Graduate medical education training may imprint young physicians with skills and experiences, but few studies have evaluated imprinting on physician spending patterns.
To examine the relationship between spending patterns in the region of a physician’s graduate medical education training and subsequent mean Medicare spending per beneficiary.
Design, Setting, and Participants
Secondary multilevel multivariable analysis of 2011 Medicare claims data (Part A hospital and Part B physician) for a random, nationally representative sample of family medicine and internal medicine physicians completing residency between 1992 and 2010 with Medicare patient panels of 40 or more patients (2851 physicians providing care to 491 948 Medicare beneficiaries).
Locations of practice and residency training were matched with Dartmouth Atlas Hospital Referral Region (HRR) files. Training and practice HRRs were categorized into low-, average-, and high-spending groups, with approximately equal distribution of beneficiary numbers. There were 674 physicians in low-spending training and low-spending practice HRRs, 180 in average-spending training/low-spending practice, 178 in high-spending training/low-spending practice, 253 in low-spending training/average-spending practice, 417 in average-spending training/average-spending practice, 210 in high-spending training/average-spending practice, 97 in low-spending training/high-spending practice, 275 in average-spending training/high-spending practice, and 567 in high-spending training/high-spending practice.
Main Outcomes and Measures
Mean physician spending per Medicare beneficiary.
For physicians practicing in high-spending regions, those trained in high-spending regions had a mean spending per beneficiary per year $1926 higher (95% CI, $889-$2963) than those trained in low-spending regions. For practice in average-spending HRRs, mean spending was $897 higher (95% CI, $71-$1723) for physicians trained in high- vs low-spending regions. For practice in low-spending HRRs, the difference across training HRR levels was not significant ($533; 95% CI, –$46 to $1112). After controlling for patient, community, and physician characteristics, there was a 7% difference (95% CI, 2%-12%) in patient expenditures between low- and high-spending training HRRs. Across all practice HRRs, this corresponded to an estimated $522 difference (95% CI, $146-$919) between low- and high-spending training regions. For physicians 1 to 7 years in practice, there was a 29% difference ($2434; 95% CI, $1004-$4111) in spending between those trained in low- and high-spending regions; however, after 16 to 19 years, there was no significant difference.
Conclusions and Relevance
Among general internists and family physicians who completed residency training between 1992 and 2010, the spending patterns in the HRR in which their residency program was located were associated with expenditures for subsequent care they provided as practicing physicians for Medicare beneficiaries. Interventions during residency training may have the potential to help control future health care spending.
Regional variations in Medicare spending have been linked to differences in health care use that cannot be fully explained by disease burden or other patient characteristics.1,2 Similar variations have been demonstrated between academic medical centers, with some centers differing by as much as 60% in the overall intensity of medical services delivered to patients with serious chronic illnesses.3 These regional and system-level variations represent the collective practice decisions of clinicians in these different systems. Some research suggests that the nature of residency training influences the nature of physician practice,4 which raises the question of whether exposure to different practice and spending patterns during residency influences physicians’ practice patterns and cost of care after training. Does it leave an imprint detectable in clinical or claims data in practice after training?
Residency training, also known as graduate medical education (GME), is a critical step in the production of the physician workforce. All states require physicians to have some GME training in the United States to become licensed, meaning every physician must train within the GME system before practice. Therefore, understanding the role of residency programs in establishing physician practice patterns related to cost of care is particularly important. In addition, Medicare- and Medicaid-funded GME represents the largest public investment in health workforce development in the United States. Medicare provides approximately $10 billion annually to teaching hospitals to support GME programs.5 Medicaid provides an additional $3 billion annually.6 Additionally, the Department of Veterans Affairs, the Department of Defense, and the Health Resources and Services Administration all provide additional explicit support for GME training each year. Given this public investment, understanding the contribution of GME programs to producing physicians who not only practice with clinical competency but also will practice in a less costly manner is important.
This study examines the relationship between spending patterns in the region of residency training and individual physician practice spending patterns after training, as well as the distribution of residency training in different spending regions.
This study was approved by the institutional review boards of the George Washington University and the American Academy of Family Physicians. Informed consent was waived because the study posed minimal risk to subjects.
The sample consisted of all Medicare beneficiaries aged 65 years or older who obtained primary care services in 2011 from a nationally representative sample of primary care physicians. The sampling frame was all physicians listed in the American Medical Association Physician Masterfile and identified as being in direct patient care in 2010, excluding residents, retirees, and physicians with unknown practice type, as well as physicians who mainly teach or conduct research. Stratifying the data by state, we drew random samples of physicians, oversampling physicians in smaller states.
Full-year 2011 Medicare claims data for all Medicare beneficiaries of the physicians in the sample were requested from the Centers for Medicare & Medicaid Services. The analysis was further restricted to primary care physicians, those listing a specialty of internal medicine or family medicine in the claims data. For a small number of physicians listing different specialties across claims, we assigned the specialty most commonly used. Hospitalists were excluded from the analysis by identifying them as individuals with more than 90% of their total evaluation and management procedures classified as inpatient.7 Physicians with fewer than 40 patients were then excluded to try to ensure reliable estimates for each physician.
Medicare beneficiaries were then linked to unique primary care physicians. Through Medicare claims, patients were uniquely assigned to primary care physicians in the sample, using the following rules8: if there was only 1 visit to a primary care physician, the patient was assigned to that physician; if there was more than 1 primary care visit, the patient was assigned to the most frequently visited primary care physician; in the case of equal numbers of visits to different primary care physicians, the patient was randomly assigned to 1 physician (this occurred for about 12% of beneficiaries); and if there were no primary care visits, then no primary care physician was assigned. The result of this process was multilevel data, with Medicare beneficiaries clustered by primary care physicians. Of the beneficiaries assigned to physicians, we excluded those with total spending higher than $100 000. Sensitivity analyses used higher ($150 000) and lower ($50 000) thresholds.
The American Medical Association Physician Masterfile was used to identify each physician’s residency training program, location, and graduation year. Location and year information was then matched with Dartmouth Atlas Hospital Referral Region (HRR) files to identify the training HRR for primary care physicians who graduated between 1992 and 2010, the years for which Dartmouth data were available. Dartmouth HRR data consisted of Part A and Part B Medicare spending per beneficiary, adjusted for age, sex, and race from 1992 to 2002. Spending was subsequently also adjusted for price.
Given the increase in Medicare spending per beneficiary over time, training HRR spending figures were standardized around a mean of zero for each year. For ease of presentation and to allow nonlinear associations, we categorized the standardized scores into 3 groups—low, average, and high spending—weighted such that each group had approximately equal numbers of beneficiaries in the 2011 Medicare claims data.
The main outcome measure was the physician’s mean Medicare Part A (hospital) and Part B (physician) spending per beneficiary for beneficiaries in their patient panel. Because these spending measures are highly skewed, the main multivariable analysis used a logarithmic transformation to normalize their distribution.
The analyses controlled for the patients’ age, sex, race/ethnicity, and number of primary care visits (defined as visits to family physicians or general internists who were not hospitalists). Race/ethnicity was included to control for differences in spending and quality of care among different racial and ethnic groups. We also calculated the Charlson comorbidity index, which is based on 17 health conditions that were found in chart reviews to be associated with mortality and was calculated from International Classification of Diseases, Ninth Revision codes in the Medicare claims data.9 Such a measure of multimorbidity was not included in the main models because of concerns that the use of a diagnosis-based measure of health may reflect treatment patterns more than patient or population health.10 All patient characteristics data were obtained from the Medicare claims files. Centers for Medicare & Medicaid Services obtains race/ethnicity information from the Social Security Administration and applies an imputation algorithm to improve the accuracy of coding.11
Physician characteristics included sex, specialty (family medicine or internal medicine), international medical graduate status, rural or urban location, percentage of care provided in hospitals, and years of practice. We used the physician main practice zip code to determine rural or urban location; categories were based on the rural-urban commuting area codes, a classification scheme using standard Bureau of Census urbanized area and urban cluster definitions in combination with work commuting patterns. The specific measure we used classified all zip code areas into 5 groups: urban, large rural, small rural, isolated rural, or frontier.12 Years in practice was determined according to the number of years since graduation from residency and was further trichotomized (1-7, 8-15, and 16-19 years) to allow subgroup analysis for examination of any decomposition of effect related to increasing time in practice.
Characteristics of the patients and physicians in the sample were calculated as a whole and separately for physicians trained in low-, average-, and high-spending HRRs. To characterize where training took place, HRRs were ordered according to Dartmouth mean spending per beneficiary for 2010 and then divided into deciles of approximately equal numbers of Medicare beneficiaries. The frequency of all physicians who completed residency in 2010 HRRs was computed for each of these deciles.
The unadjusted mean and median per-beneficiary spending of physicians for each combination of training and practice HRR categories were estimated, as well as training HRR and years of practice categories. Multivariable regression models examined the association between the natural log of the mean spending of the beneficiaries and whether the physicians trained in a low-, average-, or high-spending HRR. The model was first estimated for all physicians in the sample with and without fixed HRR effects, and second, examined for the extent to which years in practice modified the relative influence of training vs practice environment. The full model was estimated for 3 subgroups of physicians: those in practice 1 to 7 years, 8 to 15 years, and 16 to 19 years. All models included controls for patient and community characteristics, as well as individual physician characteristics. To fully capture characteristics of the HRR in which the physician practiced in 2011, the practice HRR was added as a fixed control variable with dummy variables for each of the HRRs where physicians practiced.
Alternative specifications of the full model were examined. We estimated a model that included Charlson scores. In addition, we examined the extent to which the results differed across different measures of patient use: Part B spending only and Part A spending only. In addition, as a secondary analysis, years in practice were analyzed as a continuous variable in models with and without an interaction term between years in practice and training HRR spending.
Because the sample of physicians overrepresented those in smaller states, all results were adjusted with physician-level weights to obtain national estimates. To correct for the correlation of observations treated by the same physicians, the standard errors were calculated, allowing correlation within clusters (physicians), but observations were independent across physicians. This was accomplished with the vce (cluster clustvar) command in Stata version 13.1. All tests of significance were 2-sided; significant results were defined at P < .05.
Patient and Physician Demographics
The full physician sample consisted of 38 614 physicians classified in direct patient care in 2010, representing approximately 6% of all such physicians that year (38 614 of 687 960 physicians). After limiting to physicians in a primary care specialty, not working as hospitalists, completing residency between 1992 and 2010, and with 40 or more Medicare patients, the study sample included 2851 primary care physicians. After excluding beneficiaries with more than $100 000 in 2011 Medicare spending, these physicians provided the plurality of care to 491 948 beneficiaries. In the sample, 1029 physicians were trained in low-spending HRRs, 872 in average, and 955 in high; 1032 physicians practiced in low-spending HRRs, 880 in average, and 939 in high. Sensitivity analyses that used higher ($150 000) and lower ($50 000) exclusion thresholds for Medicare beneficiary spending demonstrated minimal differences (eTable 1 in the Supplement).
Across the 3 levels of training HRR spending, beneficiaries obtaining the plurality of their primary care from physicians trained in higher-spending HRRs were significantly more likely to be Hispanic (1.9% vs 3.2% vs 7.1% in low-, average-, and high-spending training HRRs, respectively; P < .001) or black (5.9% [low] vs 8.1% [average] vs 7.4% [high]; P < .001) and reside in more urban areas (75.3% [low] vs 81.1% [average] vs 89.0% [high]; P < .001). Physicians trained in higher-spending HRRs were more likely to be general internists (31.9% [low] vs 49.3% [average] vs 66.3% [high]; P < .001), to be international medical school graduates (11.1% [low] vs 23.3% [average] vs 43.2% [high]; P < .001), and to have more years in practice (16.8% [low] vs 13.9% [average] vs 22.8% [high, who had 16-19 years in practice]; P < .001). There was a strong association between site of training and site of practice: 51.2% of physicians trained in a low-cost region practiced in one; 69.6% of those trained in a high-cost region practiced in one (Table 1).
Higher numbers of total GME residents were trained in higher-spending HRRs than in lower-spending ones (Figure). Approximately 47% of graduating residents in 2010 were trained in the top 3 deciles for average Medicare spending per beneficiary (representing 22% of HRRs serving 30% of Medicare beneficiaries).
Unadjusted mean physician spending per Medicare beneficiary for all physicians in the sample was $8262 (95% CI, $8094-$8431); the median was $3117 (95% CI, $2990-$3243). Many physicians practiced in the same type of HRR in which they trained: 567 trained and practiced in high-spending regions, whereas 674 trained and practiced in low-spending ones. For certain combinations, the numbers of physicians in the sample were relatively low; only 97 physicians trained in low-spending regions and practiced in high-spending ones. For physicians who trained and practiced in high-spending regions, mean physician spending per beneficiary was $9482 (95% CI, $9059-$9905), a significant difference of an additional $1926 (95% CI, $889-$2963) per beneficiary per year compared with that of those trained in low-spending regions who practiced in high-spending ones ($7556; 95% CI, $7036-$8076). For physicians practicing in average-spending HRRs, there was a difference of $897 (95% CI, $71-$1723) in physician spending per beneficiary between physicians trained in high- and low-spending HRRs. For physicians practicing in low-spending regions, there was not a statistically significant difference in mean spending across training HRR levels (P = .07), but there was a significant difference ($435; 95% CI, $33-$836) in the median spending per beneficiary between physicians trained in low-spending HRRs and those trained in high-spending ones (Table 2).
Differences in unadjusted mean physician spending per Medicare beneficiary between high- and low-spending training HRR varied according to physicians’ number of years in practice (Table 2). For physicians 1 to 7 years in practice, mean spending per beneficiary for those trained in high-spending regions was $2929 higher (95% CI, $1783-$4075) than for those trained in low-spending regions. The difference in spending for physicians 8 to 15 years in practice was $2050 (95% CI, $1515-$2586) and only $956 (95% CI, $199-$1713) for physicians 16 to 19 years in practice (Table 2) (eTable 2 in the Supplement provides full unadjusted outcomes for all patient and physician variables).
Adjusted Association Between Training HRR Spending and Patient Expenditures
Given the differences in physician characteristics across training HRRs, we estimated multivariable models to determine the importance of these differences in statistically accounting for the association between training HRR spending and mean physician spending per beneficiary. Because of the semilogarithmic specification of the model, the training HRR spending coefficients represent percentage changes compared with the reference group. Controlling for beneficiary-, community-, and physician-level characteristics, patient expenditures for physicians in the highest-spending training HRR category were 10% greater (95% CI, 5% to 15%) than that for physicians in the lowest-spending training HRR (Table 3). Adding a dummy variable for practice HRR reduced the magnitude of this coefficient to 7% (95% CI, 2%-12%). Across all practice HRRs, this corresponded to an estimated $522 difference (95% CI, $146-$919) in unadjusted mean physician spending between low- and high-spending training regions.
The association with training varied significantly according to the number of years in practice. For physicians in practice between 1 and 7 years, patient expenditures for those in the highest-spending training HRR category were 29% greater (95% CI, 13%-45%) than that for physicians in the lowest-spending training HRR, compared with 8% (95% CI, 2%-15%) for physicians in practice 8 to 15 years. Across all practice HRRs, this corresponded to an estimated $2434 difference (95% CI, $1004-$4111) in unadjusted mean physician spending per beneficiary between low- and high-spending training regions for physicians 1 to 7 years in practice. For physicians in practice between 16 and 19 years, adjustments removed the statistical significance for the difference in spending between training levels. In a secondary analysis analyzing years in practice as a continuous variable, each year in practice was associated with a 1.2% decrease in the difference between physicians trained in high- vs low-spending HRRs (β = −.012; 95% CI, −.023 to –.002; P = .01) (eTable 3 in the Supplement).
The other coefficients in the models indicated that patients’ age, sex, and race were associated with their Medicare expenditures. There was a positive association between a patient’s number of visits to a primary care physician and expenditures. Patient expenditures were lower in rural areas. International medical graduates, male physicians, and physicians spending more time working in a hospital setting had higher expenditures per patient. Physicians with more years of practice had overall lower mean patient expenditures.
Adding the Charlson score reduced the HRR training coefficients from 7% to 6% (95% CI, 1%-10%) (eTable 4 in the Supplement). With only Part B beneficiary expenditures for the dependent variable, physicians trained in the high-spending HRRs had patient Part B expenditures that were 5% higher (95% CI, 1%-8%) than that of their counterparts trained in low-spending HRRs (eTable 5 in the Supplement). Using only Part A (hospital) spending for the dependent variable further limited the sample to only patients who had been hospitalized in 2011, and there was no statistically significant difference in physician spending per beneficiary between physicians trained in high- vs low-spending regions.
Physician spending patterns were associated with regional spending patterns during their residency training. Physicians trained in lower-spending regions continued to practice in a less costly manner, even when they moved to higher-spending regions, and vice versa. Overall, there was approximately a 7% difference in spending between physicians trained in the highest and lowest training HRR spending groups. However, the difference in spending between physicians trained in high- vs low-spending training regions was as high as 29% for those within 7 years of completing training. This difference appeared to decrease over time to an 8% difference 8 to 15 years after completion of training and to no statistically significant difference at 16 to 19 years. These observations suggest an imprinting of care-related spending behaviors that may take place during residency. Decay of the effect over time would be consistent with a training imprint that wanes because of practice environment. These findings are notable for the initial size and continuation of the association of training with physician spending patterns for up to 15 years after training. This potential imprinting has implications for physician training and potentially for hiring, particularly because major efforts are under way to reduce health care spending.
There has been increasing interest in GME reform in recent years to ensure that federal investments produce a physician workforce matched to the public’s needs.13,14 If there is a relationship between spending patterns at the site of GME training and subsequent patterns of care, and if lower spending is not associated with lower quality of care, there may be important implications for national cost management strategies related to public funding for GME. For example, consideration might be given to prioritizing public investments in GME to institutions with learning environments in which trainees are exposed to less costly care. This association also supports training residents in new models and systems of care.
The findings of this study also suggest a need to reconsider findings of geographic imbalances in Medicare GME funding.15 GME training disproportionately occurs in higher-spending regions (Figure) so that communities will tend to recruit graduates potentially imprinted with higher-spending practice habits.
There were a number of limitations in this study. Location of training and average Medicare spending per beneficiary in an HRR were used as a proxy for training experience. Variations exist within HRRs,16 and training programs may range from high to low spending within any single HRR. This variation does not negate the findings of this study, but further research is needed to investigate the health care spending environments of residency training programs, the experiences of residents, and how these experiences influence future outcomes. Physicians caring for Medicare patients may not be representative of all physicians, and our sample of 2851 physicians specifically examined a more recent cohort of graduates, excluding those who completed residency 20 or more years ago.
An additional limitation was that the study did not measure clinical outcomes. Although research on overall regional variations in Medicare spending indicates no significant association between higher spending and better quality of care,1 the same should be tested for residency training programs to ensure that less costly care does not become a trade-off for higher-quality care.
Areas for future research include the association between spending patterns of individual training sites and those of their graduates, the relationship of residency training in high- vs low-spending systems and quality of care, and differences in the relationship between training and practice between different physician training models or specialties.
Among primary care physicians who completed residency training between 1992 and 2010, the spending patterns in the HRR in which their residency program was located were associated with expenditures for subsequent care they provided as practicing physicians for Medicare beneficiaries. This raises the possibility that interventions during residency training may be able to contribute to the control of future health care spending.
Corresponding Author: Candice Chen, MD, MPH, Department of Health Policy, Milken Institute School of Public Health, George Washington University, 2175 K St NW, Ste 500, Washington, DC 20037 (email@example.com).
Author Contributions: Dr Petterson had full access to all of the data in the study and takes responsibility for the integrity of the data and the accuracy of the data analysis.
Study concept and design: All authors.
Acquisition, analysis, or interpretation of data: Chen, Petterson, Phillips.
Drafting of the manuscript: Chen, Petterson, Bazemore.
Critical revision of the manuscript for important intellectual content: Petterson, Phillips, Mullan.
Statistical analysis: Petterson, Bazemore.
Obtained funding: Phillips.
Administrative, technical, or material support: Chen, Phillips, Mullan.
Study supervision: Phillips, Mullan.
Conflict of Interest Disclosures: All authors have completed and submitted the ICMJE Form for Disclosure of Potential Conflicts of Interest and none were reported.
Funding/Support: This project was supported by the American Board of Family Medicine Foundation. Dr Chen was supported through a National Institute of Minority Health and Health Disparities Research and Education Advancing Mission Career Transition Award (grant 5K22MD006135-02).
Role of the Funders/Sponsors: The funding organizations had no role in the design and conduct of the study; collection, management, analysis, and interpretation of the data; preparation, review, or approval of the manuscript; or decision to submit the manuscript for publication.
Disclaimer: The findings and conclusions in this report are those of the authors and do not necessarily represent the position of the American Academy of Family Physicians or the American Board of Family Medicine.
EL. The implications of regional variations in Medicare spending, part 1: the content, quality, and accessibility of care. Ann Intern Med
. 2003;138(4):273-287.PubMedGoogle ScholarCrossref
DJ. Variations in the longitudinal efficiency of academic medical centers. Health Aff (Millwood)
. 2004;(suppl var):var19-var32. doi:10.1377/hlthaff.var.19.PubMedGoogle Scholar
AJ. Evaluating obstetrical residency programs using patient outcomes. JAMA
. 2009;302(12):1277-1283.PubMedGoogle ScholarCrossref
JS. Growth in the care of older patients by hospitalists in the United States. N Engl J Med
. 2009;360(11):1102-1112.PubMedGoogle ScholarCrossref
D. Primary care physicians’ links to other physicians through Medicare patients: the scope of care coordination. Ann Intern Med
. 2009;150(4):236-242.PubMedGoogle ScholarCrossref
CR. A new method of classifying prognostic comorbidity in longitudinal studies: development and validation. J Chronic Dis
. 1987;40(5):373-383.PubMedGoogle ScholarCrossref
JE. A population health approach to reducing observational intensity bias in health risk adjustment: cross sectional analysis of insurance claims. BMJ
. 2014;348:g2392.PubMedGoogle ScholarCrossref
A. More accurate racial and ethnic codes for Medicare administrative data. Health Care Financ Rev
. 2013;29:27-42.Google Scholar
E. The geography of graduate medical education: imbalances signal need for new distribution policies. Health Aff (Millwood)
. 2013;32(11):1914-1921.PubMedGoogle ScholarCrossref