V. Physician Model Components

Published 2022

In this module:

The Health Workforce Simulation Model (HWSM) projects supply and demand for 36 physician specialty categories following the American Medical Association (AMA) specialty taxonomy (Exhibit V-1).1 This includes:

  • four primary care specialties
  • eleven internal medicine and pediatric subspecialties
  • eleven surgical specialties
  • ten other specialties

HWSM:

  • models hospitalists trained in a primary care specialty separately from primary care physicians
  • models specialist physicians practicing as hospitalists (e.g., neurohospitalists) with their specialty (e.g., neurologists)
  • combines critical care and pulmonology, since a large portion of physicians in this combined category list both critical care and pulmonology as their specialties in the AMA Physician Masterfile
  • includes physicians who list critical care surgery as their specialty under general surgery
  • includes physicians who list critical care anesthesiology as their specialty under anesthesiology
  • includes physicians who list neonatal critical care as their specialty under neonatology

Other modules provide an overview of HWSM data, methods and assumptions for supply modeling and demand modeling. In this module we discuss key inputs to modeling supply and demand for physicians. We also describe modeling of general surgery—which is modeled slightly differently from the other specialties.

Modeling Supply

Supply modeling consists of estimating the number and characteristics of current supply, modeling the number and characteristics of new entrants to supply, modeling attrition, and modeling workforce participation decisions like weekly hours worked and geographic mobility.

Estimating the Current Active Workforce Supply

The AMA Masterfile is the primary data source for estimating the number of physicians by specialty and primary geographic location (which is mapped to state and county). The Masterfile contains data on physician age and sex. Race/ethnicity data are missing for a large percentage of physicians and thus not included in the model. The analysis is limited to physicians who have completed their residency and whose status is listed as ‘active’. Active is defined by AMA as working 20 or more hours per week in professional activities.

The Masterfile could be misclassifying older, retired physicians as active. This can be due to time lags between when physicians retire and when their status is updated in the AMA file. AMA uses a variety of sources to update physician information in their file, including outreach to physicians approximately every third year. To address this issue of retired physicians still listed as active, we omitted physicians aged 75 and older from the estimate of starting supply. We also apply the retirement probabilities to the starting supply which removes additional older physicians. Without making these adjustments, projections of future supply will drop during the first few years of the projections as the status of many older physicians' changes to ‘retired’ in the simulation model. This is because the age distribution of physicians currently in the Masterfile is inconsistent with an age distribution indicated by retirement patterns.

Identifying physician specialty in the Masterfile considered both the first and second recorded specialty. For many physicians their specialty designation is straightforward—for example, listing Family Medicine in the first specialty designation and blank in the second specialty designation. Many physicians list two specialties with one a more specialized version of the other. For example, if both Internal Medicine and Cardiology are listed, the physician is categorized as a cardiologist because cardiology is a subspecialty under internal medicine. (Some physician records might have Internal Medicine as the first specialty designation and Cardiology as the second specialty, while other records might list Cardiology as the first designation and Internal Medicine as the second designation). In general, we categorize physicians under their most specialized listed specialty. Hence, if both General Surgery and Orthopedic Surgery are listed the physician is categorized as an orthopedic surgeon. Using this logic, we applied the following rules:

  1. If only one specialty is listed, use that specialty.
  2. If a primary care specialty and a non-primary care specialty are listed, use the non-primary care specialty.
  3. If a non-surgical specialty and a surgical specialty are listed, use the surgical specialty.
  4. If general surgery and another surgical specialty are listed, use the more specialized surgical specialty.
  5. If rules (1) through (4) do not apply, use the first listed specialty.

Examples of rule (5) would be someone who lists Family Medicine as the first specialty and Pediatric Medicine as the second specialty, or who lists Orthopedic Surgery as the first specialty and Plastic Surgery as the second specialty. Physicians with two different surgical specializations are rare in the AMA data, and when then they do exist the modeling team individually reviews the combinations to determine how to categorize the physician’s specialty.

The Association of American Medical Colleges (AAMC) estimates that approximately 33,300 physicians trained in primary care practice as hospitalists.2 AAMC identified practicing hospitalists by using national provider identification to combine data from 2018 Medicare fee-for-service billing records with the 2018 Masterfile. AAMC categorized as hospitalists any physicians where 90% or more of their Evaluation and Management billing is hospital-based.

Estimates by county and state are based on primary practice location address in the Masterfile. For the 5% of records with missing practice location, we use the physician’s mailing address. We categorize each physician’s rurality by the rurality of the county of his/her office location listed in the Masterfile. The National Center for Health Statistics Urban-Rural Classification Scheme for Counties assigns one of six designations to each county.3 County level supply estimates are aggregated to the state level or to the metropolitan/non-metropolitan level in each state for reporting purposes.4

Modeling New Entrants

The mechanism for modeling new physician entrants to the workforce each year is the creation of a “synthetic” cohort of physicians based on the number and characteristics of recent graduates. Each new physician is assigned a specialty, age, sex, and geographic location that reflects the distribution seen in recent years.

AMA publishes an annual report on the number of physicians completing graduate medical education (GME), with the most recent publication covering the 2020-2021 school year.5 The numbers in this report are based on graduates from programs accredited by the Accreditation Council for Graduate Medical Education (ACGME). A total of 43,590 physicians completed a residency or fellowship in 2020-2021, though many of these graduates were for additional specialization within a broader specialty category. For example, some graduates completed a program in clinical cardiac electrophysiology, which is a specialization within cardiology, an internal medicine subspecialty. Hence, to estimate the number of physicians newly entering the workforce one needs to remove double counting of physicians who will complete multiple programs. We estimate that 30,931 physicians completed their GME in 2020-2021 and newly entered one of the 36 specialty categories modeled in HWSM (Exhibit V-2).

Furthermore, AAMC estimates approximately 1,149 primary care trained physicians became hospitalists in the prior year.2 Of these, we assumed that approximately 81% are trained in general internal medicine, 16% are trained in family medicine, 2% trained in pediatric medicine, and 1% are trained in geriatric medicine.

A comparison of specialties listed in the AMA Masterfile for physicians with two listed specialties suggests that some physicians changed careers. An example is a physician trained in family medicine but also listing sleep medicine as their specialty. Not all specialties in the AMA Masterfile are part of a formal GME program, and some physicians can change specialties without requirements to complete a GME program in that new specialty. Generally, many of the physicians changing specialties moved away from primary care and into emergency medicine or the other medical specialties such as sleep medicine, preventive medicine, occupational health, aerospace medicine, pain medicine, addiction medicine, and hospice & palliative medicine, categorized in HWSM under “Other Medical Specialties.”

Estimates of the proportion of new female physicians comes from the annual AMA publication on residents in ACGME-accredited programs during the 2020-2021 academic year (Exhibit V-3).8 The estimated age distribution comes from analysis of the 2020 AMA Physician Masterfile using data on physicians who completed their training since 2000.

The state distribution for newly trained physicians is based on our analysis of physicians who have completed graduate medical education since 2000.

Modeling Workforce Attrition

HWSM simulates physician attrition probabilities based on participant responses in AAMC’s 2019 National Sample Survey of Physicians (NSSP) to questions about expected age of retirement. The survey also collects information on hours worked patterns. To conduct this survey, AAMC contracted with Toluna, an external firm that recruited active physicians from proprietary panels. The survey began February 25, 2019, and concluded on March 25, 2019, after reaching the desired quota of 6,000 participants. AAMC developed sample weights representative of practicing physicians by broad specialty category, age group, sex, and International Medical Graduate status consistent with the 2018 AMA Master File. The sample was too small to calculate retirement probabilities by detailed specialty but is sufficient to calculate probabilities by age and sex for primary care physicians (n=2,400), surgeons (n=1,116), physicians in medical subspecialties (n=1,190), and physicians in all other specialties (n=1,294). Many physicians report planning to retire at ages 60, 65, and 70 (Exhibit V-4). Few physicians remain active past age 75, and many of these older physicians have reduced work hours per week.

NSSP findings indicated physicians intended to retire earlier than modeled in previous HRSA reports. Prior estimates of physician retirement patterns were based on survey data on physician retirement expectations from approximately 2012-2015, by individual physician specialty, obtained from several states that collected such information (Florida, Maryland, New York, and South Carolina).

Data were insufficient to model whether COVID-19 has affected physician retirement probabilities. Even prior to COVID-19 there were high reported rates of burnout among physicians and other health care providers, with published reports that COVID-19 has increased rates of burnout.9 10 11 12 13 14 A 2018 OB-GYN workforce study by Doximity found that OB-GYNs had the highest burn-out rates of all medical specialties except for emergency physicians.9 They also tend to retire earlier than most specialties with a median age of 64. Hence, the use of category retirement rates could over- or under-estimate retirement probabilities for individual specialties within each category.

Modeling Hours Worked

Changing demographics of the physician workforce has implications for future FTE supply, as hours worked patterns differ systematically by physician age, sex, and specialty category. AAMC’s 2019 NSSP asked physicians how many hours they worked during their last typical week of work (excluding any week with leave). The survey also asked physicians the percent of time spend on activities (patient care, teaching, research, administration, other) and the number of weeks worked per year. We found little difference in weeks worked per year by physician age and sex. For consistency with how hours worked patterns are modeled for other health occupations in HWSM, we used total hours in professional activities and not solely hours in patient care activities. The hours estimate excludes on-call time where the physician is not working. Most physicians work more than 40 hours per week—which is the definition used for one FTE for both supply and demand modeling. Consequently, FTE supply is higher than the count of active physicians.

Hours worked patterns were modeled using Ordinary Least Squares regression with self-reported total weekly hours worked as the dependent variable. Explanatory variables consisted of age group, sex, and age group by sex interaction. The NSSP had insufficient sample size to model hours worked patterns for individual specialties, so we used four broad specialty categories like with the retirement pattern estimations (Exhibit V-5).

Modeling Cross-State Migration

HWSM accounts for annual movement of physicians across states. The main data source for physician migration is the National Plan and Provider Enumeration System (NPPES) for 2019 and 2020. First, we use logistic regression to estimate the probability that a physician under age 50 will move to another state during the year. Explanatory variables are physician age group, sex, the state’s population, and a year indicator. We include Occupational Employment and Wage Statistics (OEWS) data on mean physician salary, by state, as an explanatory variable. The physician salary data are adjusted for state cost of living using data from the Bureau of Economic Analysis. The NPPES data does not contain the physician’s age, so we merged it with Medicare data that had NPI and graduation year. We also assume that physicians are age 28 at graduation from medical school. We compare each person’s move probability to a uniform random number between 0 and 1 to simulate whether a physician moves to a new state each year over the projection horizon.

The likelihood that each physician moving will relocate to a specific state is based on the proportion of physicians moving to that state as observed in NPPES. For example, if NPPES shows 5.8% of general internists who relocated between 2019 and 2020 moved to Florida, then each general internist who moves has a 5.8% probability of moving to Florida in the HWSM.

When a physician moves to a specific state, HWSM then tags that physician by the level of rurality of the area in which s/he practices. The rurality designation is based on the current rurality distribution of the workforce in the state. For example, if 50% of pediatricians in a state work in a large core metro location, then each pediatrician moving to that state has a 50% probability of being assigned as working in a large core metro area.

Modeling Demand

The approach to model demand for physicians follows the common model outlined in the other modules. The main demand components are (1) a constructed population file for each county in the U.S. with a representative sample of the population from 2020 through 2035, (2) prediction equations of the demand (i.e., annual expected use) for health care services, and (3) physician staffing ratios. This section focuses on construction of the staffing ratios constructed for each physician specialty for the care setting where they provide services. The staffing ratios are calculated as national expected health care use in 2020 in the absence of COVID-19 divided by national FTE physicians.

Starting year demand is assumed equal to national supply for all specialties except family medicine, general internal medicine, and psychiatry. In 2020, approximately 14,945 primary care physicians and 6,464 psychiatrists would have been required to provide a level of care to remove the Health Professional Shortage Area (HPSA) designations.15 The designation is used as a conservative proxy for the current national shortage of physicians. Research published by others suggests a national shortfall of pediatric subspecialties, though there does not appear to be a national shortfall of pediatricians.16

As described in other modules, using the Medical Expenditure Panel Survey (MEPS) we developed predication equations of people’s annual visits to physicians in office and outpatient settings by physician specialty. The services metrics used to calculate the staffing ratios include visits to each specialty. Using MEPS, we developed prediction equations of hospitalizations for approximately two dozen categories of care defined by primary diagnosis codes and procedure codes (Exhibit V-6). Using the 2019 National Inpatient Survey (NIS), we developed prediction equations for expected length of stay for these same diagnosis categories. By applying the prediction equations for hospitalizations to the population database, and then applying the prediction equations for expected length of stay given a hospitalization, we calculate total inpatient days as a set of services metrics for calculating hospital-based staffing ratios. We developed prediction equations for the probability of a consult during an emergency visit using the National Hospital Ambulatory Medical Care Survey (NHAMCS). Thus, if a person with an emergency visit for a neurological condition results in the patient seeing two physicians, we assume one physician was an emergency physician and the second physician was a neurologist providing a consult.

Exhibit V-7 summarizes estimates of what national demand for services would have been in 2020 in the absence of COVID-19. These estimates reflect patterns of care use from 2015-2019 applied to the U.S. population in 2020. For example, we would have expected approximately 202,646,000 visits to family physicians in the absence of COVID-19.

A calculation of staffing ratios requires estimates of FTE physicians by care setting. In some cases, all physicians in a specialty work in a single setting—e.g., 100% of emergency physician care is provided in emergency departments; 100% of critical care physician and hospitalist services are provided in hospital inpatient settings. In many specialties, physician FTEs are distributed across settings. Thus, these physicians will have multiple drivers of demand for services (e.g., ambulatory visits, inpatient care, and possibly care in other settings such as nursing homes). Exhibit V-8 summarizes the distribution of FTE physicians across delivery settings. For some specialties, the OEWS data provides information on how jobs are distributed across settings. We use the distribution of jobs as a proxy for the distribution of FTEs. For other specialties, we use older data from the 2015 Medical Group Association Survey.

To calculate national staffing patterns (Exhibit V-9), we divide the workload measures in Exhibit V-7 by estimates of FTE physicians by setting (Exhibit V-8). To calculate demand for physicians, we divide projections of health care services demand by the numbers in Exhibit V-9.

For procedures categorized by rurality (excluding procedures where location is missing) and delivered by generally surgeons, one fourth (24.5%) of total WRVUs were in super rural or rural areas (Exhibit V-11). The distribution across rurality designations differs by procedure category. For example, for procedure category “Endoscopy - Upper G.I.” 34.8% of WRVUs provided by general surgeons occurred in rural areas.

Analyzing the above 15 procedure categories across all surgeon specialties, general surgeons generated 23% of WRVUs in urban areas and 26% of WRVUs in rural areas (Exhibit V-12). These urban/rural proportions differ by procedure category. For the category “Endoscopy - Upper G.I.”, general surgeons generated 55% of WRVUs in urban areas and 69% of WRVUs in rural areas. The increase in use of general surgeons to provide Endoscopy - Upper G.I. procedures in rural areas appears to be driven largely by the reduction in procedures provided by thoracic surgeons.

Practice patterns of general surgeons in rural versus urban areas suggest that per capita use of general surgeons to provide care to Medicare beneficiaries is 13.1% higher in rural areas compared to urban areas. This accounts for the lower availability of specialist surgeons in rural areas. Based on this, HWSM assumes overall demand for general surgeons is 13.1% higher in rural areas than urban areas regardless of insurance type of patient. These geographic differences in practice patterns raise three issues for modeling:

  • Impact on overall national demand for general surgeons: To estimate starting year (2020) demand, HWSM models the assumption that supply and demand for general surgeons are roughly in equilibrium in metropolitan areas. However, in rural areas not only is there a shortfall from maldistribution but also additional demands on general surgeons due to a lack of specialist surgeons. Based on analysis of Medicare claims data, we calculate that the per capita demand for general surgeons in rural areas is 13.1% higher than the level that would be required if general surgeons had similar practice patterns to their metropolitan-based counterparts. Applying this 13.1% increase to general surgeon demand in rural areas equates to a 490 FTE (2%) national shortfall. Hence, we model national demand for general surgeons in 2020 as 24,979 FTEs (24,489 FTE supply plus 490 FTE shortfall). This 2% shortfall is possibly a conservative effort as rural area report challenges attracting and retaining general surgeons and concerns of a growing nationwide shortfall of general surgeons.22
  • Adjustments needed to the demand projections for general surgeons by metropolitan/non-metropolitan location: Comparing actual observed patterns of general surgeon use by Medicare patients relative to the national average for Medicare patients, and extrapolating these findings to all patients regardless of medical insurance type, we estimate that HWSM demand projections should be scaled by 1.099 for general surgeon demand in non-metropolitan areas and scaled by 0.972 for general surgeon demand in metropolitan areas.
  • Adjustment to the demand projections for specialist surgeons: HWSM models the assumption that at the national level supply and demand for specialist surgeons is in equilibrium. However, because HWSM does not explicitly model the increased demand for general surgeons in non-metropolitan counties to offset lack of specialist surgeons in those areas, HWSM could overstate demand for specialists in non-metropolitan counties and understate demand in metropolitan counties. Our analysis found that controlling for subset of services where specialist surgeons and general surgeons overlap, any adjustment to demand for specialist surgeons by metropolitan/non-metropolitan location is very small (an adjustment of 5 FTEs or fewer for each of the 14 surgical subspecialties affected). When modeled across all 50 states plus the District of Columbia, and specialty bias by metropolitan/non-metropolitan status is negligible and thus not modeled.

Demand Scenarios Modeled

As with other health occupations modeled, HWSM models demand for physicians under two scenarios as described in the other modules:

  • The Status Quo scenario models a continuation of recent (2015-2019) national patterns of care use extrapolated to the future population. This scenario captures geographic variation in demographics, health risk factors, disease prevalence, insurance coverage, level of rurality, and household income that can affect demand for physician services. Similarly, the scenario captures population growth and aging—and the associated implications for disease prevalence and other health risk factors—over the projection horizon. The scenario assumes national demand equal to national supply in 2020 with the exception of primary care and psychiatry. The scenario evaluates whether the nation’s future physician workforce is sufficient to provide at least the current level of care. Interpretation of projections should be within the context that advanced practice nurses and physician assistants overlap in scope of practice with physicians. Hence, imbalances between supply and demand for physicians should consider supply adequacy for advanced practice providers.
  • The Reduced Barriers scenario estimates the number of physician FTEs required if populations that historically faced barriers to accessing health care services demonstrated care use patterns comparable to populations perceived to have fewer barriers to accessing care. This scenario modifies the health care use patterns of nonmetropolitan county residents, racial and ethnic minority populations, and people without health insurance to the health care use patterns of their peers living in metropolitan counties, who are non-Hispanic White, and who have health insurance. Each of the three components of this scenario (non-metropolitan versus metropolitan county of residence, minority versus non-Hispanic White, and uninsured versus insured) is modeled in isolation and together to quantify the magnitude of each factor on demand for services. This hypothetical scenario describes the implications on physician demand if policies and programs reduced access-based disparities to health care services. The impact of reducing barriers to accessing care is modeled only for care provided in ambulatory settings and hospital settings. Reduced Barrier Demand projections for physician services delivered in nursing homes, residential care facilities, school-based settings, nurse education, and public health settings equal those of the Status Quo scenario.
Date Last Reviewed: