Medical school admissions committees often consider geographic factors in their selection of new medical students, sometimes in an effort to meet local workforce needs by giving extra consideration to applicants with local origins. While studies have consistently shown that the geographic location of a physician's residency program is more predictive of work location than the geographic location of his or her medical school,1–7 admissions committees do not have the luxury of knowing the location of a future residency program for an applicant. As a result, committees must use the variables available to them,2 even if they are less powerful predictors of eventual place of practice. Although most stakeholders anticipate that those with local origins will be more likely to work locally after graduation than others,2,6,8–13 little is known about the long-term relationship between geographic origins and place of practice in midcareer.
Because many physicians move from their original practice locations within a few years,14–16 midcareer practice locations may more accurately represent the community in which physicians provide their longest professional service. Accordingly, this study examined the relationship between a physician's places of residence prior to medical school and the physician's practice location at midcareer, 17 to 19 years after graduation from medical school. This time frame places the physician a minimum of 10 years beyond residency training, even for someone who did a 7-year residency, and likely represents a long-term practice arrangement. This is also approximately the midpoint in a career for a licensed physician who graduates from medical school at 26 years of age and retires 39 years later at age 65.
A retrospective cohort of all 433 medical students who graduated from a single state institution (University at Buffalo) in a three-year period from 1989 to 1991 made up the subjects of the study. Application materials on file at the medical school, from the American Medical College Application Service (AMCAS), provided prior places of residence for these students for four points in time: (1) at birth, (2) at high school graduation, (3) at college graduation, and (4) on application to medical school. Demographic variables, including age and self-reported race/ethnicity, were also taken from AMCAS materials on file. The Office of Advancement at the study institution tracks and updates practice locations annually. They provided midcareer practice locations for the selected cohort as of 2008. Midcareer practice location was unavailable for 31 graduates, including one who died before 2008, and complete residence data were unavailable for three students, including two transfer students. This left a study cohort of 399. The archived records at the study institution provided the location and specialty of each student's residency program. Where the institution for internship was different from the institution for specialty residency training, the residency location was used. No residency data were available for two students in the study cohort.
Five residential locations were examined: at birth, at high school graduation, at college graduation, on medical school application, and at residency completion. For descriptive purposes, each location variable was coded as follows: 1 = local region, 0 = nonlocal. The first category, “local region,” refers to the eight counties in the western part of the state that are in immediate proximity to the medical school, an area where there are no other medical schools and which is considered the primary catchment/service area for the medical school. These eight counties have a population of 1.5 million, representing 7.8% of the state total.17
Descriptive statistics were computed. Bivariate and multivariate analyses were performed. In each bivariate analysis, the outcome variable was midcareer practice in the local region. For each of the four indicators of geographic origins and for residency location, bivariate logistic regression was used to estimate odds ratios (ORs), that is, the odds of practicing in the local region for students born in the local region divided by the odds of practicing in the local region for students born elsewhere. Note that ORs are related to but are not the same as relative risks. Multiple logistic regression analysis was also performed, predicting midcareer practice location from all four geographic origins indicators. No additional variables were included in the multiple logistic regression analysis. Two logistic regression goodness-of-fit indices are reported. The c statistic is a measure of the model's ability to correctly distinguish between successes (practice in the local region) and failures and is equal to the area under the receiver operating characteristic curve. It is interpreted as the percentage of all possible pairs of observations with differing outcomes in which the success has a higher predicted probability of success than the failure, and it has a range of 0.5 to 1.0, where 0.5 indicates that the model is no better than chance, and 1.0 indicates perfect discrimination.18 The Hosmer–Lemeshow statistic is also reported; a nonsignificant result indicates that the model adequately fits the data.18 A bivariate logistic regression model with the same outcome but only residence on medical school application as a predictor was also fit and compared with the four-predictor model using a likelihood ratio (LR) χ2 test. All analyses were performed using Stata 10 software.19
The University at Buffalo social and behavioral sciences institutional review board determined that this study was exempt from human subjects review.
The study population had the following characteristics: median age at matriculation to medical school was 22 years (range 18–54, IQR 21–24), 63.7% (254/399) were male, and 93.5% (371/397) had U.S. citizenship. The self-reported racial/ethnic composition was 71.5% (278/389) white, 10.5% (41/389) black, 6.9% (27/389) Latino/Hispanic, 10.3% (40/389) Asian/Pacific Islander, and 0.8% (3/389) American Indian. The most common residency specialties reported were medicine (26.5%, 105/397), pediatrics (11.8%, 47/397), general surgery (11.3%, 45/397), family medicine (9.1%, 36/397), psychiatry (7.3%, 29/397), and obstetrics–gynecology (6.3%, 25/397).
Approximately one-third (35.3%, 141/399) of the study population was born in the local region. A little less than half (40.6%, 162/399) graduated from high school in the local region. One-fourth (25.8%, 103/399) graduated from colleges in the local region. Nearly half (44.6%, 178/399) claimed residence in the local region at the time of medical school application. All but two individuals began a residency on graduation from medical school, with 29.7% (118/397) staying in the local region for residency training. By midcareer, 17 to 19 years after medical school graduation, just under one-quarter (23.1%, 92/399) were in practice in the local region. Of those practicing in the local region at midcareer, 84% (77/92) had at least one positive indicator of geographical origins in the region, and nearly three-quarters (73.9%, 68/92) had three or more.
Table 1 shows that for each indicator of geographic origins, comparable percentages of those who were local before medical school had stayed or returned to the local region in midcareer, ranging from 40.5% (72/178) among those with local residences on medical school application to 49.5% (51/103) among those graduating from college locally. This compares to a range of 9.1% (20/221) among those with nonlocal residences on medical school application to 13.9% (41/296) among those with nonlocal residences at college graduation. The odds of midcareer practice in the local region for those with local origins relative to those with nonlocal origins were also fairly uniform across the indicators of geographic origins, ranging from 6.1 for residence at college graduation (i.e., the odds of a physician who graduated from a local college being in local practice at midcareer were 6.1 times larger than the odds for a physician who did not graduate from a local college) to 7.3 for residence at high school graduation. Because it is used so widely in medical school admissions, it is notable that the OR for residence at the time of medical school application was similar in magnitude to those for the other indicators of geographic origins. Finally, of those graduates who did residency training in the local region, 58.5% were still working in the local region, comparable to the 56.3% of those who had lived locally for all four periods prior to medical school. Of the graduates who left the local region to do a residency, only 8.2% were back working in the local region, comparable to the 7.5% who had not lived locally during any of the four periods.
The described multivariate model fit the data adequately. The model had a c statistic of 0.78, where any value >0.70 is considered acceptable,20 and a nonsignificant Hosmer–Lemeshow statistic (χ42 = 0.92, P = .922). In the model, residence at birth (OR = 2.6, CI 1.2–5.8) and at college graduation (OR = 2.8, CI 1.5–5.0) were significant predictors of midcareer practice location, but residence at high school graduation (adjusted OR = 1.4, CI 0.5–4.3) and on medical school application (OR = 1.7, CI 0.6–4.7) were not. Predicted probabilities of practicing within the local region at midcareer can be derived from the logistic regression model. These range from 7.8% for physicians with no local origins to 58.3% for those whose geographic origins are all local. (These are very close to the actual probabilities of 7.5% and 56.3%, respectively.)
We also compared the four-predictor model described above with a model using only residence on medical school application as a predictor: LR χ42 = 80.11 for the four-predictor model and LR χ12 = 56.44 for the one-predictor model, with a difference of LR χ32 = 23.67 (P < .001), indicating that the larger model fits significantly better.
Our study of a cohort of graduates from a northeastern U.S. medical school shows that applicants' places of residence prior to medical school admission can be used to predict the probability that a medical school graduate will be practicing in the local region at midcareer. In particular, we showed that two places of residence—at birth and at college graduation—were significant predictors of midcareer practice location, and we showed that using all four places of residence captured in the AMCAS application provided better predictions than residence at the time of medical school application alone.
This information could help committees who wish to consider the workforce needs of the local region in the admissions process. In the simplest scenario, the committee can consider individually any single indicator of geographic origins—residence at birth, high school graduation, college graduation, or medical school application—as predictive of eventual practice in the local region. In a more refined scenario, a medical school can follow our example, taking historical applicant data for their institution, linking those data to workplace data collected by the medical school alumni office, and fitting a school-specific logistic regression model. They can then apply the model to new applicants, generating predicted probabilities of midcareer practice in the local area. These probabilities can, in turn, be rescaled into a score with an appropriate range (e.g., 1–5). In the most sophisticated scenario, a school-specific logistic regression equation can be modified annually as additional updated data are obtained and entered.
There are limitations to this study. Data are from one medical school only, in a metropolitan region that has been losing population in recent decades while retaining a strong cultural heritage and close family ties among residents. Thus, the attractiveness of the local region may be lower than average to outsiders and higher than average to those with local origins. The study cohort entered medical school roughly 20 years ago; conditions and preferences may have changed over the past two decades. In addition, the data used here are on graduates, not all students admitted, or all applicants. A modification in selection criteria would produce a different student population that might act differently from the population that was selected under the original criteria. Finally, the data do not take into account other elements of the selection process, including MCATs, grade point averages, or interview scores.
Some medical schools will consider local workforce needs an important priority. Others will not. Nevertheless, the current availability of admission data in electronic format makes it possible for admissions committees to integrate applicants' local geographic origins into their selection procedures in a more refined manner than in the past, producing decisions that align more closely with institutional priorities.
The authors would like to acknowledge the kind assistance of the Office of Medical Admissions and the Office of Advancement at the School of Medicine and Biomedical Sciences at the University at Buffalo. This research was presented at the AAMC Northeast Group on Education Affairs in Hershey, Pennsylvania on May 1, 2009 where it received the Innovation Award in Medical Education. This study was not directly funded, and the authors have no known conflicts of interest.
The University at Buffalo social and behavioral sciences institutional review board determined that this study was exempt from human subjects review.
1Cooksey JA, Harman CP. Dissecting the medical training-to-practice continuum: Factors associated with choosing in-state graduate medical education. Acad Med. 1999;74(10 suppl):S114–S117.
2Costa AJ, Schrop SL, McCord G, Gillanders WR. To stay or not to stay: Factors influencing family practice residents' choice of initial practice location. Fam Med. 1996;28:214–219.
3Dorner FH, Burr RM, Tucker SL. The geographic relationships between physicians' residency sites and the locations of their first practices. Acad Med. 1991;66:540–544.
4Jones GN, Rees AC. Lafayette's family practice residency program: Practice patterns of graduates. J La State Med Soc. 1996;148:359–363.
5Owen JA, Hayden GF, Bowman RC. Influence of places of birth, medical education, and residency training on the eventual practice locations of family physicians: Recent experience in Virginia. South Med J. 2005;98:674–675.
6Tryck S, Valenzeno DP. Future physician supply in Alaska. Alaska Med. 2006;48:90–99.
7Gilbert GE, Blue AV, Basco WT Jr. The effect of undergraduate GPA selectivity adjustment on pre-interview ranking of rural medical school applicants. J Rural Health. 2003;19:101–104.
8Szafran O, Crutcher RA, Chaytors RG. Location of family medicine graduates' practices. What factors influence Albertans' choices? Can Fam Physician. 2001;47:2279–2285.
9Hughes S, Zweifler J, Schafer S, Smith MA, Athwal S, Blossom HJ. High school census tract information predicts practice in rural and minority communities. J Rural Health. 2005;21:228–232.
10Owen JA, Conaway MR, Bailey BA, Hayden GF. Predicting rural practice using different definitions to classify medical school applicants as having a rural upbringing. J Rural Health. 2007;23:133–140.
11Schwartz MR. The physician pipeline to rural and underserved areas in Pennsylvania. J Rural Health. 2008;24:384–389.
12Urbina C, Solan B, Counsellor A, Mines J, Serna L, Kalishman S. “Where have all the students gone?” Retaining medical school graduates through educational innovations. Educ Health. 2003;16:279–285.
13Wade ME, Brokaw JJ, Zollinger TW, et al. Influence of hometown on family physicians' choice to practice in rural settings. Fam Med. 2007;39:248–254.
14West PA, Norris TE, Gore EJ, Baldwin LM, Hart LG. The geographic and temporal patterns of residency-trained family physicians: University of Washington Family Practice Residency Network. J Am Board Fam Pract. 1996;9:100–108.
15Godwin M, Hodgetts G, Wilson R, Pong R, Najgebauer E. Practice choices of graduating family medicine residents. Can Fam Physician. 1998;44:532–536.
16Frisch L, Kellerman R, Ast T. A cohort study of family practice residency graduates in a predominantly rural state: Initial practice site selection and trajectories of practice movement. J Rural Health. 2003;19:47–54.
18Peng CJ, Lee KL, Ingersoll GM. An introduction to logistic regression analysis and reporting. J Educ Res. 2002;96:3–14.
19Stata Statistical Software: Release 10 [computer program]. College Station, Tex: StataCorp LP; 2007.
20Chumbler NR, Williams LS, Wells CK, et al. Derivation and validation of a clinical system for predicting pneumonia in acute stroke. Neuroepidemiology. 2010;34:193–199.