Special Article
Clarifying Sources of Geographic Differences in Medicare Spending
N Engl J Med 2010; 363:54-62July 1, 2010
- Abstract
Background
Although geographic differences in Medicare spending are widely considered to be evidence of program inefficiency, policymakers need to understand how differences in beneficiaries' health and personal characteristics and specific geographic factors affect the amount of Medicare spending per beneficiary before formulating policies to reduce geographic differences in spending.
Methods
We used Medicare Current Beneficiary Surveys from 2000 through 2002 to examine differences across geographic areas (grouped into quintiles on the basis of Medicare spending per beneficiary over the same period). We estimated multivariate-regression models of individual spending that included demographic and baseline health characteristics, changes in health status, other individual determinants of demand, and area-level measures of the supply of health care resources. Each group of variables was entered into the model sequentially to assess the effect on geographic differences in spending.
Results
Unadjusted Medicare spending per beneficiary was 52% higher in geographic regions in the highest spending quintile than in regions in the lowest quintile. After adjustment for demographic and baseline health characteristics and changes in health status, the difference in spending between the highest and lowest quintiles was reduced to 33%. Health status accounted for 29% of the unadjusted geographic difference in per-beneficiary spending; additional adjustment for area-level differences in the supply of medical resources did not further reduce the observed differences between the top and bottom quintiles.
Conclusions
Policymakers attempting to control Medicare costs by reducing differences in Medicare spending across geographic areas need better information about the specific source of the differences, as well as better methods for adjusting spending levels to account for underlying differences in beneficiaries' health measures.
Media in This Article
Figure 1
Unexplained Differences in Medicare Spending per Beneficiary between the Quintile with Lowest Average Medicare Spending in 2000–2002 (Quintile 1) and Each of the Higher-Spending Quintiles, According to Six Models.Table 1
Mean Medicare Spending and Health and Other Individual-Level Values, According to Medicare Spending Quintile.Article Activity
- Article
The issue of geographic differences in Medicare spending per beneficiary has been fueled by more than two decades of research behind the Dartmouth Atlas of Health Care.1 Most notably, after adjustment for baseline differences in health status and other variables, Medicare spending per beneficiary for physician and hospital services was 52% higher in regions in the top quintile of spending than in those in the lowest quintile, with no apparent differences in quality.2,3 The Dartmouth research suggests that the primary source of the spending differences is variation in the supply of medical resources and that this variation leads to an overprovision of “supply-sensitive” services — that is, services that are provided more frequently in areas with more physicians and hospital beds than in areas with fewer physicians and beds.4-6
Given the extent of geographic differences in Medicare spending per beneficiary, policymakers need a more complete understanding of the underlying sources of the geographic differences before formulating policies to shrink them. A recent study by Sutherland and colleagues used data from the Medicare Current Beneficiary Survey (MCBS) to assess the contributions of a limited number of health and sociodemographic factors to differences in Medicare spending per beneficiary across quintiles of area-level Medicare spending intensity.7 The study showed that differences in health status accounted for 18% of the difference in per-person Medicare spending between the lowest and highest quintiles and that unidentified regional factors were responsible for more than 70% of the difference, findings that are consistent with earlier Dartmouth research.2,3
However, these findings may exaggerate the importance of regional factors if the unexplained difference was due in part to differences in unmeasured characteristics of patients. Moreover, to the extent that area-level factors account for spending differences across communities, policymakers should have more precise information identifying those factors. The current study extends previous research by investigating how much more of the geographic differences can be explained by three sets of factors not included in previous analyses, including the most recent study of regional differences7: additional measures of health (health at baseline and changes in health that occur during the year), supplementary insurance coverage, and area-level measures of the supply of health care resources.
Methods
Data Sources
We pooled data from the 2000–2002 MCBS Cost and Use files for elderly respondents (those >64 years of age) who were not institutionalized at the time of the baseline interview and who had traditional (fee-for-service) Medicare coverage for their first year in the survey.8 These 2000–2002 data are the most recent 3 years of data from an earlier study that looked at Medicare spending differences over the 11-year period from 1992 to 2002.9 We used data from these 3 years only to improve the comparability of our results with those from the recent study by Sutherland and colleagues in which MCBS data for 2004–2005 were used.
Spending data are expressed in 2002 U.S. dollars, on the basis of the medical care component of the consumer price index and are adjusted for cross-sectional differences in input prices by means of the Medicare prospective payment system hospital-wage index.10 However, the MCBS data we used could not be adjusted to remove Medicare payments to disproportionate-share hospitals and graduate medical education institutions before geographic differentials in spending were analyzed. The final sample included 6725 Medicare beneficiaries.
Annual information on the supply of medical resources by year and county was obtained primarily from Medicare's Provider of Services files and the Area Resource File.10,11 The full set of explanatory variables included measures of individual beneficiaries' baseline health and changes in health during the observation year, demographic characteristics (age, sex, race or ethnic group, and type of residence [urban or rural]), other individual determinants of demand (income and supplementary insurance coverage), and area-level measures of the supply of health care resources (numbers of hospital beds and physicians per 1000 elderly population, percentage of physicians in primary care, number of resident physicians per bed, and whether or not the nearest hospital with ≥100 beds was a teaching hospital).
Statistical Analysis
We analyzed the determinants of annual Medicare spending per beneficiary using ordinary least-squares-regression methods and Stata statistical software (version 10.1). We assessed the role of unmeasured geographic factors by using a set of dichotomous variables that indicated the ranking of each person's residence based on the weighted average of price-adjusted Medicare spending per beneficiary over the 2000–2002 period in the Dartmouth Atlas Hospital-Referral Regions (HRRs). We mapped beneficiaries' residential ZIP Codes to the HRRs and then assigned to each ZIP Code the Medicare spending per beneficiary for the HRR. These continuous values were then converted into categorical indicators representing the area's quintile ranking, with each quintile accounting for one fifth of the weighted population represented by the MCBS.
To measure the effect of various factors on the magnitude of the unexplained geographic differences in Medicare spending, we estimated a series of linear multivariate-regression models that had as the dependent variable price-adjusted Medicare spending per beneficiary and that initially included only dummy variables for the HRR quintiles. The coefficients of the quintile dummy variables measure the difference in Medicare spending per beneficiary between quintile 1 (with the lowest spending) and each of the other four quintiles. With no other variables in the model, these initial coefficients are precisely the differences in price-adjusted Medicare spending per beneficiary across the quintiles.
We then expanded the number of explanatory variables in the model in five steps. The changes in the coefficients of the quintile dummy variables as each set of additional measures were included represent the amount of the initial geographic difference that is “explained” by the additional measures. The coefficients of the quintile dummy variables from the final regression model, which includes all the independent variables, represent the amount of the geographic difference that is due to unidentified factors not represented by the model's variables.
The first two sets of variables reflect the specifications used by Sutherland and colleagues: first, demographic characteristics (age, sex, race or ethnic group, and an indicator of urban versus rural residence); and second, baseline health measures (self-reported general health status, smoking status [current or former smoker], body-mass index, and previous diagnoses of diabetes or hypertension). In subsequent sets of variables, we added other measures of baseline health and changes in health that occurred during the observation year, family income and supplementary insurance coverage, and area-level supply of medical resources.
The set of other baseline health variables also included self-reports of whether the person had previously received diagnoses of five other health conditions: myocardial infarction, coronary heart disease, another heart problem, stroke, or any nonskin cancer. We also included a variable indicating whether the interview was completed by a proxy respondent, as an indirect measure of baseline health and functional status. Changes in health were represented by indicators of whether the person died during the observation year or reported new diagnoses of diabetes, hypertension, myocardial infarction, coronary heart disease, another heart problem, stroke, or any nonskin cancer.
Results
Measuring Geographic Differences
Table 1Table 1
Mean Medicare Spending and Health and Other Individual-Level Values, According to Medicare Spending Quintile. presents the mean spending and population characteristics overall and for each quintile of Medicare spending. The mean Medicare spending per beneficiary ranged from $4,721 in quintile 1 to $7,183 in quintile 5, a difference of 52%. Many health-status measures varied significantly across the quintiles. For example, the percentage of beneficiaries in poor health was 4.5% in quintile 1, versus 7.9% in quintile 5.Figure 1Figure 1
Unexplained Differences in Medicare Spending per Beneficiary between the Quintile with Lowest Average Medicare Spending in 2000–2002 (Quintile 1) and Each of the Higher-Spending Quintiles, According to Six Models. shows how the estimated magnitudes of the unexplained geographic differences in Medicare spending per beneficiary changed as groups of variables were added to the regression model. (The Supplementary Appendix, available with the full text of this article at NEJM.org, reports the coefficients of the final regression model, which includes all the independent variables.) According to the “unadjusted” model in Figure 1, which involves adjustment only for cross-sectional and secular price differences, Medicare spending per beneficiary increased monotonically across the spending quintiles. As compared with the spending in quintile 1, spending in other quintiles was higher, with the increase ranging from $1,040 (in quintile 2) to $2,462 (in quintile 5). Although the absolute difference between spending in quintile 1 and quintile 5 was less than the $3,280 reported in the study by Sutherland and colleagues, the percent difference of 52% is similar to the 50% difference reported in that study.Explaining Geographic Differences
The other results in Figure 1 and the data in Table 2Table 2
Geographic Differences in Medicare Spending after Adjustment for Selected Variables, According to Medicare Spending Quintile. show how the geographic differences in Medicare spending changed as a result of including additional explanatory variables. Adding demographic characteristics and the initial set of baseline health characteristics reduced the magnitude of the unexplained difference between quintiles 1 and 5 from $2,462 to $1,893, or from 52% to 40%. However, 77% of the geographic difference in spending between quintiles 1 and 5 remained unexplained (Table 2), which is similar to the percentage reported by Sutherland and colleagues.Adding the expanded set of other health variables, reflecting baseline health and changes in health measures, further reduced the size of the unexplained geographic differences between quintile 1 and quintiles 3, 4, and 5 but had a minimal effect on the difference between quintiles 1 and 2, which was no longer significant. This model resulted in a reduction in the size of the unexplained geographic differences between quintile 1 and quintiles 3, 4, and 5 that ranged from 21% to 33% (Table 2), suggesting that the observed geographic differences could be explained in large part by differences in patients' demographic characteristics and health status. The addition of data on baseline health and changes in health reduced the percentage of the spending difference between quintiles 1 and 5 that was unexplained from 92% (with the model including demographic variables only) to 63%, suggesting that health measures account for 29% of the spending difference. However, the percentages of the interquintile differences that remain unexplained were still quite large, ranging from 55% to 63% (Table 2).
Finally, including in the model information on annual family income and supplementary insurance coverage, as well as the supply of health care resources, had little additional effect on the magnitudes of the unexplained geographic differences. (The type of supplementary insurance coverage differed significantly among the quintiles in the model, but annual family income and all the area-level variables reflecting the supply of health care resources did not [see the Supplementary Appendix].)
Selected Explanatory Factors
To illustrate some of the sources of the reduction in the unexplained spending differences, in Table 3Table 3
Influence of Selected Factors on Medicare Spending Differences. we report the mean values for selected variables that differed significantly between quintile 1 and quintile 5, along with the marginal effect (the regression coefficient, in dollars) on Medicare spending per beneficiary from the regression model (see the Supplementary Appendix). Multiplying the difference in the means and the coefficient (and dividing by 100) provides estimates of the amount of the difference in spending between quintiles 1 and 5 that was accounted for by each variable. Clearly, the most important variable shown was the proportion of beneficiaries who died during the year, which ranged from 2.0% in quintile 1 to 3.1% in quintile 5. Although this is not a large absolute difference, the marginal effect on Medicare spending associated with death during the year was $17,523, meaning that $198, or approximately 8%, of the $2,462 difference in per-beneficiary spending between quintiles 1 and 5 was accounted for by this variable alone. The other two variables that differed significantly in frequency between quintiles 1 and 5 were newly diagnosed coronary heart disease and a proxy respondent at baseline, accounting for $108 and $51, respectively, of the difference in spending between the quintiles. Conversely, a higher proportion of beneficiaries in quintile 1 were survivors of nonskin cancer, and the marginal effect of this fact on spending was $1,381, which increased the difference in spending between quintile 1 and quintile 5 by $38. Finally, Medicaid coverage and private supplementary insurance were both highly significant in the regression model but had opposing effects across the quintiles. The net effect of these differences regarding supplementary insurance was to raise Medicare expenditures by $89 in quintile 5 relative to quintile 1.Discussion
After adjustment for demographic characteristics, baseline health measures, and measures of changes in health, the difference in spending between the highest and lowest Medicare-spending quintiles was reduced from 52% to 33%, suggesting that the observed geographic differences could be explained in part by differences in patients' demographic characteristics and health status. We found that health status accounted for 29% of the geographic difference in per-beneficiary spending between the highest and lowest quintiles, as compared with only 18% in the recent analysis by Sutherland and colleagues. In our analyses, the percentage of the spending difference that remained unexplained was reduced but still remained substantial, at over 60%. Although the reduction in geographic differences explained by our expanded set of health measures in the models does not appear to be large in an absolute sense, the political ramifications of, for example, reducing Medicare payments by an additional 11% in high-spending areas would most likely be substantial. Thus, policymakers attempting to adjust Medicare payments on the basis of an area's geographic-spending rank need to carefully account for differences in the region's health care needs.
Our analysis also revealed that differences in the supply of medical resources are neither significant nor quantitatively important. The results do suggest that supplementary insurance coverage, whether public or private, has a substantial and significant effect on individual-level Medicare spending but does not explain much of the geographic difference in Medicare spending. Nonetheless, this finding poses a challenge for policymakers to understand and assess the roles of patients' nonhealth characteristics in causing differences in Medicare spending across individual beneficiaries.
Though significant unexplained differences in spending across areas remain, neither our study nor the Dartmouth research has captured all potential sources of these differences. Medicare spending could be influenced by differences in the organization of practices, market structure, providers' profit-seeking behavior, rates of inappropriate Medicare payment, cultural or social preferences of beneficiaries, or still-unmeasured differences in health status. Policymakers need to work cautiously, given the limits of our knowledge about the causes of geographic differences in spending.
Our results suggest that including additional and better measures of health reduces the magnitude of the unexplained geographic differences, but our additional measures could be improved on; they do not capture the severity of illness or the presence of multiple chronic conditions. In addition, some of the remaining unmeasured differences are presumably due to our inability to adjust for Medicare's additional payments to teaching and disproportionate-share hospitals. For example, a recent Medicare Payment Advisory Commission report stated that geographic differences shrink considerably when a more comprehensive health adjustment is used and when additional Medicare payments for specific policy purposes — such as adjusting for input-price differences, paying higher Medicare fees to physicians in areas with a physician shortage, and partially offsetting hospitals' expenses for graduate medical education and indigent care — are taken out of the calculations.12
Another possible limitation of this analysis is that we measured the supply of providers at the county level rather than the HRR level. However, a preliminary analysis that measured area-level variables at the level of the metropolitan statistical area (rather than the HRR) and that was limited to people living in these areas did not show a change in results. Moreover, from the perspective of the individual beneficiary seeking general medical care, the relevant market area is probably much smaller than the typical HRR, since HRRs are based on patterns of referral to teaching hospitals for heart attack and stroke. Though this may be a reasonable measure of a hospital's competitive market area, it is surely larger than the typical geographic area in which most Medicare beneficiaries receive most of their care.
Policies that focus on area-level spending without adequate adjustment for differences in beneficiaries' health status could inappropriately reward or penalize certain geographic areas. In fact, the Dartmouth researchers themselves have decided not to recommend geographically based expenditure caps, payment-rate cuts, or a renewed emphasis on regulatory approaches to reduce the supply of hospital beds and physicians.13-15 The persistence of unexplained geographic differences in Medicare spending underscores the need for continued investigation of the complex factors underlying these differences. In the absence of better information about the sources, causes, and consequences of geographic differences in Medicare spending, policymakers should resist the appeal of simple solutions for limiting expenditures in high-cost areas.
Supported by a grant from the Robert Wood Johnson Foundation Program on Health Care Financing and Organization to George Mason University.
Disclosure forms provided by the authors are available with the full text of this article at NEJM.org.
This article (10.1056/NEJMsa0909253) was published on May 12, 2010, at NEJM.org.
We thank Joel Ruhter and Matthew Cravens for providing research and computer programming assistance.
Source Information
From the Urban Institute, Washington, DC (S.Z., T.W., R.B.); and the Department of Health Administration and Policy, College of Health and Human Services, George Mason University, Fairfax, VA (J.H.).
Address reprint requests to Dr. Zuckerman at the Urban Institute, 2100 M St. NW, Washington, DC 20037, or at szuckerman@urban.org.
- References
References
1
Wennberg JE, Cooper M, eds. The Dartmouth atlas of health care 1998. Chicago: American Hospital Publishing, 1998.
2
Fisher ES ,Wennberg DE ,Stukel TA ,Gottlieb DJ ,Lucas FL ,Pinder EL . The implications of regional variations in Medicare spending. Part 1: the content, quality, and accessibility of care. Ann Intern Med 2003;138:273-287
Web of Science | Medline3
Fisher ES ,Wennberg DE ,Stukel TA ,Gottlieb DJ ,Lucas FL ,Pinder EL . The implications of regional variations in Medicare spending. Part 2: health outcomes and satisfaction with care. Ann Intern Med 2003;138:288-298
Web of Science | Medline4
Supply-sensitive care: Dartmouth Atlas project brief. (Accessed May 7, 2010, at http://www.dartmouthatlas.org/keyissues/issue.aspx?con=2937.)
5
Wennberg JE ,Fisher ES ,Skinner JS . Geography and the debate over Medicare reform. Health Aff (Millwood)2002;W96-W114
Medline6
Fisher ES ,Wennberg JE ,Stukel TA , et al. Associations among hospital capacity, utilization, and mortality of US Medicare beneficiaries, controlling for sociodemographic factors. Health Serv Res 2000;34:1351-1362
Web of Science | Medline7
Sutherland JM ,Fisher ES ,Skinner JS . Getting past denial -- the high cost of health care in the United States. N Engl J Med 2009;361:1227-1230
Full Text | Web of Science | Medline8
Adler GS. Concept and development of the Medicare Current Beneficiary Survey. In: Proceedings of the Survey Research Methods Section, American Statistical Association 1998:153-5. (Accessed May 10, 2010, at http://www.amstat.org/sections/srms/proceedings/papers/1998_021.pdf.)
9
Zuckerman S, Waidmann T, Hadley J, Berenson R. Variation in Medicare spending per beneficiary: clarifying the magnitude of geographic differentials. Washington, DC: The Urban Institute, December 2009.
10
Centers for Medicare and Medicaid Services. Provider of services file. Baltimore: CMS, 2009.
11
Health Resources and Services Administration. Area resource file. Fairfax, VA: HRSA, 2002.
12
Medicare Payment Advisory Commission. Report to the Congress: measuring regional variation in service use. Washington, DC: MedPAC, December 2009.
13
McKethan A, McClellan M. Moving from volume-driven medicine toward accountable care. Health Affairs Blog. August 20, 2009. (Accessed May 7, 2010, at http://healthaffairs.org/blog/2009/08/20/moving-from-volume-driven-medicine-toward-accountable-care.)
14
Baicker K ,Chandra A . A trillion-dollar geography lesson. Health Aff (Millwood) 2009;28:1448-1451
CrossRef | Web of Science | Medline15
Chandra A. Geography and the keys to health care reform. Health Affairs Blog. June 17, 2009. (Accessed May 7, 2010, at http://healthaffairs.org/blog/2009/06/17/geography-and-the-keys-to-health-care-reform.)
- Citing Articles (18)
Citing Articles
1
Richard A. Cooper, Matthew A. Cooper, Emily L. McGinley, Xiaolin Fan, J. Thomas Rosenthal. (2012) Poverty, Wealth, and Health Care Utilization: A Geographic Assessment. Journal of Urban Health
CrossRef2
David Hill Howard. (2012) A Better Understanding of Variation in Cancer Care. Medical Care 50:5, 363-365
CrossRef3
Salomeh Keyhani, Raphael Falk, Tara Bishop, Elizabeth Howell, Deborah Korenstein. (2012) The Relationship Between Geographic Variations and Overuse of Healthcare Services. Medical Care 50:3, 257-261
CrossRef4
Michael L. Barnett, Nicholas A. Christakis, James O’Malley, Jukka-Pekka Onnela, Nancy L. Keating, Bruce E. Landon. (2012) Physician Patient-sharing Networks and the Cost and Intensity of Care in US Hospitals. Medical Care 50:2, 152-160
CrossRef5
Holly G. Prigerson, Paul K. Maciejewski. (2012) Dartmouth Atlas: Putting End-of-Life Care on the Map but Missing Psychosocial Detail. The Journal of Supportive Oncology 10:1, 25-28
CrossRef6
Epstein, Arnold M., Jha, Ashish K., Orav, E. John, . (2011) The Relationship between Hospital Admission Rates and Rehospitalizations. New England Journal of Medicine 365:24, 2287-2295
Full Text7
2011. How Can We Improve the Quality of Care in the United States?. , 93-112.
CrossRef8
Thomas C. Ricketts, Daniel W. Belsky. (2011) Medicare Costs and Surgeon Supply in Hospital Service Areas. Annals of Surgery1
CrossRef9
J. Williams. (2011) Geographic variations in health care utilization: effects of social capital and self-interest, and implications for US Medicare policy. Socio-Economic Review
CrossRef10
Richard A. Cooper. (2011) Geographic Variation in Health Care and the Affluence-Poverty Nexus. Advances in Surgery 45:1, 63-82
CrossRef11
Theodore Marmor, Jonathan Oberlander, Joseph White. (2011) Medicare and the federal budget: Misdiagnosed problems, inadequate solutions. Journal of Policy Analysis and Management 30:4, 928-934
CrossRef12
(2011) ACCF/ASE/AHA/ASNC/HFSA/HRS/SCAI/SCCM/SCCT/SCMR 2011 Appropriate Use Criteria for Echocardiography. Journal of the American Society of Echocardiography 24:3, 229-267
CrossRef13
Pamela S. Douglas, Mario J. Garcia, David E. Haines, Wyman W. Lai, Warren J. Manning, Ayan R. Patel, Michael H. Picard, Donna M. Polk, Michael Ragosta, R. Parker Ward, Rory B. Weiner. (2011) ACCF/ASE/AHA/ASNC/HFSA/HRS/SCAI/SCCM/SCCT/SCMR 2011 Appropriate Use Criteria for Echocardiography. Journal of the American College of Cardiology 57:9, 1126-1166
CrossRef14
Jonathan Skinner. 2011. Causes and Consequences of Regional Variations in Health Care. , 45-93.
CrossRef15
Amitabh Chandra, David Cutler, Zirui Song. 2011. Who Ordered That? The Economics of Treatment Choices in Medical Care. , 397-432.
CrossRef16
(2010) Geographic Differences in Medicare Spending. New England Journal of Medicine 363:17, 1677-1679
Full Text17
Andrew A. Kramer. (2010) Taking a closer look at mechanical ventilation in the United States*. Critical Care Medicine 38:10, 2067
CrossRef18
Epstein, Arnold M., . (2010) Geographic Variation in Medicare Spending. New England Journal of Medicine 363:1, 85-86
Full Text
- Letters
