Forecasting the Future Supply of Pediatric Subspecialists in the United States: 2020–2040

The model forecasts future supply by taking the number of subspecialists in the current workforce and (1) adding the supply of fellows who complete training and then become certified by the ABP in that subspecialty, (2) subtracting pediatric subspecialists who exit the workforce by moving to nonclinical jobs or retiring (“leavers”), and (3) adding subspecialists who worked outside the United States and then return to provide clinical care (“out-of-country returners”). Additionally, the model accounts for the geographic diffusion of fellows from their training program to their first practice location and the movement of the existing workforce between different regions throughout their career. Model estimates can be viewed at the national, census region, and census division levels. The conceptual framework used to build the model is shown in Fig 2.

The model provides estimates of both the future headcount (HC) of subspecialists and HC adjusted for the proportion of time subspecialists spend in direct clinical or consultative care. As described in detail below, this adjustment uses self-report survey data and may not match clinical full-time equivalent (FTE) definitions used elsewhere.¹⁹  There is no consensus regarding the types of work (eg, clinical administrative work, call from home) that should be included in the calculation of clinical FTE. The Medical Group Management Association defines clinical FTE broadly as the number of hours worked on clinical activities.²⁰  A working group from the Association of Administrators in Academic Pediatrics and the Association of Medical School Pediatric Department Chairs (AMSPDC) is re-examining the definition of clinical FTE, with attention to billable clinical hours.²¹  Lacking a uniform definition, the model uses the term “clinical workforce equivalent.”

The model uses a dynamic microsimulation approach in which each pediatric subspecialist is treated as an “individual agent” with differing behaviors, depending on their age, gender, subspecialty, and geographic location. Microsimulation models are data-intensive and require data on (1) the current supply of pediatric subspecialists and their demographic, practice, and geographic characteristics, (2) the numbers of subspecialists entering and exiting the workforce, and (3) subspecialists’ individual behaviors such as their likelihood to exit the workforce, move between locations, or change the proportion of time they spend in clinical care.²² 

The microsimulation works by creating physicians in year 0 of the model based on the ABP’s pediatric subspecialty supply data in 2020. Then, for each year of the simulation, the model applies probabilities for each physician related to whether that physician (1) leaves the workforce, (2) moves location, or (3) changes their clinical time in that year. A “run” of an individual physician through the model is based on a computer-generated, random probability compared against the reference probability drawn from historical datasets. Each year the model adds a new supply of clinical fellows based on the number in training by subspecialty in 2019. This process is repeated for each year of the forecast.

Models were run 100 times to generate the mean supply forecast and confidence intervals (95%) around supply estimates that account for uncertainty. A Monte Carlo simulation approach was used to generate confidence intervals.²³  The formula was n = [zα/2 S/E]²  where S is the estimated SD of the data using the value of 150, which was derived from the headcount of the historical workforce (years 2010–2018) across census divisions and subspecialties, E is the desired margin of error, and zα/2 is the critical value for a 95% confidence interval. The equation yielded 85 runs, but 100 iterations were used to further improve model accuracy.

Uncertainty arises because of the random variation in behaviors in each year. For example, of the individual physicians working in 2020, most will remain in the workforce in all 100 runs of the model. Some will retire, move, or change their clinical time, leading to variation in the estimates represented by confidence intervals. Not surprisingly, confidence intervals are smaller at the national level and for larger subspecialties but wider for smaller geographies and subspecialties.

The model used the software language Python to generate the forecasts and an underlying PostgreSQL database held the data required to generate and store the forecasts. Because the model calculates outputs at the individual level, every model output represents the aggregation of individuals by subspeciality and from census divisions into census regions and national estimates. Stated differently, each output shown in the model is “rolled up” from individuals and therefore reflects more precise estimates because it uses historic behavior of real people to project the future.

A microsimulation approach is a well-established methodology used by the AAMC and HRSA in their physician workforce projections, and by the authors in other models forecasting the adequacy of the physician and nursing workforces.^11,24  Microsimulation is well-suited to modeling the pediatric subspeciality workforce because it accounts for how differences in age and gender affect physician behaviors in different subspecialites, including hours worked in clinical care, the decision to relocate practice, and the probability of exiting the workforce.²⁵ 

The data used to operationalize each of the model components and the assumptions regarding the data and model parameters are described below.

Deidentified data on the number and characteristics (eg, age, gender, subspecialty, geographic location) of pediatric subspecialists were derived from the 2020 ABP Certification Management System (ABP CMS). ABP CMS contains the most recent status reported by all physicians certified by the ABP. The model excluded physicians with an unknown age or an age greater than 70 years as the ABP is unable to determine when a physician retires; physicians with a non-United States or US Territory mailing address; and physicians who did not hold either active or permanent certification in at least 1 of the 14 subspecialties in the model. It was assumed that all physicians holding a current ABP certification were currently working, whether clinically or in other capacities.

The American Medical Association Masterfile (AMA MF) contains comprehensive information on physicians in the United States, including education, training, and professional certification. However, previous analyses comparing to the ABP certification data suggest that use of the AMA MF would likely overestimate the number of pediatric subspecialists in practice.²⁶  Therefore, despite representing a potential data source for enumerating the pediatric subspecialty workforce, it was not used in this study.

Physicians with more than 1 ABP subspecialty certification (eg, pediatric critical care medicine and pediatric cardiology) represent 1% of the pediatric subspecialty workforce. Including individuals with multiple certifications would have greatly increased the complexity of modeling and potentially overcounted the workforce. However, assigning physicians to 1 subspecialty could potentially understate the number of physicians in any given subspecialty—an effect that might be amplified for small subspecialties in less populous geographic regions. Analyses conducted to address these concerns showed that the states with the most subspecialists overall also had the most physicians with more than 1 ABP subspecialty certification. Therefore, assigning physicians to a single subspecialty would not have a significant effect on geographic regions with fewer subspecialists. To ensure that smaller subspecialties with fewer physicians were not disproportionately affected by assigning physicians to a single subspecialty, physicians with multiple certifications were assigned to the subspecialty with the fewest number of subspecialists. For example, a physician certified in both child abuse pediatrics and neonatal-perinatal medicine was assigned to child abuse pediatrics. Table 1 shows that, after making these assignments, the number of subspecialists used in the model represented 99% to 100% of total ABP certifications for the subspecialty. Pediatric cardiology had the highest proportion of physicians reassigned to other specialties; the number of pediatric cardiologists counted in the current pediatric cardiology workforce in the model represented 96% of the total number of ABP pediatric cardiology certifications.

The ABP’s Maintenance of Certification (MOC) enrollment survey is administered concurrently with re-enrollment in MOC for general pediatricians and subspecialists. Response rates for these surveys from 2009 to 2019 ranged from 63% to 99%. The survey asks pediatricians to report how many hours, on average, they worked per week in the past 6 months. It also asks pediatricians to report the proportion of time spent in 6 areas, with the amount of time across all areas totaling 100%. Those 6 areas include (1) administration, (2) direct and/or consultative inpatient and outpatient care, including patient billing and charting (with or without trainees), (3) medical education, (4) quality improvement activities, (5) research, and (6) others.

Using these questions, each subspecialist’s hours spent in clinical care were calculated by multiplying their total hours worked by their self-reported percent time in direct and/or consultative inpatient and outpatient care, including patient billing and charting (with or without trainees). Figure 3 shows how hours spent in clinical care varies, on average, as a proportion of total hours worked by gender and age across all 14 subspecialties. The model also accounts for variation between subspecialties in the proportion of time spent in clinical care by age and gender (data not shown). These adjustments allowed us to adjust (ie, reduce) HC to clinical workforce equivalent.

The MOC survey also asks, “At what age do you expect to retire?” Previous research suggests intention to exit the workforce often does not reflect actual retirement from the workforce.^27,28  Our analyses of individual subspecialists’ responses to the intention-to-retire question between 2009 and 2019 found that nearly half of physicians changed their plans with their anticipated age of expected retirement increasing as the subspecialist aged (data not shown). Because of the potential to overestimate workforce exit, MOC data on expected retirement was not used.

Instead, a longitudinal analysis of the AMA MF from 2011 to 2018 was conducted to calculate an annual probability of leaving (not just retiring) from the workforce based on the pediatric subspecialists’ age, sex, and subspecialty. Although previous analyses suggest that the AMA MF lags recording retirement,^29,30  the decision was made that it was better to underestimate, rather than overestimate, workforce exit in the model. Figure 4 shows the average annual probability of workforce exit by age is almost nonexistent <60 years across all pediatric subspecialities. The model accounts for variation in exit by age as well as gender and subspecialty (data not shown).

The National Resident Matching Program data are often used to identify the number of physicians entering training; however, previous work found that National Resident Matching Program data can underestimate the number of fellows entering training because the data exclude trainees admitted before or after the official Match.³¹  The model thus uses the ABP’s fellowship data, sourced annually for each ACGME’s accredited fellowship program (Table 2).

The model used the number of fellows entering training in 2019 for the baseline model and held it constant over the modeling period (2020–2040). We considered using an average to smooth the variation in numbers of fellows entering over time but decided it would overstate the number of entering fellows for subspecialties that had seen a recent decrease and would understate the number for those had recently increased their positions. From 2020 to 2040, this annual error would be compounded and cause the model to over- and underestimate new entrants and future supply. Thus, the model uses the number of entering fellows in 2019 and allows users to modify the number of fellows in training by using the scenarios described below.

Because trainees may extend their training for a variety of reasons, the model accounts for variation in the length of time spent in training. ABP CMS data were used to estimate the probability of remaining in subspecialty training each fellowship year and entering the workforce after passing the initial certification subspecialty examination. The model also accounted for attrition during training based on historical ABP CMS data from 2011 to 2019.

The model accounts for 3 types of geographic movements that pediatric subspecialists may make during their career: (1) fellows’ moves after they complete training, (2) moves that the existing workforce makes during their career, and (3) subspecialists’ moves from outside the country to provide care in the United States (“out-of-country returners”).

The diffusion of fellows from training to practice was derived by comparing, for each subspecialist, a fellow’s training location to their subsequent mailing address on file using data from 2017 and 2018 (n = 2530).

The bolded, diagonal data in Table 3 demonstrate that most recent subspecialists remained in the same census division where they trained, although a significant percentage moved to other divisions. The data also indicate that a small, but not insignificant percentage (between 1.3% and 5.6% of trainees), leave the United States. Although 99.0% of American medical graduates entered practice after training, only 92.9% of international medical graduates were in practice in the United States after completing training. The model accounted for different retention probabilities according to whether the trainee was an American medical graduates or international medical graduates.

Previous work suggests that about one-quarter of physicians will change the county where they practice over a decade.^32,33  However, we did not initially plan to account for these moves because (1) the number of pediatric subspecialists at the county and state level are relatively small, making estimates of moves unstable over time; (2) by using census division as the smallest unit of geography, it was assumed most moves would be within a division; and (3) the majority of moves would occur in the first stage of a pediatric subspecialist’s career, when fellows complete training and move to their first place of employment (Table 3).

During model refinement and validation, several cases emerged where some pediatric subspecialties had a significantly larger or smaller forecast at the US census region and census division level than would be expected given historic trends. Further analysis revealed that some geographic regions depended more heavily on an inflow of existing pediatric subspecialists who move to areas with lower supply from areas with higher supply. To reflect this finding, the model includes the diffusion of subspecialists between census regions and census divisions using the ABP CMS data from 2010 to 2020.

Even after the diffusion of the existing workforce was accounted for, cases still existed in which individual census division forecasts deviated from historic trends, particularly in geographic areas with smaller populations. Further examination of ABP CMS data from 2010 to 2020 found that the under-forecasted divisions depended on a greater supply of out-of-country returners (those subspecialists who previously worked outside the United States). Even though this supply is relatively small, it made a meaningful difference for less populous census divisions. Thus, out-of-country returners were also included. Table 4 shows the number of out-of-country returners added to the model by subspecialty and census division.

To allow users to compare supply projections between geographic areas with different sizes, the model includes rates of subspecialists per 100 000 children 0 to 18 years. Although not all 14 subspecialties see children of all ages (eg, neonatal-perinatal medicine, adolescent medicine), having a common population denominator allows users to compare supply projections between subspecialties. The population projections used in the model are from the University of Virginia Weldon Cooper Center, Demographics Research Group.^34,35  The model presented in this supplement used the 2020, 2030, and 2040 forecasts to interpolate the population in the years between forecasts. State population data were aggregated to the US census division, census region, and national level.

Scenarios are helpful tools to simulate what could happen if a given trend changes (ie, the number of fellows in training increases or decreases). They also enable users to explore what could happen if specific actions are, or are not, taken. A combination of peer-reviewed literature, the team’s modeling expertise, and stakeholder input from over 100 pediatricians from March to November 2022 was used to provide input on the magnitude, directionality, and duration of 10 scenarios included in the model. Table 5 outlines the 10 scenarios and the evidence used to inform these scenarios. This supplement’s prologue includes a detailed discussion of the health system, demographic, training, and workforce changes that could occur to produce these different “futures.”³⁶ 

As the baseline and scenarios were being developed and finalized, they were reviewed and validated through extensive stakeholder engagement, continuous quality checks, and “backcasting”—a process that involves checking model forecasts against actual workforce supply. The final model is the result of over 90 iterations of quality checks and refinement.

The baseline model and alternative scenarios were created with significant collaboration with pediatric subspecialist representatives from several pediatric organizations and workforce experts. These experts conducted multiple validity checks on the baseline model and on the scenarios throughout 2021 and 2022. The authors collected stakeholder feedback in the form of a semistructured, Qualtrics online survey that asked experts to assess the pediatric subspecialty model and through narrative data collected during seven meetings.

After receiving feedback from stakeholders, 2 additional scenarios were added to allow users to model the effect of larger increases and decreases in fellows in training.

As part of model validation, the number of subspecialists forecasted to be in the workforce in 2021 was compared with data on actual counts of subspecialists certified by the ABP in 2021 to identify cases in which actual numbers exceeded the bounds of confidence intervals for projected HC.

Most 2021 forecasts matched actual numbers. When differences arose, in most cases (84 of 210 forecasts), the model projection was lower than the actual number of pediatric subspecialists in some geographies. These underestimates could occur for several reasons. First, the data used to compare the forecast and the actual counts were snapshots from different months. Also, in 2021, a higher proportion of subspecialists chose to maintain their certification. This may reflect that the ABP gave MOC points to all pediatricians in recognition of the rapid quality improvement efforts undertaken to respond to the coronavirus disease 2019 (COVID-19) pandemic or decisions to stay in the workforce to address child health needs. Differences were not sufficiently significant or consistent in magnitude or direction to adjust the model.

Although the model makes significant substantive and methodological contributions to the field of pediatric subspecialty workforce planning, it is important to interpret findings in the context of some important limitations. The model draws on administrative data that may not accurately capture the supply of pediatric subspecialists in practice in the United States. Response rates to MOC surveys range from 63% to 99% and we have not analyzed potential nonresponder bias. Data on hours worked and clinical time are self-reported. Estimates of the geographic mobility are derived from mailing, not practice, addresses that may not be regularly updated by subspecialists. Workforce attrition is understudied, difficult to estimate, and likely to change given health system and economic conditions. The model is based on historic workforce trends that displayed some instability, particularly for smaller subspecialties and less populous geographic regions. The model used data from the pre-COVID-19 period to forecast future behaviors. Scenarios that reflect changes that may occur because the COVID-19 pandemic or other forces may not accurately reflect future trends.

Despite these limitations, the model is the first of its kind to model supply for all 14 subspecialties. It is accompanied by a publicly available data visualization tool that permits tailoring of scenarios and geographic areas.³⁷  The model reviewed in conjunction with other data can provide insight into the possible future of the pediatric subspecialty workforce and provide data to inform decision-making.

We thank Virginia A. Moyer and Patience Leino for their editorial support; and the pediatricians who shared their information with the American Board of Pediatrics Foundation and made this supplement possible.

AAMC

Association of American Medical Colleges

ABP

American Board of Pediatrics

ABP CMS

American Board of Pediatrics Certification Management System

AMA MF

American Medical Association Masterfile

AMG

American medical graduate

AMSPDC

Association of Medical School Pediatric Department Chairs

FTE

Full-time Equivalents

HC

headcount

HRSA

Health Resources and Services Administration

MOC

Maintenance of Certification