Quality and Cost of Care by Hospital Teaching Status: What Are the Differences?

Abstract

Policy Points.

• In two respects, quality of care tends to be higher at major teaching hospitals: process of care and long‐term survival of cancer patients following initial diagnosis. There is also evidence that short‐term (30‐day) mortality is lower on average at such hospitals, although the quality of evidence is somewhat lower.

• Quality of care is mulitdimensional. Empirical evidence by teaching status on dimensions other than survival is mixed.

• Higher Medicare payments for care provided by major teaching hospitals are partially offset by lower payments to nonhospital providers. Nevertheless, the payment differences between major teaching and nonteaching hospitals for hospital stays, especially for complex cases, potentially increase prices other insurers pay for hospital care.

Context

The relative performance of teaching hospitals has been discussed for decades. For private and public insurers with provider networks, an issue is whether having a major teaching hospital in the network is a “must.” For traditional fee‐for‐service Medicare, there is an issue of adequacy of payment of hospitals with various attributes, including graduate medical education (GME) provision. Much empirical evidence on relative quality and cost has been published. This paper aims to (1) evaluate empirical evidence on relative quality and cost of teaching hospitals and (2) assess what the findings indicate for public and private insurer policy.

Methods

Complementary approaches were used to select studies for review. (1) Relevant studies highly cited in Web of Science were selected. (2) This search led to studies cited by these studies as well as studies that cited these studies. (3) Several literature reviews were helpful in locating pertinent studies. Some policy‐oriented papers were found in Google under topics to which the policy applied. (4) Several papers were added based on suggestions of reviewers.

Findings

Quality of care as measured in process of care studies and in longitudinal studies of long‐term survival of cancer patients tends to be higher at major teaching hospitals. Evidence on survival at 30 days post admission for common conditions and procedures also tends to favor such hospitals. Findings on other dimensions of relative quality are mixed. Hospitals with a substantial commitment to graduate medical education, major teaching hospitals, are about 10% to 20% more costly than nonteaching hospitals. Private insurers pay a differential to major teaching hospitals at this range's lower end. Inclusive of subsidies, Medicare pays major teaching hospitals substantially more than 20% extra, especially for complex surgical procedures.

Conclusions

Based on the evidence on quality, there is reason for patients to be willing to pay more for inclusion of major teaching hospitals in private insurer networks at least for some services. Medicare payment for GME has long been a controversial policy issue. The actual indirect cost of GME is likely to be far less than the amount Medicare is currently paying hospitals.

The relative performance of teaching hospitals has been an important public issue since at least the early 1980s, when Congress passed the Tax Equity Fiscal Responsibility Act of 1982 (TEFRA). This legislation replaced Medicare retrospective reimbursement of hospitals with a prospective payment system (PPS). About this time, the California legislature enacted selective contracting, which allowed the state's Medicaid program to exclude hospitals from its networks. ¹ This policy diffused to other states and insurers.

To address teaching hospitals’ concerns that PPS prices did not adequately account for cost differences between teaching and nonteaching hospitals, TEFRA arguably solved the problem by implementing subsidies of graduate medical education (GME). ² , ³ Quantifying the relative benefit of teaching hospitals has proven elusive, particularly given the multidimensional nature of hospital output. That teaching hospitals, especially those affiliated with academic health centers, are more costly is a given. Their relative cost, and sources of cost differences, are more uncertain. Their relative productivity—that is, the differential value of their outputs relative to the cost differential—has remained an unresolved issue for decades.

Some public insurers, most notably traditional (fee‐for‐service) Medicare, permit free patient choice among participating providers. Then the insurer's decision is to determine the size of the payment made, not to develop a provider network. For private health insurers, purchasing agents for employers and ultimately patients, and for public insurers with provider networks, such as Medicare Advantage and state Medicaid programs with selective contracting, the issue of relative quality as well as cost is pertinent for selecting providers for inclusion in their networks. A private insurer's decision to include a hospital in its network reflects a balancing of insureds’ willingness to pay for the option of choosing the hospital against the cost of the hospital's services to the insurer. Insurance plans may limit access to major teaching hospitals for conditions or procedures for which outcomes tend to be better.

This study aims to (1) summarize and evaluate empirical evidence on relative quality and cost of teaching hospitals and (2) assess implications of the findings for public and private insurer policy.

Studies highly cited in Web of Science under such headings as “Teaching hospital AND Cost” and “Teaching Hospital AND Quality” were identified. This search led to studies cited by these studies as well as to studies that cited these studies. In addition, several literature reviews were helpful in locating pertinent studies. ⁴ , ⁵ , ⁶ Some more policy‐oriented papers were found in Google under the topics to which the policy applied. And last, some studies were recommended by reviewers.

Context and Framework

With a few exceptions, hospitals with GME programs are organized as either private not‐for‐profit (NFP), the largest ownership group with such programs, or state or local government (SLG) rather than as for‐profit (FP) entities. ⁷ The essential difference between a profit‐maximizing FP organization and NFP and SLG organizations is that the latter do not have shareholders to whom residual income accrues as with FPs. NFPs and SLGs are eligible for private donations and public grants, they do not pay corporate income taxes, and tax‐exempt bonds can be issued to cover capital expenditures. In return, NFP hospitals are to provide some community benefits. ⁸ Some models of NFP and SLG behavior predict that such hospitals will provide a combination of higher quality and/or higher quantity of care at a lower quality‐adjusted price than an FP hospital would. Some aspects of hospital care are noncontractible. ⁹ Profit‐maximizing hospitals may skimp on aspects of care that are not readily observable by consumers. ¹⁰ And at least theoretically, they would not provide care for which marginal cost exceeds marginal revenue. Alternative ownership‐based models of hospital behavior and the implications of these models are described by Malani et al. ¹¹ This article's analysis takes the dominance of non‐FP teaching hospitals as a given and asks about the extent to which being a major teaching hospital per se affects quality and cost of hospital care. Definitions of major teaching hospitals vary as shown in this article's tables. For example, major teaching hospitals are often defined as hospitals affiliated with a medical school and/or a member of the Council of Teaching Hospitals (COTH). Sometimes, the definition of major teaching hospital has been based on a threshold above a specific hospital resident‐to‐bed (RTB) ratio. The list of major teaching hospitals from the second definition overlaps the first, but the correspondence is not exact. Moreover, studies using a RTB ratio threshold to distinguish between major and minor teaching hospitals, the latter hospitals with some residents, have not all used the same threshold.

Hospitals are often designated as teaching hospitals based on the number of residents relative to beds and/or membership in an organization such as the COTH. But residents are only one of many hospital inputs (Figure 1, left). The presence of multiple inputs affects the marginal product of each input. The marginal product of each input is plausibly positive at the levels at which the input is employed. Added to this are possible differences in hospital objectives other than profit associated with hospital GME teaching intensity, which will also affect input use. Very resident‐intensive hospitals may have unique organizational objectives—for example, they may be particularly research oriented. The focus of analysis should be on differences in hospital input packages rather than solely on variation in hospital resident‐to‐bed (RTB) ratios.

Figure 1.

Hospital Production Function

Hospital output is multidimensional (Figure 1, right). There is no generally accepted hospital output index that would use weights reflecting society's preferences to aggregate output to a single index. Unfortunately, many studies have used evidence on a single element of output, mostly mortality, to conclude that hospitals with particular attributes are relatively effective. Various rankings of hospitals, which presumably reflect some sort of weighted index of output or quality, do exist. They often are based on combinations of results from patient satisfaction surveys and indicators of patient safety and clinical quality. Perez and Friedman documented inconsistencies among these rankings. ¹²

Marginal products of inputs are influenced by several factors (Figure 1, middle). First, all things being equal, marginal products are typically diminished by a more complex case mix. But major teaching hospitals’ marginal product may be less diminished than others, possibly giving them a comparative advantage in treating patients with complex conditions. Second, being an early adopter of new, effective technologies may increase a hospital's relative productivity. However, the empirical evidence on adoption of new technology by hospital teaching status is mixed. ¹³ , ¹⁴ , ¹⁵ , ¹⁶ Teaching hospitals have been early adopters of some technologies but not others. Third, to the extent that major teaching hospitals have more active research programs, new findings may be translated into more effective clinical care. On the other hand, research may, like teaching, divert faculty resources from clinical care. Fourth, relative productivity is likely to be enhanced by more widespread adherence to guidelines, as in the use of established technologies. Skinner and Staiger found that diffusion of low‐cost innovations for treating patients following heart attacks—such as use of beta‐blockers, aspirin, and timely use of reperfusion therapy—explained a large part of the variation in hospital productivity as measured by survival gains. ¹⁷ As documented in this article, major teaching hospitals have been more likely to employ these low‐cost innovations. But fifth, greater amounts of technical inefficiency (“slack”) and diseconomies of scale reduce productivity. Slack may be more likely at major teaching hospitals because of their multiple organizational objectives. Unfortunately, existing data do not allow researchers to quantify the relative importance of all of the individual factors listed in Figure 1, middle. Fortunately, some databases have allowed researchers to make refined adjustments for differences in hospital case mix.

Quality

Two measures of quality of care of hospitals are in widespread use: process of care and health outcomes. Process of care measures services provided for which there is empirical evidence that implementation of the services leads to improved outcomes, or at a minimum, there is consensus among experts that there is a likely causal relationship between service provision and outcomes. A strength of process of care measures is that they often take patient case mix into account. The process refers to the appropriateness of care, given the patient's condition. A disadvantage is that large‐scale process of care studies can be very costly to conduct, requiring professional data abstractors and physicians to review the abstracted records. Probably for this reason, large‐scale process of care studies based on abstracted medical records have not been conducted since the 1990s.

The most common outcome measure of quality is survival at various dates following the hospital admission date, most frequently 30 days. Other outcomes include postsurgical complications, failure to rescue (death following a complication), readmission rates, and outcomes specific to the diagnosis being considered. The major information sources on process of hospital care come from claims data merged with data from retrospective review of hospital records.

In the material that follows, tables providing details on study data, methods, and findings supplement the text. Some studies provide evidence under multiple subject headings. To conserve space, the tables focus on findings from the part of the study under the heading for which the study is first mentioned. Additional findings are described only in the text.

Processes of Care

A classic study of medical malpractice in hospitals is the Harvard Medical Practice Study (Table 1). The overall objective of this research was to measure the frequency of adverse events and adverse events due to negligence in hospitals based on clinicians’ review of medical records in New York state hospitals. This study contributed immensely to knowledge about medical malpractice, only the second study to document medical error rates and determining the share of such errors attributable to provider negligence. From multivariate analysis, Brennan et al., an article based on analysis of Harvard Medical Practice Study data, reported a much higher rate of adverse events in “primary teaching” than in nonteaching hospitals (odds ratio [OR] = 2.29, p = 0.02), but the adverse events in primary teaching hospitals were much less likely to be attributable by the Harvard study's independent physician reviewers to negligence (OR = 0.26, p = 0.02). ¹⁸ There were no statistical differences between “affiliate teaching” and nonteaching hospitals.

Table 1.

Processes of Care

Authors	Description	Definition of Teaching Hospital	Sample	Key Findings
Brennan et al., 1991 18	Medical records review. Measured adverse events (AE) and AE due to negligence in 51 New York hospitals.	MT: “Primary teaching,” hospitals owned or closely tied to med schools; OT: “Affiliate teaching,” other hospitals with residents.	31,429 hospitalized patients	AEs due to negligence (%): Unadj. MT: 10.7; OT: 30.1; NT: 26.9, p < .0001. Adj. relative to NT‐NFP: MT OR = 0.26; OT n.s.
Thomas et al., 2000 19	Same method as Brennan et al. but for 28 hospitals in Colorado and Utah; 2 MT, 8 OT, 18 NT.	Not explicitly defined but presumably same as Brennan.	15,000 hospitalized patients	Adj. Compared to patients in MT, patients in OT and NT hospitals more likely to suffer preventable AE. OR = 2.46 (95% CI: 1.45‐4.20).
Ayanian et al., 1998 20	Medical records review of hospital records in Illinois, Massachusetts, New York, and Pennsylvania. Adjusted for age, sex, Zip code, income, RAND Sickness at Admission Scale.	MT: RTB > 0.25; OT: Not MT but some residents.	1,767 Medicare beneficiaries hospitalized with congestive heart failure (CHF) or pneumonia	Adj. MT significantly better than NT for both conditions in overall implicit and explicit quality. OT also significantly higher than NT on both overall quality measures, but differences were smaller. Similar patterns for physician cognitive and technical diagnosis subscales. But on nurse cognitive subscales, NT had higher scores.
Keeler et al., 1992 21	Medical records review conducted in 297 hospitals in 5 states.	MT: RTB > 0.27; OT: Not MT but some residents.	14,008 Medicare beneficiaries hospitalized with CHF, acute myocardial infarction (AMI), pneumonia, hip fracture, or stroke	Explicit process: MT 0.28; OT 013; NT: −0.09. Implicit process: MT: 0.66; OT: 0.21; NT: −0.18. Excess mortality (%): MT: −2.5; OT: −0.8; NT: 0.6. Unadj. All differences between MT and NT significant at ≤0.05. All differences between OT and NT n.s.
Allison et al., 2000 22	Used data from Cooperative Cardiovascular Project (CCP). Rates calculated for patients clinically appropriate for therapy.	2 alternative definitions: Method 1. MT: RTB ≥ 0.10; OT: Not MT but some residents. Method 2. MT: Council of Teaching Hospitals (COTH) member and accredited by Council for Graduate Medical Education. OT: Not MT but with accredited residency programs. Mainly used Method 1	114,411 Medicare beneficiaries hospitalized for AMI	All unadj. (%) Use of aspirin. MT: 91.2; OT: 86.4; NT 81.4, p < 0.001. Use of ACE inhibitors. MT: 63.7; OT: 60.7; NT: 58.0, p < 0.001. Use of beta‐blockers. MT: 48.8; OT: 40.3; NT: 36.4, p < 0.001. Refusion therapy. MT 55.5; OT 58.9; NT 58.2, p = 0.29.
Chen et al., 1999 23	Used CCP. Compared care processes for patients clinically appropriate for therapy and outcomes for 60 top‐ranked hospital cardiology programs in U.S. News & World Report.	Did not precisely define teaching but 59 of 60 top‐ranked hospitals were said to be MT.	149,177 Medicare beneficiaries hospitalized for AMI	Unadj. (%) Use of aspirin: 96.2 top‐ranked; 88.6 comparable; 83.4 remaining, p < 0.01 top‐ranked vs. remaining. Use of beta‐blockers: 75.0 top‐ranked; 61.8 comparable; 58.7 other, p < .01, top‐ranked vs. remaining. Reperfusion therapy: 61.0 top‐ranked; 70.7 comparable; 65.6 other, p = 0.3, top‐ranked vs. remaining.
Landon et al., 2006 24	Used hospital performance measures reported to Joint Commission on Accreditation of Healthcare Organizations and Medicare for patients hospitalized for CHF, AMI, and pneumonia. Authors created composite scores for each disease and used factor analysis to derive 2 additional composites. Used logit to examine relationships between hospital attributes and quality. Observational unit, the hospital.	MT: Member of COTH; OT: Not MT but affiliated with medical school or with residency program	Sample size varied depending on comparison being made. Up to 3,380 hospitals in an individual analysis	Treatment and diagnosis composite (adj.), NT reference group. MT: OR = 1.37 (95% CI: 1.34‐1.39); OT: OR = 1.10 (95% CI: 1.08‐1.14). Counseling and prevention compositive (adj.) MT: OR = 0.83 (95% CI: 0.82‐0.84); OT: OR = 0.88 (95% CI: 0.87‐0.89).

Abbreviations: adj., adjusted for other factors in multivariate analysis; CI: confidence interval; MT, major teaching hospital; NT, nonteaching hospital (omitted reference group); NFP, private not‐for‐profit hospital; OR, odds ratio; OT, other teaching hospital; RTB, resident‐to‐bed ratio; unadj., unadjusted for other factors; n.s., p‐value not ≤0.05.

In a follow‐up study of Colorado and Utah hospitals using the same basic methodology, Thomas et al. documented that preventable adverse events were much higher in minor teaching and nonteaching hospitals than in major teaching hospitals (OR = 2.46; 95% confidence interval [CI]: 1.45‐4.20; major teaching, reference group). ¹⁹ An adverse event was “preventable” if the event was judged to be avoidable using any means available unless the means were not standard of care. A limitation was that there were only two major teaching hospitals in the study, one in each state, both government owned.

Several process of care studies not focused on medical malpractice, published between 1992 and 2006, merit attention, particularly because they included many hospitals in several states, and they supplemented claims data with information from medical records.

In a study of hospital care for congestive heart failure (CHF) and pneumonia in four states, using implicit and explicit measures of process quality, Ayanian et al. found that after adjusting for other factors, major teaching hospitals had better overall explicit process quality of care than nonteaching hospitals, especially on physician cognitive skills and diagnosing. ²⁰ Explicit process measures are based on particular disease‐specific criteria for appropriate care for the condition. Adherence to explicit process criteria was assessed from medical record information abstracted by a registered nurse. Nonteaching hospitals had higher scores on the nurse cognitive subscale, a measure of a nurse's reasoning process and actions taken, given the clinical issues presented to the nurse. Implicit process quality assessments were based on clinician responses to specific questions about the overall quality of care in a specific case. Implicit process of care measures are designed to provide an overall assessment of care quality. One question asked of physician reviewers of the raw data used for the implicit process measure was, “Based on what you now know about this case, would you send your mother to this hospital?” On implicit process, ratings were highest for major teaching hospitals followed by minor teaching and nonteaching hospitals, in that order.

Keeler et al. conducted a much larger process of care study (more patients and physician reviewers) of Medicare beneficiaries hospitalized in five states. ²¹ They found that both explicit and implicit measures of quality were higher for hospitals with higher teaching intensity after controlling for other factors using a computerized recursive partitioning algorithm.

Allison and coauthors’ process of care comparisons were based on data from the Cooperative Cardiovascular Project (CCP), ²² a national study of Medicare patients admitted to 4,361 hospitals for acute myocardial infarction (AMI) during 1994‐1995. The CCP collected potentially important clinical information not available from insurance claims, for example, body mass index, current smoking status, whether the persons had a terminal illness, arterial pressure, shock, and pulmonary edema on arrival at the hospital, electrocardiogram readings, chest pain lasting 460 minutes after arrival, cardiac arrest, and creatinine levels. The CCP's major advantages are in combining Medicare claims data with data from medical records’ review, which adds much more information on patient clinical attributes and care processes than is available from claims data alone.

Major teaching hospitals exhibited superior performance in terms of several process of care indicators, including rates of administration of aspirin, angiotensin‐converting enzyme (ACE) inhibitors, and beta‐blockers (when external reviews judged this therapy to be clinically appropriate); according to another study, these are the types of technologies hospitals that experienced greater gains in mortality reduction adopted. ¹⁷ Rates of reperfusion therapy were slightly lower at major teaching hospitals. The authors speculated that more complex processes at major teaching hospitals may have impeded timely administration of reperfusion therapy at these institutions. Process of care comparisons based on multivariate analysis were not presented. Also using CCP data but with a larger number of observations, Chen and coauthors reported similar results for hospitals rated “best” by U.S. News & World Report; ²³ 59 of the 60 hospitals in the sample were major teaching hospitals.

Two other sources of information of hospital‐specific processes of care are the Joint Commission for Accreditation of Healthcare Organization (JCAHO) and Medicare. JCAHO initiated its survey in 2001. Beginning in 2004, the Medicare Modernization Act required that hospitals report ten measures specific to CHF, AMI, and pneumonia to Medicare.

Landon et al. used these data to construct two types of measures of process quality: (1) disease‐specific measures and (2) results from a factor analysis of measures from all three diseases. ²⁴ On the first measure type, compared with nonteaching hospitals, major teaching hospitals provided higher quality of care for AMI but not for CHF or pneumonia. The first two factors from the factor analysis were interpreted as (1) quality of diagnosis and treatment and (2) counseling and prevention (e.g., counseling patients who smoked at the time of the hospitalization to quit). Major teaching hospitals performed better on the first but not on the second composite process of care measure.

Overall, but with exceptions noted, on process of care, being a major teaching hospital was associated with higher process quality. Empirical analysis of CCP data for AMI revealed that major teaching hospitals were exceptional in providing low‐tech, high‐quality care, which hospitals irrespective of their teaching mission can provide. ²⁵

Outcomes

Mortality. Quality studies with health outcome measures far outnumber process of care studies, especially among those published since 2000. Most health outcome studies have relied on claims data.

Among the mortality outcome studies, the strongest empirical support for higher quality of care in major teaching hospitals comes from cancer studies. Survival is the most relevant outcome for most cancers, in contrast to a knee replacement, for example, for which reduction in pain and improvement in function are likely to be more relevant outcomes.

Several studies have analyzed longer‐term (two‐ to five‐year) survival of patients diagnosed with specific types of cancer (Table 2). ²⁶ , ²⁷ , ²⁸ , ²⁹ , ³⁰ , ³¹ , ³² Not only is the longer‐term follow‐up a strength, but these studies controlled for attributes of the tumor not in claims data—for example, tumor stage, grade, depth, ³⁰ and size ²⁶ at initial diagnosis. These studies also tended to control for surgical procedure volume. Survival of patients admitted to major teaching hospitals exceeded that of patients admitted to nonteaching hospitals by more than 10%. One study contradicts these findings, but it used in‐hospital mortality as the outcome measure, and “teaching hospital” was not precisely defined. ³²

Table 2.

Cancer Mortality Studies

Authors	Description	Definition of Teaching Hospital	Sample	Key Findings
Chaudry et al., 2001 26	Compared survival outcomes of women with node‐negative breast cancer initially treated at major teaching vs. nonteaching hospitals. Used hazard model.	MT: Member of Association of Canadian Teaching Hospitals.	Random sample of 938 cases drawn from 1991 Ontario Cancer Registry	53% reduction in relative risk of death among women initially treated at MT if tumor ≤ 20 mm (RR = 0.47; 95% CI: 0.23‐0.96). n.s if tumor > 20 mm.
Cheung et al., 2009 27	Used Cox proportional hazard to analyze determinants of risk of death of patients undergoing surgical resection for lung cancer. Survival tracked up to 96 months.	MT: Member of Association of American Medical Colleges	13,609 patients selected from Florida Cancer Data System merged with data from Florida Agency for Health Care Administration	In multivariate analysis, risk of death was 14% lower at teaching hospitals (HR = 0.84; 95% CI: 0.78‐0.91).
Ramlingam et al., 2018 28	Compared survival of patients diagnosed with Stage IV lung cancer, 1998‐2010. Primary outcome measure was survival at 2 years following diagnosis. Used ordinary least squares regression with interaction terms between facility type and diagnosis year to study trends in difference in survival by teaching status.	Used National Cancer Database (NCDB) definition: community cancer program; comprehensive cancer program; academic research program (AC). Non‐AC programs combined as CC.	193,279 patients from NCDB	Survival was higher in AC in all years. Differential in survival between AC and CC was higher in the later years.
Cheraghlou et al., 2019 29	Compared survival rates of patients with diagnosis of primary nonmetastatic melanoma by facility type and volume. Used Cox proportional hazard and alternatively propensity‐score matching. Considered effect of teaching status overall and in interaction with volume.	Not precisely defined. Probably used NCDB categories.	300,365 patients from NCDB	With nonteaching, the reference group's overall survival was higher for teaching hospitals (HR = 1.05; 95% CI: 1.02‐1.62). Interacted with facility volume, teaching improved survival only for facilities in the highest quartile of volume.
Uhlig et al., 2019 30	Compared survival of patients with hepatocellular carcinoma. Used hazard model.	Used NCDB definition.	63,877 patients from NCDB	Death rate was lower for patients treated in teaching hospitals (HR = 0.79; 95% CI: 0.77‐0.81).
Boffa et al., 2020 31	Analyzed 90‐day mortality and long‐term survival of adults undergoing surgery for these cancers: esophageal, gastric, lung, pancreatic, colorectal, bladder. Used logit for outcome at 90 days and hazard for longer‐term outcomes. Patients observed for up to 7 years following surgery. Controlled for age, sex, insurance coverage, previous malignancy, tumor stage, grade, chemotherapy, and radiation receipt.	Compared outcomes between “top‐ranked” hospitals and “affiliates” of top‐ranked hospitals. 97.3% of patients at top‐ranked had academic affiliation and/or National Cancer Institute designation. Only 20.0% of affiliate hospitals did.	119,834 patients from NCDB	All else being equal, compared to top‐ranked, mortality at 90 days was higher at affiliate hospitals (OR = 1.67; 95% CI: 1.49‐1.89). Adjusted long‐term survival was higher at top‐ranked hospitals (HR = 0.77; 95% CI: 0.72‐0.83).
Syed et al., 2020 32	Studied in‐hospital mortality of patients undergoing urologic oncologic surgical procedure using Vizient Database.	MT and OT combined. Used term “academic” hospital, but term not precisely defined.	37,628 patients of all ages hospitalized for these procedures	Reported no differences in mortality between academic and community hospitals.

Abbreviations: adj., adjusted for other factors in multivariate analysis; CI, confidence interval; HR, hazard ratio; MT, major teaching hospital; NT, nonteaching hospital (omitted reference group); OT, other teaching hospital; RR, relative risk; RTB, resident‐to‐bed ratio; unadj., unadjusted for other factors.

Another condition for which detailed data are available is AMI, in particular from the CCP. Using data from the CCP, Allison et al. reported that patient characteristics (e.g., Charlson score; APACHE II score; percentages of patients with diabetes, hypertension, chronic renal insufficiency, CHF) did not differ much by hospital teaching status and were sometimes more adverse for nonteaching hospitals. ²² Differences tended to be statistically significant due to the large sample size. Risk adjusters included clinical findings not available for claims data. With risk adjustment and inclusion of covariates for the process of care (described earlier), adjusted odds ratios in the Allison et al. study were 0.91 (95% CI: 0.87‐0.96) for major teaching and 0.96 (95% CI: 0.92‐1.01) for minor teaching hospitals, indicating no statistical difference in 30‐day mortality between minor and nonteaching hospitals. Adjusted results for longer follow‐ups of mortality were not presented.

In a subsequent study focusing on the effect of a dementia diagnosis on care and outcomes of Medicare beneficiaries admitted to a hospital for AMI using CCP data, Sloan et al. obtained statistically insignificant adjusted odds ratios for both major and minor teaching hospitals for 30‐day and one‐year mortality (Table 3). ³⁴ A substantial problem with the literature lies in inconsistencies in definitions of teaching hospitals. In some studies, there is no definition. But here, Allison et al.’s Method 1 of defining teaching (see Table 2) is virtually identical to Sloan et al.’s. The difference in results from Allison et al. plausibly reflects differences in equation specification. Odds ratios on covariates for diagnosed dementia and presence of a do‐not‐resuscitate order at admission included in Sloan et al. but not in Allison et al. were well above one and highly significant. Patients admitted with an AMI diagnosis with cognitive impairments tend to receive much less aggressive care. ³⁴ Do‐not‐resuscitate orders would have been signed prior to the admission for AMI and should have affected the course of treatment and mortality. This difference in findings serves as a warning that effects of hospital teaching status on mortality following admission for AMI are sensitive to the risk adjusters included in the mortality equation.

Table 3.

Other Mortality Studies

Authors	Description	Definition of Teaching Hospital	Sample	Key Findings
Sloan et al., 2004 33	With data from the Cooperative Cardiovascular Project (CCP), focus of study was on differences in mortality and services use for patients diagnosed vs. not diagnosed with dementia. Logit analysis used.	MT: RTB ≥ 0.097 OT: RTB < 0.097. (0.097 = median value of RTB for hospitals with residents).	129,092 Medicare beneficiaries admitted for acute myocardial infarction (AMI)	After controlling for dementia, socioeconomic characteristics, admission source, comorbidities, 10 measures of severity of cardiac disease, and hospital characteristics including ownership, neither MT nor OT status affected 30‐day or 1‐year mortality.
Taylor et al., 1999 35	Study sample came from waves of the National Long‐Term Care Survey (NLTCS). Using NLTCS allowed control for cognitive and functional status prior to index admission. 4 study conditions: hip fracture (HF), stroke, coronary heart disease (CHD), and congestive heart failure (CHF). Used Cox proportional hazard for mortality analysis.	MT: RTB ≥ 0.097 OT: RTB < 0.097. (0.097 = median value of RTB. for hospitals with residents)	Number of Medicare beneficiaries by condition for admission: HF, 802; stroke, 793; CHD, 1,007; CHF, 804	Combining samples for 4 conditions, major teaching hospitals had lowest hazard of death, relative to NT‐FP hospitals: HR = 0.75; 95% CI: 0.62‐0.91. For individual conditions, only significant difference was for HF (HR = 0.54; 95% CI: 0.37‐0.79).
Silber et al., 2020 36	Compared outcomes between MT and NT hospitals for patients admitted for AMI, CHF, and pneumonia. Used propensity‐score matching. Mortality was all‐cause death rate within 30 days of admission. Included a “patient risk of death within 30 days” score as a covariate.	MT: RTB ≥ 0.25; NT: RTB < 0.05	Medicare claims and enrollment files (pairs). AMI: 43,980; CHF: 84,985; pneumonia: 74,947	30‐day mortality was lower in MT than in NT hospitals (10.7% vs. 12.0%). For the quintile of pairs at highest 30‐day death risk, death rates were 24.6% MT and 27.6% NT. Results on individual conditions were similar to pooled results.
Silber et al., 2020 37	Compared outcomes between MT and NT hospitals for patients admitted for general (GS), orthopedic (ORS), and vascular surgery (VAS). Same methods as Silber et al. 2020. 36 Used propensity‐score matching.	MT: RTB ≥ 0.25; NT: RTB < 0.05.	Medicare claims and enrollment files (pairs). GS: 86,751; ORS: 214,302; VAS: 52,125	For GS, mortality was 4.62% MT vs. 5.57% NT; patterns for VAS similar to GS. However, no statistically significant difference in mortality for ORS.
Silber et al., 2009 38	Study had 2 goals, to determine whether (1) lower mortality in teaching hospitals at 30 days stems from lower complication rates or lower death rates following complications (“failure to rescue”); and (2) results were the same for blacks and whites. Used logit with and without random and fixed effects. Adjusted for case mix with measures available on Medicare claims and interactions between case mix and demographics.	“Very major”: 0.6 < RTB ≤ 1.1; “Major”: 0.25 < RTB ≤ 0.6; “Minor”: 0.05 < RTB = 0.25; “Very minor”: 0 < RTB = 0.05.	Medicare claims and enrollment files. GS: 809,473; ORS: 954,374; VAS: 257,467	Combining all 3 surgical procedures, compared with NT hospitals, very major teaching hospitals had 15% lower odds of death at 30 days. There was no difference in complication rates, but “very major” had 15% lower odds of death at 30 days due to failure to rescue. There was no survival advantage between these 2 types of hospitals for blacks.
Burke et al., 2017 40	Examined risk‐adjusted mortality at 30 days (primary outcome) and 30‐day mortality by hospital size and 7‐day and 90‐day mortality (secondary outcomes) for 15 medical and 6 surgical conditions in the aggregate and individually. Used “linear regression.”	MT: Council of Teaching Hospitals (COTH) member; OT: Medical school affiliated but not COTH member.	21.5 million hospitalizations; Medicare beneficiaries	Unadj. 30‐day mortality. 8.1% MT, 9.0% OT, and 9.3% NT. Adj. (for patient and hospital characteristics), difference MT vs. NT, 1.2% (95% CI: 1.0‐1.4). Comparing 400+ bed MT with NT, difference was 1.2%; 100‐ to 399‐bed MT with NT, difference was 0.08%, <100 bed; OT‐NT comparison, OT 0.4% lower.
Burke et al., 2018 7	Examined (1) difference in 30‐day mortality by teaching status for patients with high, medium, and low severity of illness; and (2) whether MT mortality advantage is attributable to more advanced technology at MTs. Created severity score based on predicted 1‐year mortality. Top 10% of predicted probabilities = “high”; bottom 10% “low”; in between % “medium”. Used logit.	MT: COTH and Association of American Medical Colleges member; OT and NT categories combined into single category (reference group).	11.9 million hospitalized Medicare beneficiaries; conditions same as in Burke et al., 2017	Adj. for patient and hospital characteristics. For high severity, 7% lower odds for MT; medium severity, 13.1% lower odds; and for low severity, 17% lower odds for MT for medical conditions. For surgical patients, differences were 17% lower for high severity, 10% lower for medium severity, and no difference for low severity.
Dimick et al., 2004 42	Studied inpatient mortality and prolonged length of stay (>75th percentile) for patients undergoing 3 complex surgical procedures: esophageal resection (ESO); hepatic resection (HEP), and pancreatic resection (PAN).	In primary analysis, teaching was defined as COTH member. In sensitivity analysis, teaching alternatively defined as presence of an accredited residency program or medical school affiliation.	Data on individual hospitalizations from Nationwide Inpatient Sample merged with American Hospital Association data for hospital attributes. ESO: 1,247; HEP: 2,073; PAN: 3,337	Unadj. Inpatient mortality rates were 5.3% for teaching vs. 8.0% for nonteaching for PAN; 5.3% vs. 8.0% for HEP; and 7.7% vs. 10.2% for ESO. Adj. In multivariate analysis, which included a covariate for hospital volume, teaching vs. nonteaching inpatient mortality rates were not statistically different.
Bekelis et al., 2018 43	Used data from New York Statewide Planning and Research Cooperative System. Primary outcome measure was in‐hospital fatality rate following spinal or cranial surgery. Used an instrumental variable (IV) for patient choice of teaching vs. nonteaching hospital. Number of teaching hospitals as % of all hospitals in county. Alternatively, used propensity‐score matching (PSM).	Teaching hospital defined as hospital affiliated with a university.	180,483 patients hospitalized for cranial or spinal surgery. Vast majority of patients hospitalized in teaching hospitals, 158,288	Case fatality rate higher in teaching hospitals but difference was very small. In %, IV result: 0.3 (95% CI: 0.02‐0.04). PSM result exactly the same.
Geweke et al., 2003 44	Quality measured by survival at 10 days following admission for pneumonia. Study controlled for patient selection of hospital using a model in which distance from patient's home to alternative hospitals were a key exogenous factor. Bayesian inference used a Markov chain Monte Carlo posterior, with posterior probabilities used to make quality comparisons among hospitals. Used discharge abstract data from California Office of Statewide Health Planning and Development. Data contained detailed disease staging at admission.	Teaching hospital not defined.	74,848 patients hospitalized for pneumonia who were aged 65+, at 114 hospitals in Los Angeles County	“Private teaching” hospitals (N = 5) have highest quality. With respect to bed size, found U‐shaped relationship with highest quality being small and large hospitals.

Abbreviations: adj., adjusted for other factors in multivariate analysis; CI, confidence interval; HR, hazard ratio; MT, major teaching hospital; NT, nonteaching hospital (omitted reference group); OT, other teaching hospital; RTB, resident‐to‐bed ratio; unadj., unadjusted for other factors.

Most studies of the effect of teaching status on mortality have relied on Medicare administrative data—claims and enrollment records (Denominator File) for dates of death, in a few cases merged with household survey data. Taylor et al. analyzed mortality of Medicare beneficiaries admitted during 1984‐1994 for hip fracture, stroke, coronary heart disease (CHD, a broader category than AMI), and CHF (Table 3). ³⁵ A Hierarchical Coexisting Condition (DxCG) score was included among the risk adjusters. Up to nine comorbidities per admission were included in this score. Binary variables based on ICD‐9‐CM codes accounted for clinical differences in each of the four study conditions.

The study is noteworthy for two reasons. First, mortality outcomes were followed for up to 11 years from the index admission; by the end of the observational period, 72% of beneficiaries had died (range: 62% [CHD] to 83% [CHF]). Teaching hospitals’ outcomes might be better in the long term because of more accurate diagnosis, more timely and intensive treatment, and referrals to more capable providers after hospital discharge. Second, survey data provided information unavailable on claims but useful for risk adjustment, such as cognitive and functional status and living arrangements prior to admission. Because of the merge of Medicare data with a household survey, the analysis sample was unusually small, resulting in a loss of statistical power for a gain in accuracy of risk adjustment.

Combining patients in all four diagnostic categories, and adjusting for case mix observable to researchers and other factors, major teaching hospitals had a 25% lower hazard of death than the omitted reference group, FP hospitals. Analyzed separately, the only diagnostic category for which major teaching hospitals reduced the death hazard was hip fracture. Differences in mortality between minor teaching and FP hospitals were never statistically significant.

In a recent study with a much larger sample than Taylor et al.’s and a much higher RTB threshold for major teaching hospitals than either Allison et al. or Taylor et al., Silber et al. assessed mortality 30 days post admission for AMI, CHF, and pneumonia using Medicare claims data. ³⁶ Since the study relied on Medicare claims data, risk adjustment was probably less precise than in studies that supplemented claims with medical records abstraction or household survey data. Beneficiaries admitted for the study diagnoses to major teaching hospitals were propensity‐score matched to those admitted to nonteaching hospitals. The AMI and CHF matches were each based on 68 covariates; for the pneumonia matches, there were 76 covariates. Covariates included the patient's principal diagnosis category and the quintile of the patient's 30‐day mortality risk at admission (predicted from a separate analysis), demographic characteristics, and admission type (e.g., emergent, transfer), as well as covariates identifying types of AMI, CHF, and pneumonia.

Combining the three diagnoses, 30‐day mortality was 1.3% lower in major teaching than in nonteaching hospitals. Results for the individual diagnoses were not reported but were said to be similar. The mortality advantage for major teaching hospitals rose monotonically as patient illness severity increased, measured by predicted mortality at admission (excluding hospital type actually chosen from the list of predictors), implying that major teaching hospitals were relatively effective in reducing short‐term mortality for high‐risk beneficiaries.

Using the same methodology, Silber and coauthors examined 30‐day mortality outcomes of Medicare beneficiaries admitted to general, orthopedic, and vascular surgery units. ³⁷ On average, 30‐day survival is higher for such surgical than for medical patients. Highly skilled providers may be better at weighing risk versus benefit in such cases. Although not specifically focused on surgery, Skinner and Staiger documented the importance of skill in clinical decision making, ¹⁷ especially in the diagnostic phase, which would encompass the decision to perform a procedure.

In the Silber et al. surgery study, treatment and control groups were matched on 224, 171, and 119 covariates, respectively, including a score representing the propensity of being admitted to a major teaching hospital to deal with the patient choice issue. ³⁷ For some covariates, the matching was exact. As in the medical study, ³⁶ Silber et al. ³⁷ included a covariate for patient severity of illness at admission. For general surgery, the difference in mortality between major teaching and nonteaching hospitals was −0.95% (favoring major teaching hospitals), but for the quintile of patients with the highest ex ante risk of death within 30 days, the difference was much larger, −2.24%. For vascular surgery, differences were smaller but still statistically significant at conventional levels: −0.39% (overall) and −0.90% (highest ex ante death quintile). There were no statistical differences in 30‐day mortality for orthopedic surgery. The authors did not explain why findings for orthopedic surgery differed. The difference in findings serves to show the importance of not generalizing to all care from findings on single conditions or procedures.

In a separate study, Silber et al., using Medicare claims data on general, vascular, and orthopedic surgery admissions, found that major teaching hospitals had lower mortality, no difference in complications, and lower rates of failure to rescue when complications occurred. ³⁸ But better survival and better (that is, lower) failure‐to‐rescue outcomes were observed in major teaching hospitals for white but not Black persons. Details on reasons for difference in outcomes by race were not provided. Major teaching hospitals tend to admit higher proportions of Black persons than do other hospitals (see, for example, Skinner et al. ³⁹ ).

In a study of Medicare beneficiaries admitted to hospitals for 15 medical and six surgical conditions, using linear regression, Burke et al. reported an improvement in 30‐day mortality of 1.2% (p < 0.001) for major teaching over nonteaching hospitals. ⁴⁰ These results are consistent with the process of care study by Keeler et al., who reported a difference between observed and predicted 30‐day mortality of −2.5% for major teaching and of 0.7% for nonteaching hospitals. ²¹ The difference between minor teaching and nonteaching hospitals was much smaller.

Burke et al. classified Medicare beneficiaries according to their predicted one‐year mortality risk independently of the type of hospital to which the beneficiary was actually admitted. ⁷ The classifications were “high” (highest decile), “low” (lowest decile), and “medium risk” (the rest). After adjusting for patient and other hospital characteristics, high‐severity medical patients had 7% lower odds; medium‐severity patients, 13% lower odds; and low‐severity patients, 17% lower odds of death 30 days post admission. For surgical procedures, high‐severity patients had 17% lower odds of death at 30 days if admitted to major teaching rather than to other hospitals; for medium‐severity patients, the difference was 10%, but for low‐severity patients, there was no difference.

Several studies of specific procedures reported a survival advantage for major teaching hospitals (e.g., Konda et al. ⁴¹ ). But in multivariate analysis, Dimick et al. found no difference in in‐hospital mortality for esophageal, hepatic, and pancreatic resection. ⁴² This study did little to control for differences in patient health at admission in only including among health covariates binary variables for the number of comorbidities and for whether the admission was “urgent” or an “emergency.”

Only one published study measuring effects of hospital teaching status on outcomes, Bekelis et al., ⁴³ used an instrumental variable (IV) to deal with endogeneity of teaching. Endogeneity may arise in this context if patients with certain unspecified attributes such as poor health on dimensions not explicitly known by and/or considered by the analyst lead patients to select one type of hospital over another.

The IV was the percentage of teaching hospitals in the county. Experiments with alternative definitions of “teaching” did not materially affect the findings on teaching. This IV was strong (partial F = 125). While, in principle, the IV could be excluded from the main equation, it is likely to have been correlated with area variables that the authors should have included in the main equation, and these area variables are likely to have been correlated with the IV. If so, the study's IV strategy is not valid.

The unadjusted difference in the in‐hospital post‐neurosurgical death rate between teaching and nonteaching hospitals was 0.4% higher for teaching hospitals. Using the IV and accounting for other factors, the difference was 0.3% higher for teaching hospitals. The difference was robust to use of different statistical methods (propensity‐score matching and mixed [random] effects). The authors offered several possible reasons for the worse mortality outcomes of teaching hospitals, including lack of experience of residents and limited supervision of residents by faculty.

The most econometrically sophisticated study of mortality outcomes is Geweke et al.’s study of mortality of patients over age 65 admitted with a primary diagnosis of pneumonia to one of 114 hospitals in Los Angeles, five of which were classified as “private teaching.” ⁴⁴ The definition of a teaching hospital was not provided. Public hospitals were defined as those operated by L.A. County (a very narrow definition of a “public” hospital). The data were hospital discharge abstracts distributed by the State of California. Discharge abstracts do not contain information following discharge from the hospital. The intellectual contribution of the study is in the econometric technique for removing heterogeneity in hospital and patient attributes not observed by the researcher. Mortality was measured by whether or not the patient was alive ten days following admission. Applying their methodology, the authors concluded that private teaching hospitals and small and large hospitals offer higher‐quality care than other hospitals. No explanation was provided for why small hospitals had better survival experiences than medium‐size hospitals. Teaching hospitals’ better performance in treating pneumonia is consistent with Silber et al.’s finding ³⁶ in spite of the appreciable differences in study methods, sample, and follow‐up periods.

Other Outcomes. Silber et al.’s study of surgical outcomes reported that intensive care unit use was lower in major teaching than in nonteaching hospitals, ³⁷ as was the readmission rate at 30 days post initial admission. However, length of stay was higher at the former. In‐hospital complications were identified based on diagnostic and procedure codes available in Medicare claims; failure to rescue was defined as death following an in‐hospital complication. Comparing the experience of hospitals with an RTB ratio of 0.6 (a value well above the threshold for major teaching hospitals in most other studies) to nonteaching hospitals, for failure to rescue, the odds ratio was 0.85 (95% CI: 0.84‐0.86), a substantial difference. Moreover, the difference became larger as the predicted probability of patient death at admission increased.

Using patient safety indicators developed by the Agency for Healthcare Research and Quality in multivariate analysis, Vartak et al. documented rates of postoperative complications according to teaching intensity, finding that rates were higher for pulmonary embolism, deep vein thrombosis, and sepsis; lower for respiratory failure; and no difference for postoperative hip fracture, hematoma or hemorrhage, and physio‐metabolic derangement (Table 4). ⁴⁵ With the same data, Zafar et al. documented slightly higher rates of inpatient mortality in teaching hospitals, but slightly lower rates of major complications after receipt of emergency general surgery. ⁴⁶ Thornlow and Stukenborg found no differences by hospital teaching status in preventable complications. ⁴⁷

Table 4.

Other Outcome Studies

Authors	Description	Definition of Teaching Hospital	Sample	Key Findings
Vartak et al., 2008 45	Analyzed Nationwide Inpatient Sample (NIS) merged with American Hospital Association annual surveys. Dependent variables were Patient Safety Indicators (PSIs)—6 postop PSIs: hip fracture; hematoma; physio‐metabolic derangement, respiratory failure; pulmonary embolism; deep vein thrombosis; sepsis.	MT: COTH member OT: Not member of COTH but RTB > 0.	646 hospitals	Adj. After controlling for bed size, staffing, case mix, and other patient risk factors, MTs had significantly higher odds of postop pulmonary embolism and sepsis than NTs, but lower odds of respiratory failure with no statistical difference for other PSIs.
Zafar et al., 2014 46	Used NIS to identify patients with emergency surgical conditions. Used propensity‐score matching (PSM). To measure cost, total charges were multiplied by hospital cost‐to‐charge ratio.	Combined MT and OT. Teaching hospitals defined on basis of receipt of direct or indirect Medicare GME. Payments.	3.7 million patients	Adj. After PSM, teaching hospitals had slightly higher odds of inpatient mortality (OR = 1.04; 95% CI: 1.02‐1.06), slightly lower odds of major complications (OR = 0.99; 95% CI: 0.98‐0.99), and slightly lower length of stay (5.03; 95% CI: 4.98‐5.09 vs. 5.22; 95% CI: 5.16‐5.29).
Thornlow and Stukenborg, 2006 47	Used NIS. 5 complications/adverse events related to patient safety defined: in‐hospital deaths for low‐mortality diagnosis‐related groups (DRG); failure to rescue; selected infections caused by medical care; postop hemorrhage/hematoma; and postop respiratory failure. Elixhauser Index used to control for comorbidities.	MT: COTH member and RTB ≥ 0.25; OT: Not included.	903 hospitals	Hospital ownership and teaching status not consistent predictors of preventable events.
Chen et al., 2018 48	Using Medicare claims and Denominator File, studied variations in serious complications: high length of stay; failure to rescue; readmissions within 30 days post discharge; and Medicare payments for beneficiaries undergoing hepato‐pancreatic‐biliary surgery (HPB). Primary outcome, Medicare payments.	MT: RTB ≥ 0.363; OT: 0 < RTB < 0.363.	8,863 hospitali‐zations for HPB surgery	Comorbidities did not systematically vary by teaching status. Compared to NT, MT hospitals had lower risk‐adjusted rates of serious complications (RR = 0.8; 95% CI: 0.7‐1.0) and 30‐day mortality (RR = 0.5; 95% CI: 0.4‐0.8).
Sandhu et al., 2013 49	Assessed impact of teaching status on outcomes of patients undergoing percutaneous coronary intervention in Michigan. Used data from registry containing baseline clinical, demographic, procedural, angiographic, and medication data and procedural and in‐hospital outcomes.	MT: COTH member and RTB ≥ 0.25. Data from 7 allopathic and 2 osteopathic hospitals included. OT: Not included.	89,048 patients	Adj. After risk adjusting for many factors, including clinical findings, study found no difference in in‐hospital death, in‐hospital AMI, contrast‐induced nephropathy, or gastrointestinal bleeding between MT and NT. MT had lower rate of coronary artery bypass grafting (OR = 0.63; 95% CI: 0.49‐0.83) but increased rate of vascular complications (OR = 1.33; 95% CI: 1.21‐1.41; 95% CI: 1.21‐1.46).
Pradarelli et al., 2017 50	Using Medicare claims and Denominator File, study assessed Medicare payments and perioperative outcomes for persons undergoing abdominal aortic aneurysm repair, pulmonary resection, or colectomy.	5 teaching hospital categories ranging from “very major” (RTB > 0.599) to “very minor” to NT.	442,077 hospitalizations	“Very major” teaching hospitals generally had higher risk‐adjusted rates of complications, but lower risk‐adjusted rates of failure to rescue and 30‐day mortality than did NT hospitals.
Horwitz et al., 2018 51	Using Medicare claims data, study assessed each hospital's “excess days in acute care” (EDAC) for beneficiaries hospitalized for AMI, CHF, and pneumonia. EDAC was the difference between predicted and expected total days of acute care use (including emergency room visits, observation stays, unplanned admissions) in the 30 days post discharge/100 discharges. EDAC = 0 indicates no difference in performance of individual hospitals from average hospital. EDAC > 0 indicates individual hospitals performed poorly relative to average hospital; and conversely for EDAC < 0.	MT: COTH member; OT: Non‐COTH hospitals with RTB > 0	2,164 hospitals for AMI; 3,720 for CHF, and 4,195 for pneumonia	In multivariate analysis, for all 3 study conditions, all else being equal, major teaching hospitals had positive EDACs. Minor teaching hospitals had a negative EDAC for CHF, but there was no difference between OT and NT for the other 2 conditions.

Abbreviations: adj., adjusted for other factors in multivariate analysis; CI, confidence interval; HR, hazard ratio; MT, major teaching hospital; NT, nonteaching hospital (omitted reference group); OT, other teaching hospital; RTB, resident‐to‐bed ratio; unadj., unadjusted for other factors.

Studies of other procedures, hepato‐pancreato‐biliary (HPB) surgery ⁴⁸ and percutaneous coronary interventions (PCI), ⁴⁹ reported lower rates of complications for HPB (and lower 30‐day mortality), but higher rates of complications for PCI. In analysis of neurosurgery outcomes, Bekelis et al. reported higher rates of discharge to (nonhospital) facilities from major teaching hospitals. ⁴³ In Pradarelli et al.’s analysis of hospitalizations for abdominal aortic aneurysm (AAA) repair, pulmonary resection, or colectomy—all three high‐risk procedures—complication rates were more common in very major than in nonteaching hospitals, but the former had higher rates of rescue from these complications, which led to lower 30‐day (post‐discharge) mortality. ⁵⁰ All analyses in the Pradarelli et al. study included binary explanatory variables for basic demographic characteristics, 29 Elixhauser comorbid diseases, procedure year, and surgical approach (e.g., open versus minimally invasive).

On readmission rates, in multivariate analysis of readmission patterns for patients initially hospitalized for AMI, CHF, or pneumonia, Horwitz et al. found that major teaching hospitals had relatively more readmission days—including emergency room visits (counted as one‐half readmission day), observation days, and unplanned readmissions—than nonteaching hospitals. For minor teaching hospitals, readmission days were lower statistically than for nonteaching hospitals for CHF, but there was no statistical difference for the other two study conditions.

Cost

Different Perspectives of Cost

From a societal perspective, cost reflects minimum amounts of resources: labor, physical capital, land, and other inputs required to produce a unit of output. Determinants of hospital cost from a societal perspective are measured from a cost function. Higher cost may reflect a combination of more slack, ⁵² , ⁵³ provision of public goods/community benefits, ⁵⁴ case mix differences (measured and unmeasured), and higher care intensity.

From the consumer's perspective, out‐of‐pocket cost is most relevant. From the vantage point of insurers, public or private, cost per unit of service is the price it pays for that service. To the extent that production of a higher‐quality service costs more to produce, the insurer may expect to pay more, but in the end, the price facing insurers reflects a combination of consumer willingness to pay for the higher‐quality service, the extra cost of producing it, and relative bargaining power of sellers and buyers.

Societal Perspective

A few studies of hospital costs, published between 1981 and 1997 (e.g., Grannemann et al. ⁵⁵ and Carey ⁵⁶ ), have estimated parameters of a structural hospital cost function from hospital cost reports. These structural models use a flexible functional form with total hospital cost (often the logarithm of such cost) as the dependent variable and hospital input prices and hospital outputs as the only covariates. These models are cumbersome in containing many covariates for interaction terms, and almost all have not directly measured the effect of teaching. Several hospital cost papers have used a behavioral approach not based on an explicit economic model. ⁵⁷ , ⁵⁸ , ⁵⁹ These papers have included covariates hypothesized to be related to cost, including for teaching. This research is subject to the criticism that the cost functions are ad hoc, that is, not based on economic theory.

Sloan et al. estimated hospital cost functions based on a sample of 367 hospitals (Table 5), concluding that hospital cost was at most 20% higher in major teaching than in nonteaching hospitals; the percent difference between minor teaching and nonteaching hospitals was about half this. ⁵⁹ Examining cost differences by cost center, there was no teaching‐related difference in cost for radiology, but pathology cost was higher at major teaching hospitals. Cost per adjusted patient day for major teaching versus nonteaching hospitals was significantly higher for pharmacy (12.6%), dietary (12.1%), plant operations (25.3%), housekeeping (12.2%), nurse administration (12.7%), and medical‐surgical nursing (5.6%). Higher plant operations cost, and consequently higher housekeeping cost, may have been partly attributable to space allocated to the residency programs. Higher cost of nurse administration may reflect the presence of nursing school. Higher pharmacy and dietary cost may have reflected unmeasured case mix severity, but higher cost also could have been due to added bureaucracy in major teaching hospitals. This study gave some look “under the hood” and supports a focus on an input package associated with GME (Figure 1) rather than on the hospital's RTB ratio.

Table 5.

Hospital Cost Studies

Authors	Description	Definition of Teaching Hospital	Sample	Key Findings
Sloan et al., 1983 59	Used data from the American Hospital Association (AHA)—annual surveys, Hospital Administrative Services, and Survey of Medical Staff Organization—and Commission on Professional and Hospital Activities to study variation in total and components of hospital cost by teaching status. Case mix measured by Resource Need Index, a proprietary index used in the early 1980s.	MT: COTH member; OT: Not COTH member, but with residents.	367 hospitals, 1974‐1977	After accounting for case mix, bed size, wages, and census area, hospital cost was at most 20% higher in MT than in NT hospitals; OT‐NT cost was about half this. No teaching‐related difference in radiology cost, but MT cost was higher for pathology. Cost per adjusted patient day for MT vs. NT was significantly higher for pharmacy (12.6%), dietary (12.1%), plant operations (25.3%), housekeeping (12.2%), nurse administration (12.7%), and medical‐surgical nursing (5.6%).
Garber et al., 1984 60	Compared care cost at a single hospital with separate teaching and community facilities. Considered cost differences (measured by billed charges) overall and by predicted probability of death. Sample divided into quartiles based on predicted death probability; case mix index based on Medicare DRG weights.	—	1 hospital: Stanford	Overall, case‐mix‐adjusted teaching‐community facility difference in charges was 10.8% (95% CI: 3.8‐11). For the lowest risk of death group, no statistical difference in cost between 2 facility types. But for the highest‐risk group (patients with a predicted probability of death ≥ 0.25), there was a 70% difference in cost (95% CI: 33–101). Among 3 cost components—diagnostic, routine (e.g., room/board), and treatment—largest cost difference was in diagnostic component.
Doyle et al., 2010 61	Used data from a Veterans Administration (VA) hospital with residency programs affiliated with two medical schools, one higher ranked than the other. Each residency program was affiliated with another teaching hospital in the city. Both programs had access to the same facilities and ancillary staff. Unique aspect of the study: patients randomly assigned to one of the two clinical teams linked to residency programs.	Programs being compared were both affiliated with same VA hospital.	30,000 patients randomized	Patients assigned to medical team at higher‐ranked medical school had 10%‐25% less expensive stays than those assigned to medical team at lower‐ranked school. Health outcomes not related to team assignment. Cost differences were largest for patients with most serious conditions. Largest single cost difference was for diagnostic. Lower‐ranked program ordered more tests, took longer to order them.
Grannemann et al., 1986 55	Used 1981 AHA annual survey and 1982 Mathematica Ambulatory Care Survey to estimate parameters of multiple‐output cost function. Key covariates for various measures of hospital output (and their interactions), factor prices, hospital attributes, and region. No case mix index but case mix captured indirectly by hospital outputs.	MT: COTH member. Also included a binary for medical school affiliation.	867 hospitals	COTH members had 16% higher total cost than others; medical school affiliation binary, n.s.
Carey, 1997 56	Used data from AHA annual survey and Medicare Cost Reports for 1987‐1991 to estimate multiple‐output cost function. Dependent variable = total variable cost (log). Case mix index based on Medicare DRG weights. Alternative estimators employed.	MT: “Heavy teaching” = COTH member and affiliated with medical school; OT: “Light teaching” = affiliated with medical school but not COTH.	1,739 hospitals	Coefficient on heavy teaching uniformly statistically significant when entered as covariate, but coefficient for light teaching n.s. From ordinary least squares, implied effect of heavy teaching on short‐run hospital cost was in the 13% to 16% range.
Pettingill and Vertrees, 1982 62	Using Medicare Cost Report data for 1979, estimated effect of raising RTB (log) on hospital cost per admission (log), given case mix (DRG weights), local wage rate, bed size, and city size.	RTB covariate for teaching	5,071 hospitals	Coefficient on RTB = 0.569 (p < .001). A teaching hospital with an RTB = 0.25 would have cost per admission 17.8% higher than an NT hospital (authors’ calculation).
Dalton et al., 2001 66	Used Medicare Cost Report and related data to replicate Pettingill and Vertrees (1982) study with panel data, 1989‐1995. Estimated separate cross‐sectional equations by year and also pooled data with fixed effects.	RTB covariate for teaching.	About 5,000 hospitals	Fixed effects regression did not show that within‐hospital changes in hospital RTB were positively related to cost per admission. Inferred that measured effect of RTB on cost due to hospital and patient factors was not captured by covariates for case mix and local area wages.

Abbreviations: CI, confidence interval; HR, hazard ratio; MT, major teaching hospital; NT, nonteaching hospital (omitted reference group); OT, other teaching hospital; RTB, resident‐to‐bed ratio; unadj., unadjusted for other factors; n.s., p‐value not ≤ 0.05.

Two single‐hospital studies yield important insights about teaching‐related cost differences. In the first, Garber et al. compared costs of care of patients admitted to the hospital at a faculty (staffed by physicians with faculty appointments) versus the community service (staffed by physicians without faculty appointments) facility, both affiliated with Stanford University. ⁶⁰ Comparisons were based on billed charges. Since billing practices were the same for this single organization, the usual critique of using charges to measure resource use does not apply here. Overall, the case‐mix‐adjusted difference in charges was 10.8%. The authors divided the sample into four groups depending on their patient's expected probability of death at admission. For the group with the lowest risk of death, there was no statistical difference in cost between the two facilities. But for the highest‐risk group, there was a 70% difference. Among three cost components—diagnostic, routine (e.g., room and board), and treatment—the largest cost difference between the two facilities was for diagnostic services.

In the second study, Doyle et al. examined differences in costs and outcomes between two clinical teams, each associated with a residency program. ⁶¹ One of the teams was affiliated with a highly ranked medical school. The other team's medical school was ranked lower. What is unique about the study is that 30,000 patients were randomly assigned to one of the two programs. Thus, case mix severity was the same, as were hospital attributes, since both teams were affiliated with the same Veterans Administration hospital. Patients treated by physicians at the higher‐ranked program had 10%‐25% less expensive stays, with the cost difference largest for diagnostic testing. Physicians at the lower‐ranked program ordered more tests and took longer to order them.

Several studies have estimated a multiple‐output hospital cost function. Grannemann et al. is probably the best known and is illustrative of these studies. ⁵⁵ Among the many covariates in this study was a binary variable for COTH membership. Minor teaching and nonteaching hospitals were presumably the omitted reference group. The implied effect of being a COTH member was a 16% increase, all else being equal, in total cost. More recently, Carey analyzed determinants of total variable cost. ⁵⁶ Implied effects based on parameter estimates for major teaching hospitals, always statistically significant, ranged from 13% to 16%. Parameter estimates for minor teaching hospitals were always statistically insignificant.

One TEFRA provision was a special payment for the indirect cost of GME. This payment was intended to compensate unmeasured differences in case mix severity not accounted for by diagnosis‐related groups (DRGs). The indirect cost subsidy formula was based on Pettingill and Vertrees's analysis of variation in hospital RTB ratios on hospital cost. ⁶² The dependent and explanatory variables for RTB were expressed in logarithms. However, since RTB = 0 applied to most hospitals, the authors arbitrarily added one to each hospital's RTB ratio. The parameter estimate on RTB ratio was 0.569 (p < 0.001). Based on this parameter estimate, the authors inferred that a hospital with an RTB = 0.25 would have a cost per admission 17.8% higher than a nonteaching hospital would. Congress initially doubled the differential implied by Pettingill and Vertrees's result. This large subsidy incentivized hospitals to hire more residents and increase their Medicare patient loads. ⁶³

Pettingill and Vertrees's analysis received a substantial amount of criticism. ⁶⁴ , ⁶⁵ , ⁶⁶ One major statistical problem involved the arbitrary addition of a constant (1) to the RTB ratio. Results were shown to be highly sensitive to the level of the constant added. Moreover, when hospital fixed effects were included, Dalton et al., using panel data, showed that intertemporal within‐hospital variation in the RTB ratio did not affect hospital cost. ⁶⁶ If unobserved case severity, which in turn affects hospital cost, changes with changes in hospitals’ RTB ratios, this study provided no empirical support for this.

In sum, at one end of the spectrum, there is the implication from the Doyle et al. study that, if anything, a residency program supervised by a skilled physician team is more efficient. At the other end are estimates that the cost of a major teaching hospital is about one‐fifth higher than a nonteaching hospital. This difference reflected differences in cost components indirectly related to clinical care, such as housekeeping and plant operations, and to other clinically oriented programs such as nursing education. The finding from panel data emphasizes the deficiency in basing GME subsidies on results from analysis of a single cross section (as done by Pettingill and Vertrees).

Medicare's Perspective

When initially implemented, Medicare's switch from retrospective cost reimbursement to PPS was to be budget neutral. But decades have elapsed since PPS was implemented. PPS provided a financial incentive for hospitals to reduce cost. ⁶⁷ Also, there have been many annual updates in Medicare payment rates; subsidies, such as for disproportionate share, have been added, and teaching subsidy formulas have been modified. Reports of the Medicare Payment Advisory Commission indicate that Medicare payments are below cost, although such negative margins are average rather than marginal calculations. ⁶⁸ , ⁶⁹ Medicare prices, excluding the subsidies, may cover the cost of the marginal Medicare hospitalized patient.

Medicare covers not only hospital care, but also physicians’ and other professionals’ services, some nursing home care, home health and hospice care, and durable medical equipment. Thus, from Medicare's perspective, it is not only the cost of inpatient care that is germane. It is possible that the extra cost of inpatient care is partially offset by savings in expenditures on other services.

Taylor et al. computed Medicare cost for the index admission and total Medicare program cost for the six months following an index hospital admission in constant dollars. ³⁵ Such cost included not only payments for the index hospital admission but also hospital payments: for readmissions (within six months); to physicians—during the index admission and following hospital discharge from the hospital; and to skilled nursing facilities, hospice, home health care, and durable medical equipment. Pairwise comparisons between hospitals based on commitment to teaching in unadjusted total payments were generally insignificant. Among the four patient samples (CHF, CHD, stroke, hip fracture), the exception was for hip fracture. Payments for the index admission for hip fractures were significantly higher for major teaching hospitals as compared to FP hospitals (p < 0.001), the omitted hospital reference group, as were total Medicare payments (p = 0.05); but after adjustment for patient characteristics and Medicare subsidies, there was no statistical difference between major teaching and FP hospitals in total six months’ Medicare payments for any of the four samples. Thus, even though inpatient stays at major teaching hospitals were much more costly to Medicare, viewing payments over a six‐month period following admission, there was no cost difference.

Burke et al. used much more recent Medicare data and a much larger sample but a much shorter follow‐up period: 30 days and, alternatively, 90 days post admission. ⁷⁰ Nevertheless, the findings reinforce those of Taylor et al. The primary cost outcome was 30‐day adjusted Medicare payments in the aggregate and separately for medical conditions and surgical procedures. Total adjusted payments were $18,605 for major teaching, $18,793 for minor teaching, and $18,873 for nonteaching hospitals. The difference between major and nonteaching hospitals was −$268 (95% CI: −$436 to −$80, p = 0.005)—i.e., it was higher for nonteaching hospitals. For the index admission, payments were $8,529, $8,370, and $8,180 at major, minor, and nonteaching hospitals, respectively, the difference between major teaching and nonteaching hospitals being $349 (95% CI: $308 to $390, p < .001). These small differences (<5%) in payment disappeared using a 90‐day follow‐up period.

Silber et al. computed three measures of cost for the index hospitalization to study cost of beneficiaries admitted for medical conditions ³⁶ and surgical procedures ³⁷ during 2012‐2014. The first measure was for 30‐day “resource utilization‐based cost.” Such cost included (1) index hospitalization cost, (2) costs of emergency department, outpatient, and office visits within 30 days of admission, (3) and all cost attributable to readmissions within 30 days of the index admission. The resource cost for inpatient care is conceptually close to the dependent variables used in the cost studies described earlier (societal perspective). However, the cost was not hospital/year specific, but rather calculated for specific medical diagnoses and surgical departments. Thus, the authors needed to attribute cost to specific uses. Rather than use the hospital's cost‐to‐charge ratio for this purpose, the authors subdivided hospital stays into critical care days, step‐down unit days, and noncritical care days for each admission. Then values for each index hospitalization were assigned to each type of days. While the composition of critical care, step‐down unit days, and noncritical care days varied among hospitalizations (and among hospitals), a common set of price/cost multipliers was applied to these days. The price/cost multipliers did not vary by hospital teaching status. Physicians’ services were priced using 2014 Relative Value Units, which were converted to dollar amounts. The second measure represented total payments by Medicare, and the third, total payments less geographic factor price adjustments and subsidies for GME and disproportionate share.

Given the substantial sample sizes, all differences were statistically significant at conventional levels. Magnitudes of differences are much more interesting. In the study of medical admissions, propensity‐score‐matched differences between major teaching and nonteaching hospitals were 1.4%, 22.6%, and 9.9% for the resource cost, payment measure including subsidies, and payment measure exclusive of subsidies, respectively. The payment measure inclusive of subsidies was the notable outlier.

These results are striking. The same methodology was used to obtain the 1.4% as the 22.6% and 9.9% differences. Thus, if the 22.6% estimate of the difference in cost to Medicare between major teaching and nonteaching hospitals is upward biased, so is the 1.4% resource cost difference.

In the surgery study, on the first measure, resource cost, differences between major teaching and nonteaching hospitals were small. As a percent of nonteaching cost, the propensity‐score‐matched differences ranged from 3.3% for general surgery to 7.4% for orthopedic surgery. On the second measure, differences were large; they ranged from 28.7% for general surgery to 29.7% for orthopedic surgery. On the third, the differences were between 7.5% for orthopedic surgery and 9.3% for vascular surgery.

The differentials for the first and third measures were below cost function‐based estimates for major teaching hospitals. By contrast, those for the second measure were well above them.

Two recent studies of specific surgical procedures revealed larger differences even when the subsidies were excluded from the Medicare payment measure. Given these studies focused on less common surgical procedures than the studies by Burke et al. and Silber et al., sample sizes were much smaller. Pradarelli et al. compared total Medicare payments made on behalf of Medicare beneficiaries who underwent AAA repair, pulmonary resection, or colectomy during 2009‐2013. ⁵⁰ They divided the sample into admissions at very major (RTB ratio > 0.599), major (0.250‐0.599), minor (0.050‐0.249), very minor teaching (>0‐0.049), and nonteaching hospitals. The study included all Medicare Part A and B payments. Payments were computed for an episode of surgery that extended to 30 days past the index hospitalization's discharge date. Risk‐adjusted total Medicare payments were higher for very major teaching than for nonteaching hospitals: $45,570 versus $31,426 for AAA, p < 0.0001); $39,550 versus $29,550 for pulmonary resection; and $51,893 versus $32,746 for colectomy. However, after eliminating GME and disproportionate share subsidies, the only statistical difference in Medicare payments that remained was for colectomy: $34,949 versus $30,352, p < 0.001). Relative to risk‐adjusted payments less the social subsidies for nonteaching hospitals, the $4,597 difference in payments represented a 15.1% increase.

Chen et al. assessed Medicare payments for the index hospitalization made on behalf of beneficiaries undergoing HPB surgery during 2013‐2015. ⁴⁸ The authors computed alternative payment measures: actual; risk‐adjusted; and “price‐standardized” (adjusting for the social subsidies). Major teaching hospitals were those with RTB ratios above the US median. Actual payments were relatively high for major teaching hospitals: major teaching, $29,075; minor teaching, $22,026; and nonteaching, $19,983 (all differences between major teaching and nonteaching hospitals, p < 0.001). Risk adjusting increased the difference in total payments between major teaching and nonteaching hospitals to $10,195 from $9,092. Excluding the social subsidies reduced this difference to $8,623. Relative to the mean payment for nonteaching hospitals, this difference represented a 43.6% increase in payments!

In sum, except for the two last studies, differences in Medicare payments exclusive of GME and disproportionate share subsidies between major teaching and nonteaching hospitals tended to be lower in percentage terms than the cost studies reviewed above revealed. However, with Medicare subsidies and factor price adjustments included, differences in Medicare payments between major teaching and nonteaching hospitals were much higher than cost differentials from the cost studies. One reason for the larger differences in payments in the last two studies is that in the first of these, the comparison was between very major and nonteaching hospitals. In the second, the RTB ratio threshold for “major teaching” was higher than in most other studies, such as in Silber et al. ³⁷ , ³⁸ Some differences reflect outlier payments to the hospital groups most involved in teaching. Chen et al. cautioned that “payment policies, such as prioritization of ‘value‐based’ payment methods at high teaching intensity hospitals, should be carefully considered in light of these [their] data.” ^{49 (p2978)} Taken at face value, these much higher payments would not appear to reflect “value.”

Private Payer Perspective

Prices private insurers pay hospitals reflect a combination of costs of care incurred by hospitals (sometimes approximated by Medicare benchmarks): insurers’ willingness to pay, which presumably reflects consumers’ willingness to pay; and relative bargaining power of insurers and hospitals. Thus, prices paid to major teaching hospitals may exceed prices paid to nonteaching hospitals, not only because the cost of care is higher in the former, but also because the purchasers of care are willing to pay them more.

There is virtually no empirical analysis of differences in prices paid by private insurers to hospitals by teaching status. An exception is Cooper et al., who used claims data from the Health Care Cost Institute for the period 2008‐2012. ⁷¹ Pricing of teaching hospitals was not a focus of this paper; relevant findings on this subject come from an unpublished appendix.

The study included a covariate for “teaching hospital,” without defining “teaching.” Judging from the high percentage of admissions to teaching hospitals in this study (38%), “teaching” must have included major and minor teaching hospitals. Hospitals rated “best” during 2008‐2012 by U.S. News & World Report constituted 5.3% of the analysis sample. Given evidence from a previous study, ²³ it is likely that the best hospitals were major teaching hospitals, and the covariate for teaching hospital picked up the influence of minor teaching hospitals.

The authors ran regressions with the logarithm of the facility price and, alternatively, the sum of the physician price and the facility price/admission as dependent variables. Covariates included hospital product market competition, case mix (Charlson Comorbidity Index), ownership, bed size, a technology index, and county characteristics. With the log of facility price as the dependent variable, the parameter estimates on “best hospital” had effect sizes ranging from 13% to 15%. With the physician price included, the implied effect of best hospital was 12%. Parameter estimates on “teaching hospital” were statistically insignificant. Thus, the price differentials between major teaching and nonteaching hospitals were in the range of the differences in cost reported earlier in this paper. Private insurers’ prices tend to be much higher than those paid by public payers.

Discussion

The literature on quality and cost of teaching hospitals is vast. These are the major findings from this review.

First, most findings on process of care indicate higher quality of care at major teaching hospitals. Adverse events or complications due to medical care may be more likely at such hospitals, but they tend to be more effective in handling them. At least for AMI, there is empirical evidence that major teaching hospitals outperform others in adherence to guidelines, especially for low‐tech care, which all hospitals are capable of adopting. Process of care studies are descriptive, but they focus on whether care was appropriate conditional on the severity and other attributes of illness, based in part on clinical findings. Thus, since they explicitly take account of case mix, process of care studies are not nearly as subject to the patient selection issue of other studies—i.e., sicker patients (in ways researchers cannot observe) selecting higher‐quality hospitals and as a result, researchers not fully adjusting for case mix differences.

Second, major teaching hospitals clearly outperform others in terms of survival of persons diagnosed with various cancers. Results are strong particularly because (1) survival is a highly relevant outcome for most cancers, (2) the studies have lengthy follow‐up periods, and (3) the researchers included detailed controls for patient severity of illness at initial cancer diagnosis. Except for AMI, none of the outcome studies come near in characterizing attributes of the diagnosis as those for cancers. Results on survival after AMI by hospital teaching status, using the comprehensive CCP database, are mixed, being sensitive to particulars of equation specification.

Third, several studies analyzed short‐term (30‐day) mortality outcomes for several medical conditions and surgical procedures, finding that major teaching hospitals’ patients experience lower mortality. Results are surprisingly consistent—a 1.2% and a 1.3% reduction in mortality for major teaching versus nonteaching hospitals observed for different sets of conditions/procedures in studies by different research teams. In spite of detailed propensity‐score matching by one research team, the data came from Medicare claims and thus, in contrast to the cancer studies and CCP data on AMIs, important attributes of the individual patients’ severity of illness may be missing.

Khwaja and coauthors used CCP data to assess the importance of having severity of illness data for matching. ⁷² More specifically, they estimated a structural model of hospital choice and catheterization choice with the CCP data—data that contained medical record data on patient heterogeneity. Using the estimated structural parameters, they simulated data for which the treatment effect (catheterization) is known. With the data on patient severity of illness, matching estimators provided accurate estimates of the treatment effect. However, this was not so when matching was not based on patient illness severity and the types of measures available from claims data. Thus, while not fully rejecting the findings that major teaching hospitals are effective in lowering short‐term mortality, especially for patients with relatively high mortality risk at admission, it is important to recognize that the results might change with better measures of patient illness severity. Khwaja et al. did not assess the accuracy of specifications that include a covariate for the treatment but do not include covariates for illness severity—that is, do not match control with treatment groups. But such analysis almost surely suffers from omitted variables bias.

Fourth, several studies focused on single illnesses or procedures. Several of these studies have had shorter follow‐up periods and/or included hospitals from a single location. A sophisticated econometric study, ⁴⁴ original in its approach to removing omitted heterogeneity, of in‐hospital mortality of pneumonia patients in Los Angeles found favorable results in reducing mortality by the tenth day following admission for the five teaching hospitals in its analysis sample. It would have been useful to have seen a head‐to‐head comparison of results using the econometric technique with results from data containing more detail on clinical findings at patient admission. However, other studies focused on specific procedures found no difference in mortality, or the results favored nonteaching hospitals. There is a question whether such studies have much external validity. It seems doubtful that we can generalize about the relative quality of teaching hospitals from such studies.

Fifth, health outcomes other than mortality have been much less frequently studied. Among the outcomes are readmission rates, length of stay, complication rates, and rates of failure to rescue from complications resulting in patient death. Overall, the results are mixed.

Sixth, Medicare's indirect medical education subsidy has been based on the hospital's RTB ratio since the early 1980s. Even after four decades, there is no consensus about the indirect cost of GME. First, the cost in some cost centers is at most remotely related to medical education (e.g., housekeeping and plant operations). Such cost also varies by hospital teaching status. Nursing may be more costly for reasons other than the presence of a GME program. Nurse staffing too is related to outcomes of hospital care (e.g., Aiken et al., ⁷³ Needleman et al. ⁷⁴ ), which calls the estimates of variation in the RTB ratio on outcomes into question. Without adequately accounting for other programs commonly found in major teaching hospitals, estimates of RTB ratio effects on hospital outcomes and costs are plausibly overstated. Second, one study used panel data to assess within‐hospital changes in hospital cost as the RTB ratio changes. Intertemporal changes in the RTB ratio did not affect hospital cost. The single cross‐section differences in cost attributable to changes in the ratio plausibly reflects unmeasured differences in hospital and patient attributes, only some related to provision of GME.

Seventh, given the various social subsidies, Medicare pays appreciably more for care at major teaching hospitals. Fortunately, the higher price of inpatient care appears to be offset at least in part by reductions in payments on behalf of Medicare beneficiaries post hospital discharge.

Given these findings, what are the implications for private (health insurer) and public policy?

A place to start is with consideration of the quality of the empirical evidence. Long‐term survival is quite relevant for cancer and other common chronic conditions. In such cases, private insurers would want to include hospitals with superior long‐term survival benefits in their networks. The empirical evidence in favor of major teaching hospitals on process of care is also compelling. At least at the time the process of care studies were conducted, major teaching hospitals were more likely to have performed diagnostic and therapeutic services appropriate to the patients’ underlying conditions.

Short‐term mortality is particularly relevant for risky procedures. It seems reasonable to surmise that more patients undergoing common orthopedic procedures, such as replacement of a hip, knee, or shoulder, are not primarily concerned with survival in the weeks following the procedure. Those who are at particularly high risk of death following receipt of the procedure are likely to forgo the procedure and hence are not in the data. Nevertheless, findings on process of care and several of the mortality studies, especially the cancer studies, are favorable to major teaching hospitals, which implies that private payers would want to include such hospitals in their networks, at least at some price.

Supposing that the differences in mortality are causally related to differences in hospital teaching status, there is the question of maximum patient willingness to pay for reductions in mortality. Silber and colleagues addressed this issue, ³⁶ , ³⁷ at least indirectly. They found, for example, that to achieve a 1% decrease in 30‐day mortality post admission, for general surgery and vascular surgery, the added costs per admission were $965 and $3,567, respectively. No computation was done for orthopedic surgery given they found no statistical difference in 30‐day mortality for such surgery. For total payments, the marginal program costs (per 1% reduction in mortality) per admission were $7,668 and $20,302. The total payments less the geographic factor price adjustment and the GME and disproportionate care subsidies were $2,186 and $6,141. Although differences in Medicare payment are documented, we are still left with the problem of determining maximum willingness to pay for a 1% reduction in 30‐ or 90‐day mortality risk. There are estimates of the value of a statistical life year. ⁷⁵ But can values for a year be extrapolated to a month, i.e., to one‐twelfth of a life year? It seems doubtful that an extra month of life, much of which is spent in the hospital and in post‐stay recovery, is worth much, but in the end, this is an empirical question yet to be addressed.

Public subsidies of GME should be guided by the principle that social benefit of such programs exceeds their cost. As the preceding example demonstrates, benefit valuation is lacking. One may argue, as Figure 1 illustrates, that improved patient outcomes, however measured, is not the only social benefit of such programs, and observed cost differences by GME teaching status are not all attributable to medical education. In addition, there is reason to doubt that observed quality differences only or even mainly reflect the hospital's RTB ratio; rather, major teaching hospitals may have more capable medical staffs or clinical faculty in being more capable diagnosticians, being more up to date in new clinical research findings, and being more adherent to practice guidelines.

Even if these serious problems are ignored, if the public shares in the benefit, all insurers and/or taxpayers should bear the program cost, not just Medicare. Reliance on a Medicare subsidy has at least two potential adverse consequences. First, if Medicare's price, inclusive of the indirect GME subsidy, exceeds marginal cost, teaching hospitals have a financial incentive to expand their residency programs. One study of this issue found that after Congress placed a limit on the subsidy, hospitals continued to increase the number of residents they employed. ² But this finding does not rule out hospital responses in the longer term. Second, the social subsidies boost Medicare prices relative to the prices paid by other payers. The price is further distorted because Medicare's subsidy covers the cost of other payers that do not pay for hospital activities such as GME. A profit‐maximizing firm in an imperfectly competitive product market, facing different demand curves from different consumer groups (in this context according to source of payment), ranks consumers according to their marginal revenue starting with the consumer generating the highest marginal revenue. ⁷⁶ If Medicare raises its price and hence hospital/marginal revenue accruing from a large number of patients, Medicare beneficiaries will have better access to hospitals. In response, other payers, attempting to preserve access of their insureds, are expected to also seek price increases. Under supplier‐induced demand, responses to an increase in the Medicare price are more complicated. ⁷⁷

As a recent editorial (the title of which implies there are scale diseconomies that have not been rigorously documented) put it, “Shifting to more closely realign the incentives of AMCs (academic medical centers) will not be easy. The clinical enterprises have become analogous to gigantic tankers, creating momentum to continue on the current course.” ^{78 (p204)} Even if policy change is feasible, there is no empirical basis that raising the RTB ratio by a unit would improve hospital quality.

Several limitations of the literature should be acknowledged. First, with one exception, findings are based on observational data. Second, the literature makes it difficult to make apples‐to‐apples comparisons. Studies differ in their definitions of major and minor teaching hospitals, in outcomes, in the length of follow‐up periods, in patient populations, and in observational periods. Any one of these differences could account for differences in patient outcomes. It would be helpful to have a generally accepted definition of a major teaching hospital. As it is, studies’ thresholds for “major teaching” vary from an RTB ratio of about 0.1 to 0.6.

Third, there is always a concern about measurement of severity of case mix. If case mix measures are insufficiently precise to capture true higher case mix severity of teaching versus nonteaching hospitals, quality differences between teaching and nonteaching hospitals will tend to be understated and cost differences overstated. Such biases could arise when a case mix index based on DRG weights is used to measure case mix severity, as in a few studies included in this review (e.g., Carey, ⁵⁶ Pettengill and Vertrees ⁶² ). The premise that case mix severity is higher in teaching hospitals, which many advocates for teaching hospitals maintain, has not always been supported empirically (see, for example, Allison et al. ²² ). Nevertheless, particularly given that such potential biases underlie the rationale for special public subsidies of indirect medical education, this concern should not be dismissed.

To be assured that patient populations are the same, ideally, quality and cost comparisons would be based on randomized data—that is, patients randomly assigned to teaching versus nonteaching hospitals. However, randomized data were only available for one study in this review. ⁶¹ Ideally, studies would be based on randomized data from a nationally representative sample of hospitals rather than a single setting as in the study I reviewed. But obtaining such data would be exceedingly costly. Second best is inclusion of risk adjustors from patients’ medical records or registries that are unavailable from claims data, as in the cancer studies and studies based on the CCP. However, maintaining registries and abstracting medical records is expensive; hence, such data are unavailable for most diagnoses and procedures. Third best is to eliminate unmeasured heterogeneity in case mix severity econometrically. However, this was done in only one study in this review, ⁴⁴ probably because of the approach's technical complexity. A much less complex approach involves the use of instrumental variables. ⁴³ Use of IVs is much more straightforward computationally, but IVs are only valid under specific circumstances. Fourth best is to perform the types of adjustments best exemplified by the Silber et al. studies. Two of their adjustments are most pertinent: (1) propensity‐score matching of hospitals by teaching status using many covariates; and (2) computation of resource cost, which allowed variation in input quantities to differ by hospital teaching status but with a set of price/cost weights that did not vary by hospital teaching status. Input quantities are far more likely to vary by patient case mix severity than are prices or costs per unit of service. Fifth best, and most widely used, is inclusion of various covariates available from administrative data and, if possible, data from general longitudinal surveys merged with the administrative data, to account for variation in case mix severity. Overall, the authors of the studies appear to have been aware of potential biases from imperfect risk adjustment, but some did a better job of risk adjusting than others.

Fourth, the studies are largely descriptive. Documenting differences in benefit and cost of teaching programs is important, but we lack important details as to why quality and cost varies by hospital teaching status. There is evidence from process of care studies that major teaching hospitals tend to be better in applying knowledge gained from clinical research. One study implies that better outcomes from teaching hospitals may come from the cognitive abilities of their physician teams, particularly in diagnosing. However, there is no decomposition of the relative importance of the factors potentially accounting for differences in productivity listed in Figure 1 in any study.

Nor do we have much empirical evidence on effects of public policies on the choices that teaching versus nonteaching hospitals make. One exception is a study of the effects of report cards on hospital behavior by Dranove et al. ⁷⁹ The authors demonstrated that nonteaching hospitals tended to reduce the share of high‐mortality‐risk cardiac patients they admitted in response to implementation of statewide report cards in New York and Pennsylvania, presumably to improve their ratings. By contrast, average severity of such patients admitted to teaching hospitals (defined as hospitals with at least 20 full‐time residents, a threshold that would include minor and major teaching hospitals) increased substantially.

The Balanced Budget Act of 1997 (BBA) substantially cut Medicare GME subsidies as well as implementing other cuts. ⁸⁰ , ⁸¹ The statutory change provided a natural experiment to study hospital behavior in general and teaching hospital behavior in particular. Chandra et al. analyzed the effect of the BBA on residents’ salaries and on the trend in the number of residency positions offered by hospitals. ² They found no effect on residents’ salaries and no change in the positive trend in number of positions post BBA. An earlier study provided evidence on the likely reason. ⁸² The study documented that at the margin, in spite of the Medicare cuts, hiring residents remained profitable. Other sources of profitability (from private payers, Medicaid) were not affected by the BBA. Other studies documented no change in process of care or in patient outcomes post BBA implementation in the short run. ⁸³ , ⁸⁴ In the longer term, some adverse effects of the BBA on patient outcomes have been documented. ⁸⁵ , ⁸⁶ Unfortunately, the outcome studies did not provide empirical evidence on teaching hospital behavior, since their focus was on comparisons between hospitals that received large versus small cuts under the BBA and Medicare policies implemented subsequent to the BBA. Teaching hospitals in the outcome studies were a control variable, although they tended to suffer larger cuts. The studies did not explicitly analyze whether teaching hospitals’ patient outcomes were more or less responsive to reductions in the teaching subsidy.

Still another potential example comes from the literature on insurer‐hospital bargaining and insurer network formation that has emerged, especially in the past five years. Unfortunately, these studies have not explicitly analyzed behaviors of teaching hospitals. Ho and Lee, in analyzing network formation, indicated three reasons that private insurers would want to exclude high‐cost hospitals (plausibly many major teaching hospitals) from their networks. ⁸⁷ One, sick consumers may decide against a plan that excludes a high‐cost, high‐quality hospital. ⁸⁸ Based on the empirical evidence, a rational cancer patient would eschew such plans. Two, relatively healthy persons might choose high‐cost, high‐quality hospitals, but if out of network, such patients will be steered to less expensive hospitals. And three, by actually excluding some hospitals or threatening to exclude them, the insurer may be able to obtain lower prices from the hospitals that remain. Empirically, we do not know the extent to which insurer‐hospital interactions actually affect major teaching hospitals’ case mixes and transaction prices.

Normatively, it would seem that steering patients to less resource‐intensive hospitals would be a win‐win proposition for insurers and consumers, who ultimately bear the cost, but not for the resource‐intensive hospitals. Results reported previously in this article suggest that patients newly diagnosed with cancer would be reassured by having access to a major teaching hospital; this is plausibly far less likely to be so for a patient undergoing common, non‐life‐threatening procedures.

No studies to date have assessed spillovers from teaching to nonteaching hospitals and the reverse. It is possible that since the process of care studies were conducted, nonteaching hospitals adopted best practices previously more likely to have been adopted by major teaching hospitals. Alternatively, there could be new best practices, and the nonteaching hospitals are perpetually behind in adopting them. Entry of FP specialty hospitals may force major teaching hospitals in the same market to adopt practices of the specialty hospitals, including marketing to patients who are less costly to treat.

Fifth, this review did not consider community benefit. Not‐for‐profit hospitals receive their tax advantages in return for providing community benefit. If teaching hospitals provide benefits other than patient care that the community values, this in addition to higher quality would be reflected in the maximum amount the community would be willing to pay teaching hospitals.

No estimates of community benefit have been presented in this study for the following reasons: First, community benefit is defined by what hospitals want to provide and report, at best the hospital's perception of what community members’ preferences are. Second, there is no consensus as to what constitutes community benefit. Even though the addition of Schedule H to US Form 990 in 2007 has led to more uniformity in items to be included in community benefit, empirical evidence on the relative contribution of teaching hospitals to such benefit is conflicting, ⁸⁹ , ⁹⁰ , ⁹¹ , ⁹² , ⁹³ possibly because of differences in equation specification among studies. Major teaching hospitals do provide a disproportionate (to their US market share) amount of charity care. Charity care is by far the largest single element of community benefit. Various public health programs by hospitals tend to be very small, measured by dollars expended by hospitals. ⁹³

In sum, whether or not major teaching hospitals give “more bang for the buck” is the wrong question. The question cannot be answered at this level of generality. Empirical evidence from process of care implies that quality of care is higher at major teaching hospitals on average. The empirical evidence suggests, under some circumstances, such as for cancer, outcomes are better at such institutions, but at what additional cost? Reports of differences in mean differences in performance obscure important underlying variation in patient health at admission. Some evidence implies that major teaching hospitals may be more effective than nonteaching hospitals in treating patients at high risk of imminent death. Thus, while it might be desirable to have such patients matched to major teaching hospitals, it should often be desirable to steer other patients to minor or nonteaching hospitals. The literature to date is insufficiently developed to provide much guidance on this issue. This will take more detailed empirical research on individual conditions to determine when referral to a major teaching hospital is advantageous. Provision of GME is associated with higher hospital cost; but the differences reflect several factors, not necessarily and not only the hospital's ratio of residents to beds. Even if there were a causal relationship between teaching intensity and cost, it is not advantageous for public payers such as Medicare to bear this cost alone. Rethinking public subsidies of graduate medical education is long overdue.