion of calendar year effects, making it possible to examine whether more exposed hospitals experience a relative decrease in research activity following the reform, relative to less exposed hospitals.

Appendix Figure B5 shows a similar upward trend for total publications. There is a discrete shift in 1998, which is likely explained by the use of the Web of Science in constructing the publication dataset, as discussed above. A more pronounced upward trend is observed for translational research and “other” clinical research articles, whereas the trajectories for laboratory research and clinical trials are relatively flat.²⁰

Appendix Table D1 compares the teaching hospital sample and the AMC sample. Relative to hospitals in the analysis hospital file, AMC sample hospitals produce significantly more grant applications and publications.

For a subset of the empirical exercises that follow, we examine the impact of hospital financing cuts on clinical outcomes in hospitals. Our main measures of clinical outcomes stem from Chandra et al. (2016) analysis of the relationship between hospital quality and market size. In particular, these authors construct hospital-level measures of 30-day risk-adjusted survival for four frequent conditions: acute myocardial infarction (AMI), congestive heart failure, pneumonia, and hip/knee replacements (a common pair of surgical procedures). Specifically, they construct condition-specific measures using claims and survival outcomes in three-year rolling windows.²¹ Since the time period in the Chandra et al. (2016) study spans the pre- and post-BBA periods, for each hospital, we use the observation in 1996 (corresponding to an average over the pre-BBA years 1994, 1995, and 1996) and the observations (corresponding to an average over the post-BBA years 2000 through 2005). Of the 780 hospitals in the teaching hospital sample, we successfully match 700 hospitals to the clinical outcomes dataset. Table 1, Panel D provides summary statistics for these survival outcomes, while Appendix Figure B6 presents histograms for the distribution of the survival rates for each of the conditions.

4. Empirical strategy and results

Our primary empirical approach is rooted an extensive literature that examines the impact of changes in public funding (e.g., Duggan and Morton, 2006, 2010; Blume-Kohout and Sood, 2013). This methodology enables us to isolate the effect of the BBA by examining variation across hospitals in their expected exposure to the reform. We expect changes in research outcomes (e.g., increases due to hiring or within-physician substitution towards research) to be more pronounced in hospitals that faced a large decrease in the level of Medicare reimbursements (relatively high BBA Bite) relative to those that were minimally impacted by the reform (relatively low BBA Bite):

$$y_{ht}=\beta BB{A}_{-}Bit{e}_h\times PostBB{A}_t+{\delta }_h+{\zeta }_t+{\epsilon }_{ht}$$ (2)

where $y_{ht}$ is a measure of research outcome for hospital $h$ and year $t$, and $AfterBB{A}_t$ is an indicator equal to 1 after 1997.²² As indicated in our prior description of the BBA Bite, a key benefit of this methodology is that it relies on pre-BBA hospital characteristics to measure exposure to the reform. This enables us to exclude endogenous hospital responses (e.g., changes in the number of patients or residents) that could influence the estimated impact of the reform and therefore our estimates.

We account for other factors that are likely to influence research activity within hospitals. For example, we control for time-invariant differences across hospitals (e.g., location, management style) using hospital fixed effects. We also include year fixed effects $\left({\zeta }_t\right)$ to account for secular changes, such as technological advancements and new scientific discoveries. As described in our conceptual framework, BBA-driven changes in research outcomes are more likely to occur within more treated hospitals and more likely to be hospital-year specific.

One threat to identification is that our main results may biased by the doubling of the NIH budget between 1998 and 2003: research-intensive hospitals may have been more likely to benefit from the NIH budget increase (e.g., because they were most prepared to increase research production) and may have been more exposed to the BBA.²³ To address these concerns, we include a proxy for hospital research intensity (RI): the average annual number of publications between 1990–1991 divided by 1000.²⁴ The distribution of RI among hospitals in the teaching hospital sample is depicted in Fig. 2. High RI hospitals have a higher BBA Bite (see Appendix Figure B7). We also find that hospitals with higher RI receive more indirect medical education and disproportionate share subsidies, have more residents, and produce more research (as measured by grant applications, grants funded, and publications) prior to the BBA (see Appendix Table B2).

Fig. 2.

Distribution of research intensity.

Notes: This figure shows a histogram of the Research Intensity for the teaching hospital sample, where Research Intensity is the annual number of publications averaged over 1990–1991 and divided by 1000. The solid line indicates the mean (0.015) of this variable and the dotted lines indicates the median (0.002). For ease of interpretation in this figure, hospitals with Research Intensity greater than 0.1 are recoded as 0.1.

Source: Healthcare provider cost reporting information system; Inpatient prospective payment system payment impact files.

Using this proxy for hospital research intensity, we modify Eq. (2) to directly estimate and control for RI hospitals’ response to the NIH budget increase. Our main specification becomes:

$$y_{ht}=\beta BB{A}_{-}Bit{e}_h\times PostBB{A}_t+\gamma R{I}_h\times PostNIHDoublin{g}_t+{\delta }_h+{\zeta }_t+{\epsilon }_{ht}$$ (3)

where $R{I}_h$ denotes the average number of publications between 1990–1991, $PostNIHDoublin{g}_t$ is an indicator equal to 1 after 1998, and all remaining variables are similar to those found in Eq. (2).

Table 2 examines the effect of the BBA when RI controls are excluded (odd columns) and included (even columns). To ease interpretation of the magnitudes, we present elasticities in the fifth and six rows. Column 1 shows that the BBA has a significant and positive effect on grant applications. Column 2 reveals that these effects, though smaller in magnitude, remain significant and positive even after controlling for the doubling of the NIH budget: in particular, following the BBA’s enactment, the total number of grant applications increases by 0.026 percent for each percent increase in BBA Bite. This is tantamount to an approximately 6 percent increase in grant applications among hospitals whose BBA Bite is above the median (see Appendix Table B3).

Table 2.

Effect of BBA and NIH budget doubling.

	Grant applications		Grants funded		Publications
	(1)	(2)	(3)	(4)	(5)	(6)
BBA Bite ×Post BBA	19.065*** (4.284)	9.338** (4.497)	15.545*** (3.325)	7.170** (3.430)	6.148 (5.380)	1.510 (5.893)
RI ×Post NIH Doubling		1.009*** (0.207)		0.869*** (0.209)		0.481*** (0.120)
Elasticity (BBA Bite)	0.053	0.026	0.047	0.021	0.017	0.004
Elasticity (RI)		0.017		0.016		0.008
Adjusted R2	0.023	0.039	0.018	0.038	0.029	0.030
Mean of Outcome	0.751	0.751	0.439	0.439	1.991	1.991
Nb. Observations	12,480	12,480	12,480	12,480	12,480	12,480
Nb. Hospitals	780	780	780	780	780	780

Notes: This table reports difference-in-differences estimates of the effect of the BBA and NIH budget doubling on the number of grant applications, grants funded, and publications, in the teaching hospital sample. Observations are at the hospital-year level. Outcomes are transformed with the inverse hyperbolic sine function. BBA Bite is a hospital’s BBA Bite and RI is a hospital’s research intensity (average number of annual publications between 1990–1991 divided by 1000). Estimates are from OLS regressions, and each coefficient is from a separate regression. Robust standard errors (clustered at the hospital level) are in parentheses. Sources: Healthcare Provider Cost Reporting Information System; Inpatient Prospective Payment System Payment Impact Files; NIH IMPAC II; PubMed, Web of Science.

^*p < 0.10,

^**p < 0.05,

^***p < 0.001.

Event studies that explore the timing of these estimated effects support these findings. In particular, we estimate:

$$\begin{gathered} Y_{ht}=\alpha +\sum _z{\beta }_z\times 1(z)\times Hig{h}_{-}BB{A}_{-}Bit{e}_h+ \\ \sum _w{\gamma }_w\times 1(w)\times Hig{h}_{-}R{I}_h+{\delta }_h+{\tau }_t+{ϵ}_{ht} \end{gathered}$$ (4)

where $Hig{h}_{-}BB{A}_{-}Bit{e}_h$ and $Hig{h}_{-}R{I}_h$ are indicators for whether a hospital $h$’s BBA Bite and RI, respectively, are above the median, ${\delta }_h$ represents hospital fixed effects, and ${\tau }_t$ represents year fixed effects. $z$ represents the “lag”, or the years relative 1997, which is the year in which the BBA is implemented. $w$ represents the years relative to 1998, the year in which the NIH budget doubling begins.

Fig. 3 plots the estimates of ${\beta }_z$ from Eq. (4) for total grant applications and corresponds to a dynamic version of Table B3, Columns 1 and 2. The figure also displays 95-percent confidence intervals and a dashed gray line that represents the year in which the BBA was enacted. The estimated coefficients illustrate that hospitals with high BBA Bites exhibit trends in grant applications similar to hospitals with low BBA Bites in years prior to the BBA. Further, using the methodology outlined in Roth (2021), we find little evidence that differential pre-trends (that we might be underpowered to detect) among high BBA Bite and low BBA Bite hospitals would substantially bias our results.²⁵ Notably, the level of grant applications increases differentially in the wake of the BBA and continues to increase afterward. The increase in total grant applications is gradual, further providing evidence that these aggregate effects may mask significant heterogeneity in the response across different types of grants. In subsequent analyses, we provide additional evidence suggesting that hospitals respond to Medicare payment cuts through both the intensive margin (i.e., within-physician substitution towards research activities funded by non-Medicare sources) and the extensive margin (i.e., shifting their hiring practices).

Fig. 3.

Effect on the number of grant applications.

Notes: This figure plots the response in grant applications in the teaching hospital sample. Each dot corresponds to OLS coefficient estimates stemming from the event study specification in which asinh(number of grant applications) is regressed onto year fixed effects, hospital fixed effects, as well as interaction terms between an indicator for a hospital’s High BBA Bite status interacted with the number of years before/after the BBA. Panel A plots estimates from a specification that excludes indicators for a hospital’s High RI status interacted with the number of years before/after the start of the NIH budget doubling. Panel B plots estimates from the specification described in Eq. (4) and includes indicators for a hospital’s High RI status interacted with the number of years before/after the start of the NIH budget doubling. The 95% confidence interval (corresponding to robust standard errors, clustered around hospital) around these estimates is shown.

Source: Healthcare provider cost reporting information system; Inpatient prospective payment system payment impact files; NIH IMPAC II.

Next, we examine the impact on the number of grants funded. We find similar results. Table 2, Column 2 and Figure B8 show that following the BBA, a percent increase in a hospital’s BBA Bite is associated with an approximate 0.021 percent increase in the number of grants funded. This corresponds to an approximately 5 percent increase in grants funded among hospitals whose BBA Bite is above the median (see Appendix Table B3).

Turning to the effect on total publications, Table 2, Column 3 shows that the BBA has economically small and insignificant effects on the total number of subsequent publications. However, a challenge of only examining research aggregates is that AMCs and researchers may alter the nature of their research projects in response to financial strain, in which case examining hospital-level grant and publications total counts may mask how institutions and individuals reallocate their efforts among research projects with varying types and differing levels of impact. In the following sections, we turn to a more detailed investigation into how a hospital’s research funding levels, impact, and compositional mix shift in response to the BBA. All the specifications include the research intensity controls as in Eqs. (3) and (4), though the corresponding estimates are not reported.

4.1. Effects on research funding levels

Table 3 presents the estimates of the impact of the BBA on detailed measures of research funding levels, as measured by grant applications. In particular, we examine the impact of cuts in Medicare reimbursement on different types of grants (by grant cycle and principal investigator type). The estimates presented in Column 2 suggest that the positive response in total grant applications is driven by an increase in applications for new proposals. In contrast, there is no statistically significant effect on the number of “competing continuations”, which correspond to applications seeking to extend previously awarded grants for an additional funding cycle.

Table 3.

Effect on the number of grant applications.

		Grant cycle		Principal investigator experience				Principal investigator degree
	Total (1)	New (2)	Renewal (3)	Seasoned (4)	Novice (5)	Incumbent (6)	New Hire (7)	MD (8)	PhD (9)	MD-PhD (10)
BBA Bite ×Post BBA	9.338** (4.497)	13.13** (4.604)	−0.505 (2.841)	20.89*** (4.018)	8.714** (3.942)	21.78*** (4.945)	13.09** (4.047)	8.686** (4.424)	11.79** (4.178)	14.60*** (3.789)
Elasticity	0.026	0.037	−0.002	0.061	0.025	0.063	0.038	0.025	0.034	0.050
Adjusted R2	0.039	0.057	0.010	0.111	0.035	0.132	0.043	0.033	0.068	0.101
Diff. Wald test p-value (Coefficient)		0.001		0.004		0.121			0.509	0.237
Diff. Wald test p-value (Elasticity)		0.001		0.003		0.115			0.542	0.113
Mean of Outcome	0.751	0.705	0.372	0.479	0.574	0.466	0.559	0.519	0.533	0.328
Nb. Observations	12,480	12,480	12,480	12,480	12,480	12,480	12,480	12,480	12,480	12,480
Nb. Hospitals	780	780	780	780	780	780	780	780	780	780

Notes: This table reports difference-in-differences estimates of the effect of the BBA on the number of grant applications, in the teaching hospital sample. Observations are at the hospital-year level. Outcomes are transformed with the inverse hyperbolic sine function. Column 1 presents results from OLS regressions. As shown in equation (3), regressions include hospital fixed effects, year fixed effects, and RI ×Post NIH Doubling, where RI is the average number of publications between 1990–1991 divided by 1000. Columns 3–10 present estimates from seemingly unrelated regressions. Each coefficient is from a separate regression. The elasticity of 0.026 implies that a 1 percentage increase in BBA Bite is associated on average with a 0.026 percent yearly increase in grant applications following the BBA’s enactment. The fifth row shows p-values from Wald tests that compare coefficients on BBA Bite ×Post BBA across different columns and the sixth row shows p-values from Wald tests that compare elasticities across different columns (Column 2 vs. Column 3; Column 4 vs. Column 5; Column 6 vs. Column 7; Column 8 vs. Column 9; Column 8 vs. Column 10). Robust standard errors (clustered at the hospital level) are in parentheses. Sources: Healthcare Provider Cost Reporting Information System; Inpatient Prospective Payment System Payment Impact Files; NIH IMPAC II.

^*p < 0.10,

^**p < 0.05,

^***p < 0.001.

Columns 4 and 5 show that the effect on total grant applications is driven primarily by an increase in those proposed by seasoned (rather than novice) investigators. Event studies echo these findings: grant applications rise more quickly and strongly among seasoned investigators (see Appendix Figure B10). Columns 6 and 7 indicate that the increase in total grant applications is driven by both incumbents and new hires (the coefficients indicate a larger increase among incumbents relative to new hires, but the difference is not statistically significant). Taken together, these estimates are consistent with the view that AMCs may respond to Medicare payment cuts along both the intensive and extensive margins: along the intensive margin, by encouraging incumbent investigators to apply for new grants; along the extensive margin, by hiring additional, seasoned researchers who may bring grants with them.

Estimates in Columns 8 through 10 explore the effects of the financing cuts on research conducted by MDs, PhDs, and MD-PhDs. In all cases, we estimate a strong, positive, and statistically significant effect of the BBA Bite on the level of subsequent grant applications. The PhD results are slightly larger in magnitude than the MD results: a 1 percent increase in BBA Bite yields a 0.034 percent increase in research conducted by PhD-holding principal investigators. In contrast, research conducted by MDs increases by 0.025 percent (though the difference between PhD and MD coefficients is not statistically significant). This accords with our priors: PhD-holding investigators are more likely to engage in research at the laboratory bench, and MDs are traditionally associated with clinical and translational research. The finding that research conducted by MDs significantly increases can be explained by the fact that NIH grantees holding MD or MD-PhD degrees increasingly share similar scientific interests and methodological approaches with investigators who received traditional PhDs in the various disciplines within the life sciences domain (Ali and Gittelman, 2016).

Appendix Table B4 demonstrate that similar patterns for grants funded persist when looking across grant cycle (new applications versus competing continuations) and principal investigator degree type.

4.2. Effects on research impact

In addition to shifting the level of subsequent research activities, an additional dimension through which hospitals may have responded to the BBA is by shifting the quality of these activities. Table 4 report the effects of the shock on publication rates adjusted for three measures of impact: publication-to-publication citation impact (Columns 2–7), patent-to-publication citation impact (Columns 8–9), and “disruptiveness” (Columns 10–11).

Table 4.

Effect on publication impact.

	Total	By citation percentile						Cited in a patent		By disruption
		≤25	26–50	51–75	76–95	96–99	>99	Yes	No	Consolidating research	Disruptive research
	(1)	(2)	(3)	(4)	(5)	(6)	(7)	(8)	(9)	(10)	(11)
BBA Bite ×Post BBA	9.271*	7.742	7.947*	11.655**	13.623***	11.630***	6.170	3.654	8.353	9.737**
	(5.889)	(4.914)	(4.723)	(4.503)	(4.928)	(4.081)	(2.883)	(4.798)	(5.721)	(6.713)	(4.409)
Elasticity	0.004	0.026	0.022	0.022	0.033	0.041	0.058	0.017	0.010	0.023	0.031
Adjusted R2	0.030	0.030	0.024	0.022	0.031	0.042	0.048	0.013	0.036	0.027	0.014
Diff. Wald test p-value (Coefficient)		0.012	0.048	0.089	0.008	0.021	0.072	0.577		0.784
Diff. Wald test p-value (Elasticity)		0.012	0.047	0.087	0.008	0.012	0.002	0.574		0.563
Mean of Outcome	1.991	1.347	1.153	1.089	0.962	0.469	0.232	0.976	1.876	1.822	0.800
Nb. Observations	12,480	12,480	12,480	12,480	12,480	12,480	12,480	12,480	12,480	12,480	12,480
Nb. Hospitals	780	780	780	780	780	780	780	780	780	780	780

Notes: This table reports difference-in-differences estimates of the effect of the BBA on the number of publications, by impact, in the teaching hospital sample. Observations are at the hospital-year level. Outcomes are transformed with the inverse hyperbolic sine function. Estimates are from seemingly unrelated regressions and each coefficient is from a separate regression. As shown in equation (3), regressions include hospital fixed effects, year fixed effects, and RI ×Post NIH Doubling, where RI is the average number of publications between 1990–1991 divided by 1000. Each publication’s disruptiveness is measured by using the Funk and Owen-Smith (2017) disruptiveness index, denoted by d. Column 8 refers to publications with d≤95th percentile and Column 9 refers to publications with d > 95th percentile. The elasticity of 0.004 implies that a 1 percentage increase in BBA Bite is associated on average with a 0.004 percent yearly increase in publications following the BBA’s enactment. The fifth row shows p-values from Wald tests that compare coefficients on BBA Bite ×Post BBA across different columns and the sixth row shows p-values from Wald tests that compare elasticities across different columns. For example, for Columns 2 to 7, the fifth row compares estimates in Column 1 vs. estimates in the focal column. For Columns 8 and 9, the fifth row shows p-values from Wald tests that compare Column 8 and 9 estimates. For Columns 10 and 11, the fifth row shows p-values from Wald tests that compare Column 10 and Column 11. Standard errors are in parentheses, and are clustered at the hospital level. Sources: Healthcare Provider Cost Reporting Information System; Inpatient Prospective Payment System Payment Impact Files; Funk and Owen-Smith (2017); PubMed, Web of Science.

^*p < 0.10,

^**p < 0.05,

^***p < 0.001

Turning first to the effect on publication-to-publication citation impact and patent-to-publication citation impact, much of the scholarship in the economics of innovation implicitly treats citation impact as a proxy for impact or “importance” (Hall et al., 2005; Azoulay et al., 2011), although this assumption has come under scrutiny (Funk and Owen-Smith, 2017). If the decrease in publication volume documented above was driven by articles which do not make much of a ripple, leaving the impact on highly-cited articles unchanged, we might be tempted to conclude that the effects of the reform on research activities are at most second-order, relative to the effects on the cost and quality of health care delivery in these same institutions.

As described in Section 3, publications are assigned to six mutually exclusive citation-based bins. We run the regression specified in Eq. (2) using each bin as a separate outcome. Looking across Columns 2 through 7, we cannot reject that the implied elasticities are essentially identical. We next examine the effect of the BBA on the number of publications that are subsequently cited by a patent, which we view as a proxy for the extent to which the work of AMC scientists shapes biopharmaceutical firms’ R&D efforts. Columns 8 and 9 reveal similar effects across the number of publications that are and are not, respectively, cited by a patent. Overall, the evidence in Table 4 indicates that the effect of the reform are relatively homogeneous along the “citation as quality” dimension.

The remaining columns of Table 4 assess the extent to which the BBA disproportionately affects “disruptive” research. Using the $d$ index proposed by Funk and Owen-Smith (2017) to partition the set of articles for each hospital into “disruptive” publications ($d$ in the top ventile) and “consolidating” publications ($d\le 95\text{th}$ percentile). We find a slightly larger positive effect on disruptive research (Column 11) relative to consolidating research (Column 10), though the difference between the coefficients is not statistically significant. Taken together, the results suggest that the BBA ushers in an uniform increase in the level of both low- and high-impact scientific projects.

4.3. Effects on the composition of the research portfolio

Given the unique research capabilities and priorities of AMCs, a natural next question is how the BBA shaped the composition of hospital research. Indeed, event study evidence hints that research aggregates may be masking compositional changes: Appendix Figure B9 shows that in the years immediately following the reform, the BBA has a significant and positive effect on the number of publications. This trend may have resulted from two factors: scientists might have shifted more of their time to research, leading to the publication of studies that were previously set aside. Additionally, efforts by AMCs to secure grant funding (e.g., increased efforts to hire productive researchers attracted to AMCs) likely resulted in a higher number of publications.

In Table 5, we explore the possibility of compositional changes by examining the effect on laboratory research, translational, and clinical research. The results indicate that basic laboratory research is largely unaffected by the BBA: while a 1 percent increase in BBA Bite is associated with a 0.017 percent decrease in publications of this type. However, the estimates are statistically indistinguishable from 0, likely due to the fact that our analysis is focused on hospitals that are primarily focused on translational and clinical research. Consistent with this view, we observe a quantitatively and significantly more positive response for the level of translational research: the BBA causes translational research publications to increase by 0.024 percent for each percent increase in BBA Bite. This amplifying effect also extends to publications that “build on translational research” (Column 3) or “inspire translational research” (Column 4), both of which increase significantly more than translational research publications.

Table 5.

Effect on publication composition.

	Laboratory research	Translational research			Clinical research
	Bench MeSH	Translational MeSH	Builds on translational MeSH	Inspiring translational MeSH	Clinical trials	Other
	(1)	(2)	(3)	(4)	(5)	(6)
BBA Bite ×Post BBA	−6.146 (5.206)	8.618* (4.755)	22.493*** (4.357)	13.859** (4.466)	24.496*** (4.650)	10.560** (5.087)
Elasticity	−0.017	0.024	0.065	0.039	0.069	0.030
Adjusted R2	0.001	0.020	0.080	0.032	0.098	0.027
Mean of Outcome	0.826	1.062	0.798	0.804	0.950	1.398
Nb. Observations	12,480	12,480	12,480	12,480	12,480	12,480
Nb. Hospitals	780	780	780	780	780	780

Notes: This table reports difference-in-differences estimates of the effect of the BBA on bench, translational, and clinical trial research in hospitals, in the teaching hospital sample. Observations are at the hospital-year level. Outcomes are transformed with the inverse hyperbolic sine function. Estimates are from OLS regressions and each coefficient is from a separate regression. As shown in equation (3), regressions include hospital fixed effects, year fixed effects, and RI ×Post NIH Doubling, where RI is the average number of publications between 1990–1991 divided by 1000. Column 1 refers to publications that are not disease-oriented, are not clinical trial publications, and rely on either a molecular biology technique, a model organism, cellular structures and macromolecules, or biochemical and cellular processes (based on MeSH terms). Column 2 refers to publications that are disease-oriented, are not clinical trial publications, and rely on either a molecular biology technique, a model organism, cellular structures and macromolecules, or biochemical and cellular processes (based on MeSH terms). Column 3 refers to publications that report the results of a clinical trials, or are tagged by a human MeSH term and also cite a translational publication. Column 4 refers to publications that are translational and is cited by a clinical trial publication (or one that contains a human MeSH term). Column 5 refers to publications that are indicated as clinical trials based on MeSH terms or the publication type field in PubMed. Finally, Column 6 refers to publications that are disease-oriented, not clinical trials, and not tagged by any bench MeSH keywords. The elasticity of −0.017 implies that a 1 percentage increase in BBA Bite is associated on average with a 0.017 percent yearly decrease in publications following the BBA’s enactment. Standard errors are in parentheses, and are clustered at the hospital level. Sources: Healthcare Provider Cost Reporting Information System; Inpatient Prospective Payment System Payment Impact Files; Azoulay et al. (2020); Marx and Fuegi (2020); PubMed; Web of Science.

^*p < 0.10,

^**p < 0.05,

^***p < 0.001.

Looking next to clinical research publications, we find that the level of clinical research — as measured using the clinical trial definition (Column 5) and “other” clinical MeSH-based (Column 6) definition — increases substantially: a 1 percent increase in BBA Bite is associated with a 0.069 percent increase in clinical trial publications and 0.030 percent increase in “other” clinical research. As with translational research, the difference between changes in basic laboratory research and clinical research is statistically significant.

In sum, the results presented in Table 5 paint a nuanced picture of compositional change in response to the health care financing cuts. The financing reform leaves bench research relatively unchanged, while triggering a sizable increase in investigations that can plausibly be labeled “translational” or “clinical”. These results are consistent with the argument that the BBA led hospitals do respond along the extensive margin (i.e., through shifts in hiring translational and clinical researchers) and that any intensive margin substitution effects likely involve physician substitution of patient care or research activity towards “bedside” research—i.e., clinical trials.²⁶

4.4. Countervailing impact on clinical care

Our results so far are consistent with the view that financial pressures might lead these scientists to allocate their time and effort away from patient care activities, towards research. As a result, the increase in subsequent research activity may be countered by declines in the quality of clinical care. A handful of previous studies have found that AMCs are associated with better outcomes (Keeler et al., 1992; Taylor et al., 1999; Ayanian and Weissman, 2002; Burke et al., 2017). In this section, we explore whether the BBA dampened these clinical care advantages by impacting patient outcomes within the hospitals in our sample.

We estimate the impact of the BBA on risk-adjusted survival rates in the “long difference” dimension of the data:

$$\Delta Y_h^c=\beta BB{A}_{-}Bit{e}_h+Discharge{s}_h+{\epsilon }_h$$ (5)

where $\Delta Y_h^c$ corresponds to the change in the risk-adjusted 30-day mortality rate in hospital $h$ for condition $c$. Acting as a proxy for hospital size, $Discharge{s}_h$ is the log total number of patient discharges in 1995.

Table 6 reports the results. While we find no evidence that the BBA significantly impacts patient survival and survival at 30-days is a crude measure of care quality, the confidence intervals are large enough that we cannot rule out substantial changes in survival. For example, we cannot rule out increases in patient survival rates following a heart attack by 0.4 or 44 percent relative to the mean when a hospital’s BBA Bite increases from 0 to 1. At the same time, it is possible that the BBA could have caused survival rates to decline by 1.26 (or 140 percent) for each unit change in BBA Bite. Nevertheless, it seems notable that we were unable to detect consistent declines in patient outcomes concurrently with the increase in research activity experienced by hospitals more exposed to the reform and our findings are consistent with a number of studies indicating that the BBA did not have a substantial effect on patient outcomes among AMCs and non-AMCs (McCall et al., 2003; Seshamani et al., 2006a,b).²⁷

Table 6.

Effect on clinical outcomes.

(Changes in Risk-adjusted Survival Rates).

	Heart attack (1)	Heart failure (2)	Hip/Knee (3)	Pneumonia (4)
BBA Bite	−0.4286 (0.4227)	−0.4028 (0.2537)	0.4099 (0.2783)	−0.0743 (0.3146)
Ln(Discharges in 1995)	0.0016 (0.0025)	0.0018 (0.0015)	−0.0003 (0.0014)	0.0046** (0.0019)
Adjusted R2	−0.0012	0.0016	0.0004	0.0076
Mean of Outcome	0.0270	0.0106	−0.0005	0.0147
Nb. Observations	700	700	700	700

Notes: This table displays the effect on changes in risk-adjusted survival rates among hospitals in the teaching hospital sample. The hospital sample used is the subset of the teaching hospital sample that is matched to the clinical outcomes dataset from Chandra et al. (2016). Observations are at the hospital-level. Outcomes are the difference in average survival rates between the post-BBA time period and the pre-BBA time period. Estimates are from OLS regressions and each coefficient is from a separate regression. Standard errors are in parentheses and robust. Sources: Healthcare Provider Cost Reporting Information System; Inpatient Prospective Payment System Payment Impact Files; Chandra et al. (2016).

^*p < 0. 10,

^**p < 0.05,

^***p < 0.001.

4.5. Robustness checks

AMC Sample Analyses.—

In Appendix D, we perform our analysis on the AMC sample, a subset of hospitals with high research and teaching intensity. This analysis reveals results that are fully consistent with our main results. In many cases, the estimated magnitudes are even greater than those found in the main teaching hospital sample, though in some cases, the estimates are not significant, largely due to the smaller sample size. For example, among research-intensive hospitals, a 1 percent increase in a hospital’s BBA Bite translates into a 0.031 percent increase in the total number of subsequent grant applications (Appendix Table D2) and a 0.037 percent increase in the total number of grants funded (Appendix Table D3). We find a similar increase in the number of publications related to translational and clinical research (Appendix Table D5). As a robustness check, we confirmed that our results in the primary teaching hospital sample are not being driven by hospitals not found in the AMC sample.

Age Neutrality of the NIH Budget Doubling.—

In our main specification, we address concerns that the NIH budget doubling may bias our results by including controls for hospital research intensity. The existing evidence demonstrates that the positive and significant impact of the BBA on hospital research investments persists. However, a natural concern is that the increase in the NIH budget may still disproportionately affect hospitals that were more exposed to the BBA through a different channel—e.g., by targeting the type of research produced by hospitals most exposed to the reform. To investigate these concerns, we explore whether the doubling of the NIH budget was “age-biased”. Empirically, we examine whether NIH budget increases disproportionately benefited research on health conditions affecting elderly populations. Medicare-eligible individuals are almost exclusively above 65 years old. Age-bias in the NIH research portfolio could advantage high-BBA Bite hospitals since their investigators have privileged access to patient populations relevant for the study of these conditions. Appendix Table E1 shows that funding appears to have flown slightly more towards research on diseases that targeted the elderly (vs. those that targeted the non-elderly), but the differences are small in magnitude and not statistically significant.

Subsidy BBA Bite Analysis.—

In Appendix F, we examine the robustness of the core results previously presented in Section 4 to an alternative measure of BBA Bite. Recall that hospital-level variation is driven by differences in the hospitals’ reliance on Medicare payments and in particular, Medicare subsidy payments. In this robustness check, we measure of the BBA Bite as the share of 1995 hospital patient revenues that come from IME and DSH payments (Appendix Figure F1 graphs the distribution of this version of the BBA Bite). The regression results in Appendix Figures F1 through F4 are quite similar to those in Tables 3 through 5, both in terms of the magnitude of the elasticities and statistical significance. The corresponding event-study graphs, displayed in Appendix Figures F3 through F5 are similar to our core set of results. As described in Section 3.B, we prefer measuring BBA Bite as the difference in simulated revenue per discharge and actual PPS revenue per discharge, weighed by the Medicare share of total discharges, as it allows us to more directly estimate the average hospital income loss as a result of the BBA.

Instrumental Variables Estimation.—

Appendix G considers an alternative strategy for capturing the exogenous variation in the PPS reimbursement formula. Following Shen (2003), we construct a simulated post-BBA revenue per discharge change that we use to instrument the actual post-BBA revenue per discharge change in a series of “long-difference” regressions. The difference in simulated and actual post-BBA revenue per discharge changes is documented in Appendix Figure G1. Appendix Table G1 reports the two-stage least square regression results for our primary research outcomes. The results are consistent with our core results: hospitals that experience a larger financial loss as a result of the BBA experience an increase in research outcomes. The effects are both seen in the short-run (Panel A) and long-run (Panel B), though the effects are only statistically significant for publications in the long-run.

Intended Effects of the BBA.—

Our primary analysis focuses on simulating the revenue per discharge that hospitals would have lost under the BBA, had the implemented changes been in effect in 1995. Notably, we focus on implemented changes to the Medicare payments, not the intended effects. However, as noted above, several post-BBA reforms were enacted to provide relief to hospitals, ultimately creating a wedge between the intended and implemented Medicare payment changes. To illustrate, DSH payments were reduced by 2 percent in 2001, rather than the 4 percent reduction as outlined in the BBA. We probe the robustness of our main results by simulating the revenue that hospitals would have lost under the BBA, had the intended effects of the reform been in effect earlier. The results in Appendix H are similar to the main results, both in terms of the magnitude of the elasticities and statistical significance.

5. Conclusion

We investigate how Medicare financing cuts shape institutions that play a central role in the biomedical research ecosystem: Academic Medical Centers. Using a differences-in-difference approach, we find that research activities, measured using a variety of metrics, increase in response to a reform that decreases clinical care margins. Importantly, we document evidence that these effects do not cut evenly along the vertical chain of biomedical research. Laboratory-based “bench” research appears relatively unaffected. On the opposite end of the spectrum, clinical research activity (including industry-sponsored clinical trials) increases markedly in more exposed institutions, relative to less exposed ones. This is also the case for translational research, a type of “basic patient-oriented research” (Ahrens, 1992) that is both hard to appropriate and difficult to perform outside of the AMC setting. Notably, we fail to find much evidence of countervailing impacts on the quality of care within these institutions.

Our evidence is consistent with a mechanism where in response to a funding decline, AMCs increased research by shifting their hiring practices and substituting away from patient care activities towards clinical research activities. Our results support the view that researchers (or at least their employers) can “induce demand” in a manner similar to physicians shifting their practices’ emphasis away from Medicare beneficiaries onto patients covered by private insurance (He and Mellor, 2012).

For the past 30 years, McGuire and Pauly (1991) model of physician behavior has provided health economists with a conceptual framework to analyze the effect of price changes for medical care in the presence of multiple payers. Our results can be interpreted through that lens: they are consistent with the substitution effect dominating the income effect, leading researchers to increase their research activities. However, this interpretation does not do justice to the subtleties of the academic medical setting. For example, in a competitive grant system (even one with rising paylines as in the immediate aftermath of the BBA passage), it is difficult for researchers to “induce demand” in a manner similar to a doctor treating Medicare patients or those covered by private insurance. Research activities are subject to lumpy adjustment costs and make the allocation of time less fractional than in the clinical care setting: a physician who stepped off the research funding treadmill might find herself unable to “prime the pump” of preliminary research results necessary to make her NIH grant proposals viable. An additional friction in the substitution process, and one regularly emphasized by academic medical leaders, is that lower Medicare reimbursements hampered AMCs’ ability to cross-subsidize research activities. And yet, the combination of soft money appointments and relatively generous indirect cost recovery rates creates a channel for teaching hospitals and AMCs to quickly adjust to the financing shock—through the creation of new academic positions.

Our study suffers from limitations and suggests opportunities for future research. First, we analyze the effect of the BBA at the hospital-level. This is sensible given the hospital-level nature of the financing shock, but leaves us unable to directly examine how and to what extent individual investigators can substitute between patient care and different types of research activities. Yet, our decomposition of the effect on grant applications between seasoned and novice investigators points to a lack of response by inexperienced researchers. This finding echoes perennial concerns regarding the “greying” of the NIH investigator population (Azoulay et al., 2013). These concerns might be particularly salient in the context of AMCs, where newly-minted physician-scientists could face stronger barriers to set themselves on the path towards research independence, relative to their colleagues working exclusively at the laboratory bench (Rosenberg, 1999). Policymakers seeking to relieve these frictions could encourage funders to decrease the importance of preliminary results in the grant application process, possibly extending to a new class of grant mechanisms where investigators, especially novices, face no or little burden of initial proof for the ideas they propose.

Second, our findings on AMCs, coupled with recent work on universities (Babina et al., 2023), suggest the resilience of research institutions in response to financing shocks. However, we are cautious in extrapolating our results to other settings: we can only speculate on the effects of other targeted or broad-based funding shocks, such as a drop in philanthropic donations or endowment losses (Dranove et al., 2017). Our sense is that such shocks might induce hospitals to rely on alternative funding sources, but may not necessarily lead to the uneven effects we uncover. Clarifying the conditions under which research institutions respond to different types of funding shocks is an important area for future research.

As governments endeavor to enact funding policies to reduce inefficiencies within the health care system, there have been concerns about negative implications for institutions that occupy a central role in the biomedical research ecosystem. Our results demonstrate the importance for a more comprehensive and systematic analysis of the trade-offs in an institutional environment where seemingly disparate activities are tightly linked. Notably, a decline in public sector financing (from Medicare reimbursements) led to increased reliance on private sector funding (through clinical trials), which is associated with its own set of opportunities and constraints. Attending to the innovation impacts of well-intentioned funding policies is essential to ensure that new policies will, in fact, improve patient and societal welfare.

Appendix A. Supplementary data

Supplementary material related to this article can be found online at https://doi.org/10.1016/j.respol.2024.105163.

Data availability

The authors do not have permission to share data.