This Article Abstract Submit a response to this article Services E-mail this article to a friend Alert me to new issues of the journal

C H R O N I C  C A R E
P R O V I D E R  P E R F O R M A N C E
W E B  E X C L U S I V E   

7 October 2004  
  

Use Of Medicare Claims Data To 
Monitor Provider-Specific 
PerformanceAmong Patients With 
Severe Chronic Illness

Analyses of seventy-seven of America’s “best hospitals” document
extensive variation in the amount of care provided to
patients with three common chronic conditions

By John E. Wennberg, Elliott S. Fisher,
Thérèse A. Stukel, and Sandra M. Sharp

ABSTRACT:

This study illustrates that Medicare claims can be used to measure population-based, provider-specific rates of resource inputs, utilization, and Medicare spending. The target populations are seventy-seven cohorts of chronically ill Medicare enrollees who received most of their care from seventy-seven well-known U.S. hospitals. Striking variations are documented in resource inputs and use of services during the last six months of life. The patterns of care seen in the progression of chronic illness correlate highly with care received during previous periods. We believe that hospital-specific measures can be helpful in identifying providers with acceptable quality indices who are also relatively efficient in managing chronic illness.

The accelerated increase in Medicare per capita spending and evidence that effective care is underused and hospitals, intensive care units, and physician services are overused have focused attention on the need to improve the quality and efficiency of care for chronically ill Medicare beneficiaries. Several current federal initiatives promote disease management for enrollees in traditional fee-for-service (FFS) Medicare, some relying on disease management companies, others on the providers themselves. Provider-specific information on performance is key to these initiatives’ success. There is a growing list of quality indicators that can identify the underuse of effective care. However, illness-adjusted, provider-specific indicators to measure the possible overuse of care in managing chronic illness are much less well developed. This is unfortunate, because variation in the frequency of use of hospitals, intensive care units (ICUs), and physician services in managing chronic illness is most closely correlated with Medicare per capita spending among regions.¹

In this paper we show how Medicare claims can be used to evaluate the efficiency of specific providers in their management of cohorts of chronically ill Medicare patients who receive their care under traditional FFS Medicare. We measure and compare the frequency of use of health care and associated costs among cohorts according to the hospital where they receive most of their care.² We first compare the performance of seventy-seven well-known academic medical centers (AMCs). We describe the variations in the amount of care used among cohorts of patients with solid tumor cancers, congestive heart failure (CHF), and chronic obstructive pulmonary disease (COPD), comparing the illness-adjusted frequency of use for physician visits, hospitalizations, and ICU stays. We concentrate on events occurring in the last six months of life. We then show that the provider-specific frequency of use of services for one chronic disease is closely correlated with use of services for patients with other chronic conditions. Moreover, the frequency of use of care and Medicare spending during the last six months of life are highly correlated from year to year and with use by the same patients during earlier periods. We also profile the use of care and resources (physician labor inputs) among patient cohorts assigned to seven hospitals that were chosen because they ranked at the top of U.S. News and World Report’s list of the best U.S. geriatric hospitals in 2001.³

The stability of hospital-specific patterns in managing chronic illness over time indicates that provider-specific measures of resource allocation, frequency of use of hospital care and physician services, and Medicare spending based on historical research files maintained by the Centers for Medicare and Medicaid Services (CMS) should prove useful to policymakers, payers, and providers in identifying low-cost providers, estimating provider-specific actuarial costs, and providing benchmarks of efficient practice.

Study Data And Methods

The feasibility of developing provider-specific measures in FFS medicine depends on the fact that most patients with serious chronic illnesses use the same group of providers for most of their care. Claims data can thus be analyzed to assign patients to the hospital most often used (hereafter referred to as the index hospital). Details of our methods have been published elsewhere.⁴ The following reviews several aspects of the methods that are key to understanding the data.

How the study population was defined. For this study we assigned patients according to the hospital most frequently used during the last two years of life. We began by identifying all deaths among Medicare enrollees in 1999–2000 by searching the Medicare Denominator file. (The file contains a record for each enrollee in Part A or B or both, which includes date of death.) Only those with two full years of Parts A and B entitlement prior to death were included. Eligible decedents were then linked to the Medicare Provider Analysis and Review (MEDPAR) file (which contains a record for each hospitalization) to identify those with one or more medical (nonsurgical) hospitalizations in the last two years of life. Patients were then assigned to the hospital most frequently used; in the case of a tie, assignment went to the hospital with the discharge date closest to the date of death. For this study, the hospital cohorts were restricted to (1) those with one or more of eleven chronic conditions that Lisa Iezzoni and her colleagues have shown to have a high probability of death in the hospital, and (2) seventy-seven hospitals that appeared on U.S. News and World Report’s list of the best U.S. hospitals for geriatric care and for heart and pulmonary diseases in 2001.⁵

How the measures were constructed. Utilization. The numerators for utilization rates were calculated from three CMS data sources: the MEDPAR file (a 100 percent sample of hospital discharges); the Hospice File (which records all hospice stays); and a 20 percent sample of the Physician/Supplier Procedure Summary Master File (which includes claims for physician and laboratory services). Rates were calculated for six-month intervals over the two years prior to death. The measures of use include number of days spent in a hospital (hospital days), number of days spent in an intensive or coronary care unit (ICU days), number of physician visits (by specialty), percentage of patients seeing ten or more different physicians, and percentage of patients enrolled in hospice care. Measures of the intensity of terminal care include the percentage of deaths occurring during an inpatient stay and the percentage of deaths occurring during a hospitalization that involved one or more ICU admissions. The measures of resource inputs include estimates of the physician labor input and Medicare spending for Part A hospitalization and for Part B (physician and laboratory) services.

Physician labor inputs. To develop measures of physician labor input, we used data from a national study of physician productivity to obtain estimates of the average annual number of work relative value units (W-RVUs) produced per physician on a specialty-specific basis.⁶ Using Medicare Part B data, we calculated the specialty-specific sum of W-RVUs provided to a cohort over a given period. Finally, the sum of W-RVUs was divided by the average annual number of W-RVUs per physician to estimate the “standardized full-time-equivalent (FTE) physician labor input” per 1,000 cohort members.

Medicare spending. Medicare spending for each hospital was calculated by summing Medicare program reimbursements for clinical services for each cohort. Payments for Part B physician services include program reimbursements for all physician services included in the Physician/Supplier File.

Using methods of population-based epidemiology, we based all utilization rates on the total experience of the cohort, not just on services provided by the index hospital and associated providers. However, since the percentage of total hospital care provided by the index hospital is high (Exhibit 1), the variations in illness-adjusted use of care primarily reflect clinical choices made by physicians associated with that hospital.

Case-mix adjustment. Case-mix adjustment is important because the frequency of use of care differs according to the age, sex, and race of the patient; the specific chronic conditions the patient may have; and the patient’s severity of illness. To account for differences in severity of illness, we restricted the study population to patients in their last six months of life. To account for differences in the relative predominance of different chronic conditions across hospitals, we used Iezzoni’s approach to identifying chronic illnesses.⁷ Using diagnoses that appeared in the last hospital admission, we determined the presence of up to eleven chronic conditions; this was used to adjust for differences among cohorts in the underlying case-mix of disease. Finally, we adjusted for age, sex, and race. (Statistical methods for case-mix adjustment and age, sex, and race are described elsewhere.)⁸ The univariate analysis investigating the association between use rates among the seventy-seven hospital cohorts was conducted using Pearson product moment correlations.

Results

Characteristics of cohorts assigned to the seventy-seven hospitals. Exhibit 1 shows the percentage of all patient days for hospitalization for medical (nonsurgical) diagnosis-related groups (DRGs) during the last two years of life that occurred at the hospital to which the patient was assigned. The data are for deaths in 1999 and 2000. A total of 90,616 patients were assigned to the seventy-seven hospitals. Although loyalty declined as patients experienced multiple discharges, the decline was modest. For patients with two discharges, loyalty was 84 percent; for those with ten or more, it was 81 percent. The exhibit also shows data for the hospital with the lowest and highest percentage of patient days at the assigned hospital. Loyalty at the University of Pennsylvania hospital for all discharges was 81 percent; for J.D. Archbold Hospital it was 96 percent. The number of decedents assigned to each index hospital ranged from 386 to 3,506.⁹

Patterns of care. Variation in use across hospitals among patients with cancer, CHF, and COPD. We first examine patterns of care among the seventy-seven U.S. News and World Report hospital cohorts. Exhibit 2 summarizes the variation for patients with cancer, COPD, and CHF, providing the highest and lowest rates as well as the ratio of the highest to lowest (extremal ratio).¹⁰ It also shows the median rate and the rates for the region ranked twentieth (25 percent of rates are higher) and fifty-eighth (75 percent of rates are higher), as well as the ratio of these rates (the interquartile ratio) and the coefficient of variation. (The coefficient of variation, which is expressed as a percentage, is the standard deviation divided by the mean.) The mean use rates for cancer patients were lower than for COPD and CHF patients. For patient days in hospital and for physician visits, the mean rate for COPD and CHF patients was about 21–28 percent greater than for cancer patients. There is striking variation in use across hospitals within a given illness group. For patient days per decedent, the extremes varied 2.7-fold for cancer and 3.6-fold for CHF; for physician visits, they varied 4.8-fold for cancer to 6.5-fold for CHF. The greatest difference in mean use rates according to illness was for ICU days, with cancer patients receiving much less care than those with COPD or CHF. Again, there was great variation among the hospitals in how they managed care: Rates for cancer patients varied more than 14-fold; for COPD patients, 7.2-fold; and for CHF patients, 6.3-fold.

Consistency in provider-specific treatment patterns. Hospitals that have high use rates for one condition tend to have high rates for other conditions. Exhibit 3 illustrates the association between hospital days for cancer patients and for patients with CHF. The rates are highly correlated (R² =.64), even though for most hospitals CHF patients use more bed days than cancer patients (as indicated by the clustering of dots above the diagonal line).¹¹ The relationship between patient days for COPD and cancer patients followed a similar pattern. Physician visit rates were highly correlated, independent of condition (data not shown). For example, the R² statistic for physician visits for cancer and CHF patients was .69; for CHF and COPD, it was .70. ICU days for cancer patients compared to CHF and COPD patients showed the lowest correlations (R² =.46 and .44, respectively), and use rates among cancer patients were much lower in most hospitals (data not shown). By contrast, the patterns of variation for CHF and COPD patients were virtually the same, with R² =.85 and utilization profiles similar to that seen in Exhibit 4. The strong correlations indicate that with the exception of ICU use among cancer patients, the system of care—namely, the hospital to which the patient is assigned—is much more important than the nature of the patient’s chronic illness in determining the frequency of use of care.¹²

Consistency of hospital-specific use rates over time. Measuring the relative efficiency of hospitals in managing chronic illness is an important step in designing strategies to improve efficiency. Provider-specific performance measures for the last six months of life may serve this purpose because they provide an “illness-corrected” estimate for resource allocation, use, and spending that is highly correlated with patterns of practice in previous periods in the management of chronic illness. For example, as shown in Exhibit 5, the correlation between mean per person spending for inpatient care in the last six months of life and mean spending for inpatient care in the nineteenth to twenty-fourth months before death is .70 (p < .0001), even though average spending per person was five times greater during the last six months of life than during the nineteenth to twenty-fourth months before death. The correlation between spending in the last six months of life and spending during the seventh to twelfth months before death was .85; and with the thirteenth to eighteenth months, it was .79. Similarly strong correlations were seen for days spent in a hospital (Pearson R², .71, .81, and .77 for comparable periods) and for patient days in an ICU (.71, .83, and .72).

We also examined the consistency in relative rates of spending for end-of-life care among the hospital cohorts between years. Relative spending levels during the last six months of life at a given hospital appeared to be stable over at least a four-year period of observation. For example, inpatient end-of-life spending for deaths occurring in 1998 at the study hospitals was highly correlated with spending for deaths occurring in 2001 (R² =.77). Similar patterns were seen for other measures of use.

Profiling performance among the seven “best” geriatric hospitals. Exhibit 6 presents data on the use of care and medical resources among the seven hospitals that ranked at the top of the U.S. News and World Report 2001 list for geriatric care.¹³ These quality rankings are based on institutional measures (such as nurses per bed and inpatient mortality) and reputation as determined by expert opinion.¹⁴ Despite the high quality ranking of all of these hospitals, when viewed from the perspective of population-based use rates, the patient cohorts assigned to these hospitals receive very different amounts of care. During the last six months of life, patients receiving most of their care from Mount Sinai Medical Center spent almost twice as many days in the hospital as patients cared for by the Mayo Clinic hospitals (St. Mary’s and Rochester Methodist). The intensity of use of technology, measured by the number of ICU days, was three times greater for patients at the University of California, Los Angeles (UCLA) Medical Center, than for patients who received most of their hospital care at Massachusetts General hospital; Mount Sinai Medical Center and UCLA patients experienced more than twice as many physician visits as patients at Duke University Hospital.

The quality of terminal illness care was also very different. For example, patients assigned to St. Louis University Hospital were almost 70 percent more likely to spend time in intensive care during the hospitalization in which they died than were people who died as inpatients at Mayo Clinic hospitals. Hospice use by members of the Johns Hopkins cohort was 2.5 times greater than for Mount Sinai (31.1 percent versus 12.4 percent). The number of patients dying in the hospital also varied from a low of 32 percent of all deaths to more than 52 percent. The varying probability that death occurs in the hospital calls into question the validity of U.S. News and World Report’s use of inpatient mortality rates as a basis for evaluating hospital quality in managing chronic illness.¹⁵

The AMCs ranked at the top of the U.S. News and World Report list for geriatric care also differed in the amount of physician labor they invested in caring for patients. During the last six months of life, labor input for physicians primarily responsible for managing chronic illness ranged from 12 physicians per 1,000 patients at Duke University and the Mayo Clinic to 20 and 22 per 1,000 at UCLA and Mount Sinai. Labor input for primary care physicians serving UCLA and Duke University Hospital populations were 4 and 3 FTE physicians per 1,000 cohort members, while patients at St. Louis University and Mount Sinai Hospitals received two times more (8 FTEs per 1,000 cohort members). For medical specialists, labor input per 1,000 members among the Mayo Clinic and Johns Hopkins cohorts was relatively low (6 FTE physicians), while for UCLA and Mount Sinai the input rates were nearly two times greater (12 and 11 FTEs).

The patterns of deployment of physician labor also differed with regard to the mix between primary care physicians and medical specialists. The FTE input of specialists serving the UCLA population was 2.8 times greater than that for primary care physicians (Exhibit 6). By contrast, for St. Louis University and Johns Hopkins Hospitals, the labor input of specialists was slightly less than that for primary care physicians.

Discussion

Heretofore, provider-specific information on the size of the patient population served by a given provider and on the per capita inputs of resources and health services use have been available routinely only to managers of group- and staff-model health maintenance organizations (HMOs). In this study we have shown that claims data can be used to define chronic disease populations that rely on specific hospitals and thereby provide a means to measure provider-specific, population-based rates of use of resources, physician services, and hospital care for patients enrolled in traditional FFS Medicare.

The performance measures provide a very different perspective than that provided by U.S. News and World Report’s measures, which are based on reputation and on institution-specific measures such as number of nurses per bed. In addition to documenting marked variation across hospitals, our performance measures make transparent the relationship between management decisions that determine the size of the professional workforce and the numbers of hospital beds and other resources, on the one hand, and the costs and use of care, on the other.

Appropriateness of hospital-level analyses. Organizing our analysis around the hospital is appropriate because decisions made at this level have a profound effect on the amount of care provided and its per capita costs. Physicians associated with the hospital are responsible for deciding who is admitted as well as the amount and type of care patients receive. In turn, the probability of being hospitalized and admitted to an ICU is related to the capacity of the hospital compared with the size of the population it serves. Hospital management, including senior administration, medical staff leadership, and hospital boards, is responsible for major decisions that affect the capacity of local or regional health care systems, including the numbers of beds and outpatient facilities, and for hiring the nonphysician workforce, including nurses. They are also responsible, at least indirectly, for the size of the physician workforce because they are ultimately responsible for extending admitting privileges and, in some cases, for contracting with or acquiring primary care physician practices.

On the other hand, what is going on at the physician level is also important. Practice styles among physicians, even those practicing at the same hospital, can differ, and knowledge of these differences can be useful for efforts to improve quality and efficiency in a given hospital.¹⁶ Both private- and public-sector strategies for improving the efficiency of care would benefit from such information.

The differences among hospitals—particularly those within the same market— point to the potential value of claims data in identifying what Arnold Milstein has called “longitudinally efficient providers”: those who use less care and fewer resources and incur lower costs in managing patient populations over time.¹⁷ The interpretation that hospitals providing fewer resources and less frequent care to their patients are, on average, more “efficient” than those providing more depends on the findings of Elliott Fisher and colleagues that an increasing rate of care intensity is not necessarily associated with marginal benefit to the patient.¹⁸ In a previous study Fisher and colleagues found no improvement in life expectancy, functional status, or patient satisfaction among populations with chronic illnesses living in regions with high rates of care compared to those with low rates. They report a similar finding among patients treated at AMCs.¹⁹ On the basis of the failure to find evidence that more care is better (indeed, the troubling indications that more may be worse), AMCs with low-intensity patterns can be viewed as benchmarks for (relatively) efficient care.

Opportunities for care management. The availability of provider-specific information should open up new opportunities for the active management of FFS health care. Because of the stability of provider-specific rates of resource inputs, spending, and use, the historical Medicare claims data—even with the current eighteen-month time lag—should be useful for evaluating performance and estimating relative actuarial costs. Studies comparing Blue Cross Blue Shield and Medicare use rates show them to be highly correlated, which indicates that in managing chronic illness, efficient providers for Medicare patients are likely to be efficient providers for those insured by the private sector.²⁰ For these reasons (and because Medicare estimates are based on large samples), information based on Medicare data may prove useful for selective contracting and other “value health purchasing” strategies by private-sector payers.

Provider-specific information that profiles performance in managing chronic illness may be useful in implementing provisions of the Medicare Prescription Drug, Improvement, and Modernization Act (MMA) of 2003. An important example is Section 721, which calls for chronic care improvement programs modeled after private-sector disease management programs. Success will depend on how well case management, nurse coaching, and other techniques of disease management are implemented. It will also depend on the use of Medicare claims data to identify high-risk patients for “targeted outreach”; information that identifies high-cost providers should be useful in planning outreach strategies.

Ultimately, success in the “longitudinal” management of populations with chronic illness will depend on the integration of care across various sectors: acute care hospitals, primary care, nursing home care, home health care, and hospice care. It will also depend on a financing strategy that promotes the rational allocation of resources among the sectors of care. The claims data provide the basis for tracking patients though the sectors of care to benchmark efficient practices and estimate resource inputs and actuarial costs. The data should therefore also be useful in implementing MMA Section 646, the Medicare Health Care Quality Demonstration Program. This program places responsibility for integrating and coordinating care on participating provider organizations and offers them the opportunity to propose modifications in traditional FFS reimbursement to support systems integration.

Study limitations. The Medicare data themselves are limited in several ways. The estimates provided in this study were limited to acute care hospital services, physician and laboratory services, and hospice care. This limitation should be overcome by including Medicare files covering skilled nursing facility (SNF) services, home health care, outpatient facilities, and durable medical equipment (DME). However, since Medicare does not pay for long-term nursing home stays, information on use in this sector is limited to physician services. The Part B physician service research files currently available are based on a 20 percent sample, which limits the value of Medicare claims in evaluating physician performance. This limitation could be overcome if the CMS would agree to provide files covering all Part B services. The data are based on historical files, which necessitates at least an eighteen-month delay. However, MMA contains provisions for making FFS claims data available much sooner.

The case-mix measures we used in this study were limited to the diagnoses present on the hospital discharge records. Our study population, however, was restricted to patients at each hospital that, in retrospect, had identical prognoses, and we accounted for several major factors that could further influence utilization rates—for example, specific chronic conditions such as cancer, which are associated with lower utilization rates, as well as age, sex, and race. The consistency of the findings for the specific hospitals across clinical conditions suggests that unmeasured case-mix differences are unlikely to explain the dramatic differences observed.

Medicare claims can be used to provide illness-adjusted, population-based measures of resource inputs, use, and Medicare spending for cohorts of FFS Medicare patients who receive most of their care from a given hospital. We measured the performance of seventy-seven hospitals deemed among America’s “best hospitals” to document extensive variation in the amount of care provided to patients with three common conditions. We suggest that these measures could prove useful for identifying relatively efficient providers and in configuring (and evaluating) provider networks and designing (and evaluating) “pay for performance” strategies.

The authors thank Kristen K. Bronner for assistance with graphics and Martha M. Smith for editorial assistance. Grant support was provided by the Robert Wood Johnson Foundation and the National Institute on Aging (P01 AG019783).

NOTES

1. J.E. Wennberg, E.S. Fisher, and J.S. Skinner, “Geography and the Debate over Medicare Reform,” Health Affairs, 13 February 2002, content.healthaffairs.org/cgi/content/abstract/hlthaff.w2.96 (15 June 2004).
2. The findings reported here build on two previous publications that dealt with hospital-specific measures. See E.S. Fisher et al., “Hospital Readmission Rates for Cohorts of Medicare Beneficiaries in Boston and New Haven,” New England Journal of Medicine 331, no. 15 (1994): 989–995; and J.E. Wennberg et al., “Use of Hospitals, Physician Visits, and Hospice Care during Last Six Months of Life among Cohorts Loyal to Highly Respected Hospitals in the United States,” British Medical Journal 328, no. 7440 (2004): 607–611.
3. For a description of the methodology used by U.S. News and World Report to create its “America’s Best Hospitals” list, see A. Comarow, “Methodology: Behind the Rankings,” 2004, www.usnews.com/usnews/health/hosptl/methodology.htm (6 August 2004).
4. Wennberg et al., “Use of Hospitals, Physician Visits, and Hospice Care.”
5. L.I. Iezzoni et al., “Chronic Conditions and Risk of In-Hospital Death,” Health Services Research 29, no. 4 (1994): 435–460.
6. Medical Group Management Association, Physician Compensation and Production Survey: 2002 Report Based on 2001 Data (Englewood, Colo.: MGMA, 2003).
7. Iezzoni et al., “Chronic Conditions and Risk of In-Hospital Death.”
8. Wennberg et al., “ Use of Hospitals, Physician Visits, and Hospice Care.”
9. A detailed data table is available online (Supplemental Exhibit 1). For each hospital, it gives the total number of decedents; the number in the cancer, CHF, and COPD cohorts; as well as the percentage of patient days for medical conditions that occurred at the hospital. The supplemental exhibit is available at
content.healthaffairs.org/cgi/content/full/hlthaff.var.5/DC2.
10. The data table in Supplemental Exhibit 2 provides for each hospital cohort the rate and the 95 percent confidence intervals for hospital days, days in the ICU, and physician visits during the last six months of life. Ibid.
11. A patient can appear in more than one disease-specific cohort (that is, the patient has more than one chronic illness). To avoid bias, in conducting the correlations we excluded patients from the cancer cohort who also had CHF or COPD or both, and patients from the CHF cohort who also had COPD.
12. The common denominator is likely the supply of resources relative to the size of the patient population. For example, at the regional level, hospital beds per 1,000 people (of all ages) is highly correlated with Medicare hospitalization rates for cancer, COPD, CHF, and other chronic illnesses, a pattern consistent with the hypothesis that hospital capacity exercises a threshold effect on clinical decision making independent of specific chronic illness. Although similar, direct population-based measures of provider-specific capacity per 1,000 can be calculated only for staff-model HMOs such as Kaiser Permanente, the fact that occupancy rates are high for the hospitals studied here indicates that available beds are fully used; hence, the patient day rate is an indirect measure of hospital bed capacity.
13. Supplemental Exhibit 2 provides rates and 95 percent CI for end-of-life measures for each of seventy-seven hospital cohorts. See Note 9.
14. The U.S. News and World Report ranking is based on an adaptation of Avedis Donabedian’s quality model based on structure, process, and outcomes. The structural dimension includes the resources available for treating patients such as nurses per bed, presence of specific technologies, and intensive care beds, derived in large part from the American Hospital Association (AHA) annual survey of hospitals. The process dimension of “the quality agenda is the net effect of physicians’ clinical decision-making” such as “clinical choices about…admission to hospital or one of its units and length of stay.” However, since such “national measures of process are difficult to obtain,” process is evaluated on the basis of reputation as determined by interviewing physician experts under the assumption that in naming a “hospital as one of the ‘best’ he or she is, in essence endorsing the process choices made (by physicians) at that hospital.” The outcome variable used for geriatrics is all-cause inpatient mortality rates. See Comarow, “Methodology.”
15. The varying probability that death will occur in the hospital means that statistics based on inpatient data will not accurately reflect the true risk of death for cohort members; moreover, the varying probability of being hospitalized, given a chronic illness, means that case fatality rates based on institutional measures are biased, even if mortality were based on thirty-day follow-up after discharge.
16. In preliminary studies in Boston, we have noted substantial differences in the frequency of use of care and Medicare spending between patients primarily managed by primary care physicians compared with those who are primarily managed by medical specialists.
17. Arnold Milstein, Pacific Business Group on Health, testimony before the U.S. Senate Committee on Health, Education, Labor, and Pensions, 28 January 2004, labor.senate.gov/testimony/milstein.html (3 September 2004). For example, there is much variation in use and Medicare spending during the last six months of life among the populations assigned to the four U.S. News and World Report hospitals located in the Boston area. For deaths occurring in 1999–2000, the Lahey Clinic Hospital provided fewer hospital and ICU days and fewer visits than Beth Israel–Deaconess, Massachusetts General, and Brigham and Women’s Hospitals.
18. E.S. Fisher et al., “The Implications of Regional Variations in Medicare Spending, Part 1: Utilization of Services and the Quality of Care,” Annals of Internal Medicine 138, no. 4 (2003): 273–287, and “The Implications of Regional Variations in Medicare Spending, Part 2: Health Outcomes and Satisfaction with Care,” Annals of Internal Medicine 138, no. 4 (2003): 288–298.
19. E.S. Fisher et al., “Variations in the Longitudinal Efficiency of Academic Medical Centers,” Health Affairs, 7 October 2004, content.healthaffairs.org/cgi/content/abstract/hlthaff.var.19.
20. J.E. Wennberg and D.E. Wennberg, eds., The Dartmouth Atlas of Health Care in Michigan (Hanover, N.H.: Center for the Evaluative Clinical Sciences, Dartmouth Medical School, 2000).

John Wennberg (john.wennberg{at}dartmouth.edu) is director of the Center for the Evaluative Clinical Sciences and the Peggy Y. Thomson Professor for Evaluative Professor for Evaluative Clinical Sciences at Dartmouth Medical School in Hanover, New Hampshire. Elliott Fisher is codirector of the Veterans Affairs (VA) Outcomes Group, VA Medical Center, in White River Junction, Vermont. Thérèse Stukel is a senior scientist and research director at the Institute for Clinical Evaluative Sciences in Toronto, Ontario. Sandra Sharp is a research associate at the Center for the Evaluative Clinical Sciences, Dartmouth Medical School.

DOI: 10.1377/hlthaff.var.5
©2004 Project HOPE–The People-to-People Health Foundation, Inc.