health workforce microsimulation model documentation · 2020-05-20 · exhibit 5 characteristics...
TRANSCRIPT
Health Workforce Microsimulation Model Documentation
Version 5.19.20
May 2020
Tim Dall Executive Director
Ryan Reynolds Senior Consultant
Ritashree Chakrabarti Senior Consultant
Will Iacobucci Senior Consultant
Kari Jones Associate Director
Life Sciences
IHS Markit | HWMM Documentation
© 2020 IHS Markit. All rights reserved. ii May 2020
Contents
Introduction .......................................................................................................................................... 1 Background ............................................................................................................................................ 1 Microsimulation model overview ............................................................................................................. 3 Healthcare Demand Microsimulation Model ....................................................................................... 5 Overview ................................................................................................................................................ 6 Population files ....................................................................................................................................... 7 Healthcare use patterns ....................................................................................................................... 10 Health workforce staffing patterns ........................................................................................................ 18 Scenarios ............................................................................................................................................. 19 Input summary ..................................................................................................................................... 21 Health Workforce Supply Model........................................................................................................ 22 Starting supply input files ..................................................................................................................... 22 New entrants ........................................................................................................................................ 23 Hours worked patterns ......................................................................................................................... 24 Labor force participation ....................................................................................................................... 27 Retirement ........................................................................................................................................... 27 Geographic migration ........................................................................................................................... 31 Scenarios ............................................................................................................................................. 32 Workforce implications of strategies to prevent or manage chronic disease ............................... 32 Model validation, strengths, and limitations .................................................................................... 35 Validation activities ............................................................................................................................... 36 Model strengths .................................................................................................................................... 36 Model limitations .................................................................................................................................. 37 References .......................................................................................................................................... 40
IHS Markit | HWMM Documentation
© 2020 IHS Markit. All rights reserved. iii May 2020
Exhibits
Exhibit 1 Integrated Health Workforce Microsimulation Model ............................................................ 4
Exhibit 2 Health occupations and specialties modeled........................................................................ 5
Exhibit 3 Schematic of Healthcare Demand Microsimulation Model ................................................... 6
Exhibit 4 Population database mapping algorithm............................................................................... 8
Exhibit 5 Characteristics available for each person in representative population sample .................. 9
Exhibit 6 Sample regressions: adult use of cardiology services ....................................................... 12
Exhibit 7 Patient characteristics on rate of primary care office visits for adults ................................ 13
Exhibit 8 Logistic regression for emergency department consultation .............................................. 15
Exhibit 9 Illustration of probability of emergency department consultation ....................................... 16
Exhibit 10 Average prescriptions per healthcare visit ........................................................................ 17
Exhibit 11 HDMM calibration: physician office visits .......................................................................... 18
Exhibit 12 Demand model input data summary ................................................................................. 22
Exhibit 13 Data sources for number and characteristics of new entrants ......................................... 24
Exhibit 14 OLS regression example: weekly patient care hours for general internal medicine ....... 25
Exhibit 15 OLS regression coefficients predicting weekly hours worked for select occupations ..... 26
Exhibit 16 Odds ratios predicting probability active ........................................................................... 27
Exhibit 17 Physician retirement patterns by age and sex .................................................................. 29
Exhibit 18 Probability male physician is still active by specialty and age ......................................... 30
Exhibit 19 Overview diagram of the Disease Prevention Microsimulation Model ............................. 34
Exhibit 20 Overview diagram of body weight component in DPMM ................................................. 35
IHS Markit | HWMM Documentation
© 2020 IHS Markit. All rights reserved. iv May 2020
Acronyms used in report
AACN American Association of Colleges of Nursing
AAPA American Academy of Physician Assistants
ACS American Community Survey
ADA American Dental Association
AMA American Medical Association
APRN Advanced practice nurse
BLS Bureau of Labor Statistics
BRFSS Behavioral Risk Factor Surveillance System
CDC Centers for Disease Control and Prevention
CMS Centers for Medicare and Medicaid Services
DPMM Disease Prevention Microsimulation Model
HDMM Healthcare Demand Microsimulation Model
HRSA Health Resources and Services Administration
HWSM Health Workforce Supply Model
IPEDS Integrated Postsecondary Education Data System
LPN/LVN Licensed practical/vocational nurse
MEPS Medical Expenditure Panel Survey
NAMCS National Ambulatory Medical Care Survey
NCLEX National Council Licensure Examination
NCSBN National Council of State Boards of Nursing
NCCPA National Commission on Certification of Physician Assistants
NHAMCS National Hospital Ambulatory Medical Care Survey
NIS National Inpatient Sample
NP Nurse practitioner
NSSRN National Sample Survey of Registered Nurses
PA Physician assistant
PCMH Patient centered medical home
RN Registered nurse
SNF Skilled Nursing Facility
Note: Earlier versions of this technical documentation are available upon request from [email protected].
IHS Markit | Title of Report
© 2020 IHS Markit. All rights reserved. 1 May 2020
Health Workforce Microsimulation Model Documentation
Version 5.19.2020
Introduction This report provides technical documentation of the health workforce microsimulation models developed by IHS
Markit, with contributions to model development from the various organizations for which studies have been
conducted using these models. The following section provides background information and an overview of the
workforce models. Next, we document the data, methods, assumptions and inputs for the demand model—
referred to as the Healthcare Demand Microsimulation Model (HDMM)—as well as the supply model—referred
to as the Health Workforce Supply Model (HWSM), and provide a brief overview of the Disease Prevention
Microsimulation Model (DPMM) used to model the workforce implications of strategies to prevent or manage
chronic disease.1 The final section describes work to validate the models, model strengths and limitations, and
areas of ongoing and future research. An appendix contains additional information about model inputs.
We continue to maintain and refine the models as new data and research become available; additionally, we
continue to develop new modules and scenario modeling capabilities. This documentation is intended to help
make the models transparent and provide the opportunity for feedback to improve these models. This report is
updated periodically to reflect refinements to the models and updated data sources. Hence, application of the
model to previous studies might have used earlier data sources than documented in this report.
Background
The workforce models described here are unique in their approach, breadth and complexity. Health workforce
projection models have been used for decades to assist with workforce planning and to assess whether the
workforce is sufficient to meet current and projected future demand (or need) at the local, regional, state, and
national levels. The models described here use a microsimulation approach where individual people (patients and
clinicians) are the unit of analysis. While microsimulation models have been used to study complex policy and
health issues2–6, the models described here are the first broad application of microsimulation modeling for
developing health workforce projections.
Approaches used historically in the U.S. to model the demand for health workers include: (1) convening expert
panels that consider patient epidemiological needs and provider productivity7; (2) extrapolating care use and
delivery patterns from beneficiaries in health maintenance organizations8,9; (3) extrapolating trends based on an
econometric approach of the correlation between provider-to-population and population characteristics and
economic measures10–12; and (4) developing demand models that use historical patterns of healthcare use and
delivery to create detailed provider-to-population ratios.a Such “macro” approaches that model demand at the
population level have limited ability to model policy changes or paradigm shifts in care delivery because most
coverage and treatment decisions are determined by individual patient circumstances. While approaches used
historically for modeling demand vary widely, the approach to supply modeling has been relatively similar across
studies, and models the likely workforce decisions of provider cohorts as they enter and progress through their
careers. Similar modeling approaches have been used across health professions.
Modeling approaches used in the past faced many challenges—data limitations, computing resources, and gaps in
research and understanding of health workforce issues. The use of microsimulation modeling to study the
a For example, workforce models used by the Health Resources and Services Administration from the 1990s to approximately 2012.
IHS Markit | HWMM Documentation
Confidential. © 2020 IHS Markit. All rights reserved. 2 May 2020
healthcare system was proposed in the early 1970s by Yett and colleagues, but data and computer computational
constraints prevented the full implementation of such a model.13 Improved computing power and wider access to
data and research have enabled development of more sophisticated workforce models that provide more reliable
projections and that can be forward looking in terms of a changing healthcare delivery and policy landscape. The
microsimulation models described here were designed to help address limitations of earlier models.
These microsimulation models have been adapted to model national, state and local area supply and demand for
many organizations. These include:
• Federal Bureau of Health Workforce (to model physicians, advanced practice providers, nurses, oral health
providers, behavioral health providers, and many other health occupations) at the national, state, and urban/rural
levels;14,15
• States—including Arkansas (primary care providers), Florida (physicians), Georgia (nurses, physicians, and
physician assistants), Hawaii (multiple occupations), Maryland (select physician specialties), New York (multiple
occupations), South Carolina (multiple occupations), Texas (multiple occupations), and Vermont (multiple
occupations);16–22
• Trade and professional associations;23–26
• Hospitals and health systems—including market assessment and regional planning, and the workforce implications
of strategies to restructure the healthcare delivery system;27–32 and
• Independent analyses.33,34
DPMM, which models strategies to prevent or manage chronic disease and the resulting implications for
healthcare use and provider demand, has also been used for work with:
• Life sciences companies -- to model burden of disease and strategies to prevent or delay onset of diabetes,
cardiovascular disease and other chronic conditions associated with obesity;35–38 and
• Trade associations and non-profit organizations -- to model burden of chronic disease and strategies to reduce
future burden including lifestyle interventions to promote improved diet and increased physical activity, smoking
cessation programs, improved screening and treatment, and improved medication adherence (to control blood
pressure, cholesterol, and blood glucose levels).39,40
The goals behind development and maintenance of these microsimulation models include:
• Providing the most accurate workforce supply and demand projections possible, as well as timely updates to reflect
the latest data, trends, policies, and research in the field;
• Informing strategies and policy decisions with health workforce implications;
• Integrating supply and demand across many occupations and specialties into a dynamic model; and
• Adapting the models to state and sub-state levels.
IHS Markit | HWMM Documentation
Confidential. © 2020 IHS Markit. All rights reserved. 3 May 2020
Microsimulation model overview
To provide maximum flexibility for adapting the model to different populations and to unique supply and demand
scenarios, these models use a microsimulation approach. As depicted in Exhibit 1, there are three major modeling
components: (1) modeling demand, (2) modeling supply, and (3) modeling disease management and prevention.
Consistent with recommended standards, we developed and validated self-contained modules that describe
different components of the healthcare system.41
• Demand: HDMM has three major components: (a) characteristics of each person in a representative sample of the
current and future population (demographics, socioeconomics, health-related behaviors, presence of chronic
conditions, insurance type/status, etc.), (b) healthcare use patterns that relate patient characteristics to annual use of
healthcare services by delivery setting and medical condition/provider specialty, and (c) staffing patterns that
translate demand for healthcare services into requirements for full time equivalent (FTE) providers by
occupation/specialty and by care delivery setting. Healthcare use and staffing patterns are influenced by changing
demographics and trends in care reimbursement and delivery.
• Supply: HWSM simulates workforce decisions for each person in a representative sample of providers based on
the person’s demographics, profession and specialty, and characteristics of the local or national economy and labor
market. Components include: (a) characteristics of the starting supply, (b) characteristics of new entrants to the
workforce, (c) attrition, (d) geographic mobility, and (e) work patterns.
• Disease management: DPMM simulates treatment/intervention scenarios to quantify their impact on preventing or
delaying onset of chronic disease and sequelae.
These three models are partially integrated as depicted by the dotted lines in Exhibit 1. For example, the available
supply influences staffing patterns; provider demand influences career decisions of individual providers; and
disease prevention and management strategies influence patient health outcomes and the derived demand for
services and providers. The three models are programmed in R, which is open source software.
IHS Markit | HWMM Documentation
Confidential. © 2020 IHS Markit. All rights reserved. 4 May 2020
Exhibit 1 Integrated Health Workforce Microsimulation Model
The health occupations and medical specialties included in this model are summarized in
Integrated Health Workforce Microsimulation Model
Source: IHS Markit © 2020 IHS Markit
IHS Markit | HWMM Documentation
Confidential. © 2020 IHS Markit. All rights reserved. 5 May 2020
Exhibit 2. Not all occupations are included in the supply analysis, often because of data limitations on entry and
exit from low compensated occupations with low barriers to entering the profession.
IHS Markit | HWMM Documentation
Confidential. © 2020 IHS Markit. All rights reserved. 6 May 2020
Exhibit 2 Health occupations and specialties modeled
Health occupations and specialties modeled
Occupations & Specialties Occupations & Specialties, cont.
Physicians & physician assistants Advanced practice nurses
Primary Care Nurse anesthetists
Family Medicine Nurse midwives
General Internal Medicine Nurse practitioners (by specialty)
Geriatric Medicine Nursing
General Pediatrics Registered nurses
Medical Specialties Licensed practical/vocational nurses
Allergy & Immunology Nurse assistants/aides (incl. home health)
Cardiology Behavioral health (incl. psychiatrists and NPs/PAs)
Critical Care/Pulmonology Psychologists
Dermatology Addiction counselors
Endocrinology Social workers
Gastroenterology Mental health counselors
Hematology & Oncology School counselors
Infectious Disease Marriage and family therapists
Neonatal-perinatal Oral health
Nephrology General dentists
Rheumatology Specialist dentists
Surgery Dental hygienists
General Surgery Pharmacy
Colorectal Surgery Pharmacists
Neurological Surgery Pharmacy technicians
Obstetrics & Gynecology Pharmacy aides
Ophthalmology Respiratory care (therapists & technicians)
Orthopedic Surgery Rehabilitation Services
Otolaryngology Occupational therapists & assistants
Plastic Surgery Physical therapists & assistants
Thoracic Surgery Therapeutic Services
Urology Chiropractor
Vascular Surgery Podiatrists
Other Specialties Vision Services
Anesthesiology Opticians
Emergency Medicine Optometrists
Neurology Nutritionists
Pathology Select diagnostic laboratory professions
Physical Medicine & Rehabilitation Select diagnostic imaging professions
Psychiatry Long term services and support professions
Radiation Oncology
Radiology
Other Med Spec
Hospitalist
Source: IHS Markit © 2020 IHS Markit
Healthcare Demand Microsimulation Model This section provides a brief overview of HDMM and describes creation of the major components: the population
file, healthcare use prediction equations, and provider staffing parameters. Data sources and methods for
producing national, state, and county demand projections are described. A description of the scenarios HDMM
was designed to model is also provided.
IHS Markit | HWMM Documentation
Confidential. © 2020 IHS Markit. All rights reserved. 7 May 2020
Overview
HDMM models demand for healthcare services and the number of providers required to meet demand for
services. Demand is defined as the level and mix of healthcare services (and providers) that are likely to be used
based on population characteristics and economic considerations such as price of services and people’s ability and
willingness to pay for services. HDMM was designed also to run a limited set of scenarios around “need” for
services. Need is defined as the healthcare services (and providers) required to provide a specified level of care
given the prevalence of disease and other health risk factors. Need is defined in the absence of economic or
cultural considerations that might preclude someone from using available services. Other scenarios model the
evolving care delivery system.
HDMM has three major components: (1) a population database with information for each person in a representative
sample of the population being modeled, (2) healthcare use patterns that reflect the relationship between patient
characteristics and healthcare use, and (3) staffing patterns that convert estimates of healthcare demand to estimates
of provider demand (Exhibit 3). Demand for services is modeled by employment or care delivery setting. Demand is
also modeled by (a) diagnosis category for hospital inpatient care and emergency department visits, and (b)
healthcare occupation or medical specialty for office, outpatient and home health visits. The services demand
projections are expressed in terms of workload measures, and demand for each health profession is tied to one or
more of these workload measures. For example, current and future demand for primary care providers is tied to
demand for primary care visits, demand for dentists is tied to projected demand for dental visits, etc. External
factors—such as trends or changes in care delivery—can influence all three major components of HDMM.
Exhibit 3 Schematic of Healthcare Demand Microsimulation Model
Schematic of Healthcare Demand Microsimulation Model
Source: IHS Markit © 2020 IHS Markit
IHS Markit | HWMM Documentation
Confidential. © 2020 IHS Markit. All rights reserved. 8 May 2020
Population files
The population files used in the model contain person-level data for a representative sample of the population of
interest. The population of interest might be the entire U.S., an individual state, a county within a state, or some
other geographic unit such as a region, metropolitan area, or hospital service area defined by a set of ZIP codes.
When a population file is created for a specified area, demand estimates can be produced for subsets of the
population—e.g., subsets defined by insurance type, patient demographic, or other tracked characteristic of the
population. Prior to 2019, the population database was created at the state level and could be aggregated to the
national level. Starting in 2019, the population files were constructed for each of the 3,142 counties or county
equivalents in the U.S. The county population files can be summed to produce either state or national estimates
and by National Center for Health Statistics (NCHS) urban-rural county designation.42 The population file is
updated each November to incorporate the latest versions of the following data sources:
• American Community Survey (ACS). Each year the Census Bureau collects information on approximately three
million individuals grouped into roughly one million households. For each person, information collected includes
demographics, household income, medical insurance status, geographic location (e.g., state and sub-state [for
multi-year files]), and type of residency (e.g., community-based residence or nursing home).
• U.S. Census Bureau Population Estimates. The U.S. Census Bureau produces current population totals for each
county by demographics including five-year age groups, sex, and race/ethnicity.
• Behavioral Risk Factor Surveillance System (BRFSS). The Centers for Disease Control and Prevention (CDC)
annually collects data on a sample of over 500,000 individuals. Similar to the ACS, the BRFSS includes
demographics, household income, and medical insurance status for a stratified random sample of households in
each state. The BRFSS, however, also collects detailed information on presence of chronic conditions (e.g.,
diabetes, hypertension) and other health risk factors (e.g., overweight/obese, smoking). One limitation of BRFSS is
that as a telephone-based survey it excludes people in institutionalized settings (e.g., nursing homes) who do not
have their own telephone. We combine the latest two years of BRFSS files to provide records for approximately
one million individuals. Since BRFSS reports some variables biennially (e.g., hypertension, which is omitted from
the even year files), we used a predictive equation to estimate the probability of having those conditions in even
years based on known characteristics of the individual.
• Medicare Beneficiary Survey (MCBS). Starting in 2017, the health characteristics of the residential care
population were modeled using individuals in the MCBS living in residential care facilities (with the 2017 MCBS
data being the most recent available). Prior to 2017, individuals living in residential care were merged with the
BRFSS—thus taking on the health risk profile characteristics of a community-based population that is healthier, on
average, than the population in residential care facilities.
• CMS’s Long-Term Care Minimum Data Set (NHMDS). Starting in 2017a, we used the NHMDS to develop a
representative sample of residents in nursing homes in each state. This data source contains information on disease
prevalence and health risk factors for each person residing in a nursing home. From the NHMDS we drew a
random sample of resident records where the size of each sample was determined based on CMS published data of
the average number of nursing home residents in each state by age group.
Creation of the state population database merges information from these sources using a statistical matching
process that combines patient health information from the BRFSS, MCBS and NHMDS with the larger ACS file
a Previously, we used the 2004 National Nursing Home Survey (NNHS) combined with CMS estimates of nursing home residents in each state to develop a representative sample of the nursing home population in each state. The NNHS collected information on chronic conditions and health risk factors of this
population.
IHS Markit | HWMM Documentation
Confidential. © 2020 IHS Markit. All rights reserved. 9 May 2020
that has a representative population in each state (Exhibit 4). Creation of county-level population files uses a
similar process that is described later.
For the non-institutionalized population, each individual in the ACS was matched with someone in the BRFSS
from the same gender, age group (15 age groups), race, ethnicity, insured/uninsured status, household income
level (8 income categories), and state of residence.a Individuals categorized as residing in a residential care facility
or nursing home were randomly matched to a person in the MCBS or NHMDS, respectively, in the same state,
age group, gender, and race and ethnicity strata. Under this approach, some BRFSS, MCBS or NHMDS
individuals might be matched multiple times to similar people in the ACS, while some BRFSS or NHMDS
individuals might not be matched. The match probability for BRFSS and MCBS reflects the surveys’ sample
weights, with survey participants having higher sample weight more likely to be sampled.
Exhibit 4 Population database mapping algorithm
Exhibit 5 summarizes the population characteristics available in each source file and the characteristics used for
the statistical match process. This detailed information for each person captures systematic geographic variation in
demographics, socioeconomic characteristics, and health risk factors (e.g., obesity, smoking, diabetes and
cardiovascular disease prevalence) that reflect regional differences in diet, physical activity, and other health-
related behavior.
a The first round of BRFSS-ACS matching produced a match in the same strata for 94% of the population. To match the remaining 6%, the eight income levels were collapsed into four (1% matched), then the race/ethnicity dimension was dropped (1% matched), and then the same criteria as the first round was applied except
State was removed as a strata (remaining 4% matched), and finally for the fifth round only demographics were included (remaining 0.1% matched).
Community based
Residential care
facilities
Nursing
homes
CMS Nursing Home Minimum Data Set
Medicare Current Beneficiary Survey
Behavioral Risk Factor Surveillance
System
Population demographics Population health characteristics sourcesPopulation database mapping algorithm
Source: IHS Markit © 2020 IHS Markit
IHS Markit | HWMM Documentation
Confidential. © 2020 IHS Markit. All rights reserved. 10 May 2020
Exhibit 5 Characteristics available for each person in representative population sample
Characteristics available for each person in representative population sample
Population Characteristics Match Strata Source
AC
S-B
RF
SS
AC
S-M
CB
S
AC
S-N
MM
DS
AC
S (
2018)
BR
FS
S
(2017 &
2018)
MC
BS
(2017)
NM
MD
S
(2017)
Demographics Children age groups: 0-2, 3-5, 6-13, 14-17
Adult age groups: 18-34, 35-44, 45-64, 65-74, 75+
✓b ✓ ✓ ✓ ✓ ✓ ✓
Sex: male, female ✓ ✓ ✓ ✓ ✓ ✓ ✓
Race/ethnicity: non-Hispanic white, non-Hispanic black, non-Hispanic other, Hispanic
✓ ✓ ✓ ✓ ✓ ✓ ✓
Health-related lifestyle indicators a
Body weight: normal, overweight, obese ✓ ✓ ✓
Current smoker status ✓ ✓ ✓
Socioeconomic conditions and insurance
Family income (<$10,000, $10,000 to <$15,000, $15,000 to < $20,000, $20,000 to < $25,000, $25,000 to < $35,000, $35,000 to < $50, 000, $50,000 to < $75,000, $75,000+)
✓ ✓ ✓
Medical insurance type (private, public, self-pay) ✓ ✓ ✓ ✓
In a managed care plan ✓
Chronic conditions
Diagnosed with asthma ✓ ✓ ✓
Diagnosed with arthritis, heart disease, diabetes, hypertension a ✓ ✓ ✓
History of cancer, heart attack, or stroke a ✓ ✓ ✓
Geographic location
State (or other geographic area such as county) ✓ ✓ ✓ ✓ ✓ ✓ ✓
Living in a metropolitan area ✓ Notes: a Characteristics available only for adults. b Fifteen age groups are used for the statistical match process: ages 0-19, 20-24, 25-29, 30-34, 35-39, 40-44, 45-49, 50-54, 55-59, 60-64, 65-69, 70-74, 75-79, 80-84, and 85+. Then, individual ages are used to create the nine age groups above for modeling demand for healthcare services. The smaller number of age groups used for modeling demand for healthcare services reflects smaller sample size in the data sources used for modeling patterns of healthcare use.
Source: IHS Markit © 2020 IHS Markit
The ACS provides a representative sample of the population in each state for the most current year available, with
sample weights that can be aggregated to produce state (or national) totals. Developing demand forecasts for
future years requires incorporating state-specific population projections developed by state governments or other
organizations such as universities, and national population projections developed by the U.S. Census Bureau.
Using the population projections, we developed new sample weights for each individual that when aggregated
produce population estimates for each future year consistent with published population projections. The model’s
status quo demand scenario assumes that base year prevalence rates of health and health behavior characteristics
within each demographic group (by age, gender, race and ethnicity) remain the same over the projection
horizon—though HDMM can model scenarios where disease prevalence and health behavior characteristics
change within demographic strata such as modeling a population health scenario related to changes in modifiable
health risk factors.
Based on this constructed state population file, the next step is to develop the population file at the county level.
The U.S. Census Bureau produces annual data on the total population in each county by five-year age bands, sex
and race/ethnicity. We re-weight the sample weights for each metropolitan and non-metropolitan individual in a
state’s population file to match the demographics of the population characteristics in each metropolitan and non-
IHS Markit | HWMM Documentation
Confidential. © 2020 IHS Markit. All rights reserved. 11 May 2020
metropolitan county, respectively, using the published Census Bureau population data. This produces a weighted
sample that is representative of the demographics in each county. Further, county-level estimates of disease
prevalence are calibrated at the individual level to match with external published information for each county.
BRFSS data from state BRFSS surveys is the primary source for external county-level statistics used for
calibrating prevalence of diseases and risk factors in the population files.
The resulting constructed population file contains a representative sample of adults and children in each county by
demographics, insurance type, prevalence of disease and health risk factors, with household income and residence
type (community, residential care, or nursing home) reflective of the demographics in the county.
Healthcare use patterns
Projected future use of healthcare services, based on population characteristics and patterns of health-seeking
behavior, produce workload measures used to project future demand for healthcare providers. HDMM uses
prediction equations for healthcare use based on recent patterns of care use, but also can model scenarios where
healthcare use patterns change in response to emerging care delivery models, policy changes, or other factors.
Health seeking behavior is generated from econometrically estimated equations using data from ~170,000
participants in five years (2013-2017) of pooled files of the Medical Expenditure Panel Survey (MEPS). Pooling
multiple years of data increases sample size for regression analysis for smaller health professions and lower
frequency diagnosis categories. Over time, as a new year of data becomes available and is added to the analytic
file the oldest year in the analysis file is dropped. We used the 2017 Nationwide Inpatient Sample (NIS), with ~8
million discharge records, to model the relationship between patient characteristics and length of hospitalization
by primary diagnosis category.
Many of the population characteristics such as demographics and socioeconomic circumstances are likely
correlated with cultural and other factors (e.g., access constraints) that influence use of healthcare services and are
omitted from the regressions due to data limitations. Consequently, the observed relationship between annual use
of healthcare services and observed patient characteristics reflects correlation rather than causation.
Negative Binomial regression was used to model annual office visits, annual outpatient visits, and annual home
health/hospice visits. Prior to 2019, Poisson regressiona was used to model annual visits by provider occupation or
specialty. From 2019, various regression models were evaluated in response to issues of over-dispersion in the
Poisson model and the negative binomial regression model was selected as the alternative. This change had
negligible impact on the demand projections but conceptually is more appropriate given the large percentage of
patients with no visits to certain types of providers. These regressions were estimated separately for children
versus adults. Separate regressions were estimated by physician specialty or non-physician occupations—e.g.
dentists, physical therapists, psychologists—for office-based care. Likewise, separate regressions were estimated
for occupations providing home healthcare. The dependent variable was annual visits (for office, outpatient, and
home health). The explanatory variables were the patient characteristics available in both MEPS and the
constructed population file (Exhibit 6).
Logisticb regression was used to model annual probability of hospitalization and annual probability of emergency
department visit for approximately two dozen categories of care defined by primary diagnosis code. The
a Poisson regression is often used when the dependent variable (annual visits) is a count variable with a skewed distribution—i.e., many people have 0, 1, or 2,
visits, but the number of people with higher volume of visits (3, 4, 5, etc.) declines at the higher volume levels.
b Logistic regression is often used when the dependent variable is binary (yes/no). The sample size of MEPS is too small to accurately model patients with multiple
hospitalizations and multiple emergency department visits—especially when modeling at the diagnosis category level.
IHS Markit | HWMM Documentation
Confidential. © 2020 IHS Markit. All rights reserved. 12 May 2020
dependent variable for each regression is whether the patient had a hospitalization (or ED visit) during the year for
each of the condition categories (these categories were defined using the ICD-9 and ICD-10 codes). For
hospitalized patients, we used Poisson regression with NIS data to model hospital length of stay given the
condition category and patient information (age, sex, race/ethnicity, insurance type, presence of diabetes, and
urban-rural residency).
The model contains several hundred prediction equations for healthcare use, with examples of the regression
output for cardiology care presented in Exhibit 6 and for primary care presented in Exhibit 7. The numbers in
Exhibit 6 reflect either rate ratios (for office and outpatient visits, or inpatient days) or odds ratios (for ED visits
and hospitalizations). For all types of cardiology-related care there is a strong correlation with patient age
(controlling for other patient characteristics modeled). For example, relative to patients age 75 or older, patients
age 65-74 have only 80% as many office visits but have 18% more outpatient visits, although only the office visits
estimate is statistically different from 1.0 (where a ratio of 1.0 would indicate no statistical difference with the
comparison category). Patients age 65-74 have lower odds of a cardiology-related ED visit (i.e., primary diagnosis
was cardiology-related), and lower odds of a cardiology-related hospitalization. However, the length of
hospitalization averages 94% as long as the hospitalization for the age 75 or older patient.
Blacks tend to have fewer office visits than whites, but higher odds of ED visits or hospitalizations and longer
average length of hospital stay. Obesity is associated with increased use of cardiology-related services. Smoking
is associated with fewer office and outpatient visits to a cardiologist but higher rates of ED visits (likely reflecting
correlation rather than causality in the case of ambulatory care, as smoking is a risk factor for heart disease but
could be correlated with aversion to visit a doctor). Lower income is associated with less use of ambulatory care
and more use of ED visits and hospitalization. Having any medical insurance is associated with much greater use
of ambulatory care, and if the insurance is Medicaid then there is even greater use of cardiology services across all
care delivery settings. The presence of chronic medical conditions—and especially heart disease, hypertension,
and history of heart attack—are associated with much greater use of cardiology services across care delivery
settings. In general patients living in either small/medium metro or suburban large metro fringe areas tend to have
fewer ambulatory visits compared to those living in a large core metro area. Regression equations for other types
of care (whether by medical specialty or condition category) exhibit similar patterns that are consistent with
expectations and the health research literature.
IHS Markit | HWMM Documentation
Confidential. © 2020 IHS Markit. All rights reserved. 13 May 2020
Exhibit 6 Sample regressions: adult use of cardiology services
Sample regressions: adult use of cardiology services
Parameter a Office visitsb Outpatient visitsb Emergency visitsc Hospitalizationsc Inpatient daysd
Age
18-34 years 0.10** 0.35** 0.44** 0.19** 0.80**
35-44 years 0.20** 0.49** 0.69** 0.47** 0.74**
45-64 years 0.38** 0.83 0.67** 0.57** 0.84**
65-74 years 0.80** 1.18 0.84* 0.85 0.94**
75+ years 1.00 1.00 1.00 1.00 1.00
Male 1.09** 1.18* 0.77** 1.04 1.00**
Race- Ethnicity
Non-Hispanic White 1.00 1.00 1.00 1.00 1.00
Non-Hispanic Black 0.77** 1.08 1.20** 1.19* 1.11**
Non-Hispanic Other 0.96 0.78 1.02 0.96 1.01**
Hispanic 0.90* 0.57** 0.88 0.90 0.98**
Body Weight
Normal 1.00 1.00 1.00 1.00
Overweight 1.04 1.00 1.06 1.08
Obese 1.10* 1.04 1.27** 1.04
Current Smoker 0.80** 0.76* 1.22** 1.13
Household Income
<$10,000 0.96 1.63** 1.35** 1.37**
$10,000 to <$15,000 0.95 1.21 1.38** 1.26*
$15,000 to < $20,000 0.95 0.89 1.26* 1.29*
$20,000 to < $25,000 0.91 1.02 1.44** 1.24
$25,000 to < $35,000 0.92 1.11 1.26** 1.21
$35,000 to < $50,000 0.85** 1.08 1.20* 1.09
$50,000 to < $75,000 0.93 1.16 1.16 1.02
$75,000 or higher 1.00 1.00 1.00 1.00
Insurance
Has insurance 2.58** 2.87** 0.71** 1.05** 1.04**
In Medicaid 1.22** 1.38** 1.58** 1.49** 1.10**
In managed care plan 1.01 0.80** 1.08 0.96
Diagnosed with
Arthritis 1.21** 1.54** 1.08 1.04
Asthma 1.14** 1.23 1.28** 1.17*
Diabetes 1.20** 1.15 1.12* 1.51** 1.14**
Heart disease 7.73** 9.58** 2.49** 3.62**
Hypertension 2.00** 1.43** 5.22** 2.81**
History of cancer 1.23** 1.46** 1.07 1.07
History of heart attack 1.73** 2.05** 2.31** 2.73**
History of stroke 1.18** 1.12 2.07** 2.33**
Urban-Rural Areas e
Non-core 1.07 1.12 1.06 1.03
Micropolitan 0.93 0.83 0.95 1.00
Small metro 0.83** 0.88 0.93 0.98 1.03** Medium metro 0.72** 0.78 1.06 0.90
Suburban 0.77** 0.85 0.88 1.10
Large metro core 1.00 1.00 1.00 1.00
Notes: Statistically different from 1.00 at the 0.05 (*) or 0.01 (**) level. a For children the age categories are 0-2, 3-5, 6-12, and 13-17). The adult regressions include everyone age 18 and older. Variables not available for use in the regression equations for children are body weight, smoking status, and diagnoses of the chronic conditions listed (except for asthma which is included). b Rate ratios based on negative binomial regression of MEPS data. Dependent variable is annual visits to cardiologist. c Odds ratios based on logistic regression of MEPS data. Dependent variable is whether a patient had an emergency visit or hospitalization with a cardiology-related primary diagnosis code. d Rate ratios based on Poisson regression of NIS data. Dependent variable is length of stay conditional on hospitalization for cardiology-related primary diagnosis. e NCHS urban-rural categories can be aggregated to metro and non-metro areas; non-core and micropolitan are mapped to non-metropolitan area and the rest are mapped to metropolitan area. The reference population for comparison is age 75 or older, female, non-Hispanic white, normal body weight, non-smoker, household income of $75,000 or higher, uninsured, without the diagnosed conditions listed, residing in large metro core areas.
Source: IHS Markit © 2020 IHS Markit
IHS Markit | HWMM Documentation
Confidential. © 2020 IHS Markit. All rights reserved. 14 May 2020
Office visits by adults to a family medicine (FM) or general internal medicine (GIM) provider are presented for
comparison (Exhibit 7). The bars represent the percent difference in annual office visits contributed by each
characteristic controlling for other patient characteristics and relative to the reference population. Many of the
patient characteristics correlated with use of primary care services are similar to characteristics associated with
greater use of cardiologist services—e.g., the presence of chronic conditions like cardiovascular disease and
diabetes. Higher family income and residing in a metropolitan are associated with greater use of GIM services but
lower use of FM services.
Exhibit 7 Patient characteristics on rate of primary care office visits for adults
IHS Markit | HWMM Documentation
Confidential. © 2020 IHS Markit. All rights reserved. 15 May 2020
For care provided in the emergency department we link demand for emergency physicians to total demand for
emergency visits (so 10% growth in visits would translate to 10% growth in demand for emergency physicians
under the status quo scenario). Specialist physicians sometimes provide consults for emergency visits, and the mix
of patients and their diagnoses are expected to change over time. Using the 2015 and 2016 NHAMCS we
estimated a logistic regression where the dependent variable was whether during the visit a second physician was
seen. As summarized in Exhibit 8, the explanatory variables include specialty category (defined by visit primary
diagnosis), patient demographics (age, sex, and race), insurance status and whether insured through Medicaid, and
whether the patient lives in a metropolitan or non-metropolitan location. As illustrated by the odds ratios, the
likelihood that a specialist physician will be consulted during the visit differs by condition category, but in general
a second physician is most likely to be consulted if the patient’s primary diagnosis is related to nephrology,
neonatal medicine, vascular surgery, or cardiology. Patients with a primary diagnosis related to dermatology,
otolaryngology, or rheumatology are much less likely to see a second physician during their ED visit. Consults are
more likely for older patients, males, insured, not on Medicaid, and living in a metropolitan area.
For illustration, applying the logistic regression results to a female patient age 65-74, non-Hispanic white, and
living in a metropolitan area produces the following probabilities of having a consult tied to the primary diagnosis
for the emergency visit (Exhibit 9). The probabilities range from a high of 34% if the primary diagnosis is in the
category of nephrology, to a low of 8% is the primary diagnosis is in the category of otolaryngology or
rheumatology.
IHS Markit | HWMM Documentation
Confidential. © 2020 IHS Markit. All rights reserved. 16 May 2020
Exhibit 8 Logistic regression for emergency department consultation
Logistic regression for emergency department consultation
Parameter Odds Ratio 95% Confidence Interval
Diagnosis category (General Surgery comparison group) a
Cardiology 2.67 2.18 3.26
Dermatology 0.78 0.61 0.99
Endocrinology 1.44 1.10 1.88
Gastroenterology 1.14 0.94 1.37
Hematology 2.57 1.92 3.43
Infectious Disease 1.00 0.76 1.30
Neonatal Medicine 2.98 1.35 5.89
Nephrology 3.55 2.21 5.60
Neurological Surgery 1.50 0.94 2.31
Neurology 1.15 0.92 1.43
Obstetrics & Gynecology 2.21 1.76 2.77
Ophthalmology 1.14 0.76 1.66
Orthopedic Surgery 0.95 0.80 1.14
Otolaryngology 0.64 0.42 0.93
Other Specialties 1.21 1.00 1.47
Plastic Surgery 0.86 0.38 1.70
Psychiatry 2.36 1.96 2.86
Pulmonology 1.36 1.15 1.60
Rheumatology 0.64 0.45 0.89
Thoracic Surgery 1.85 1.55 2.22
Urology 1.07 0.90 1.29
Vascular Surgery 2.74 1.05 6.38
Female 0.90 0.84 0.97
Age (45-64 comparison group)
0-2 0.34 0.27 0.41
3-5 0.44 0.34 0.55
6-12 0.47 0.39 0.57
13-17 0.67 0.56 0.80
18-34 0.62 0.56 0.69
35-44 0.69 0.60 0.78
65-74 1.32 1.16 1.49
75+ 1.67 1.49 1.87
Race/ethnicity (non-Hispanic white comparison group)
Hispanic 1.46 1.33 1.61
Non-Hispanic black 1.03 0.94 1.13
Non-Hispanic other 1.29 1.07 1.55
Has medical insurance 1.35 1.18 1.54
Insurance is Medicaid 0.83 0.76 0.91
Lives in metropolitan area 3.09 2.72 3.53
2015 (vs 2016) 0.88 0.82 0.94
Source: Logistic regression analysis of the 2015 and 2016 NHAMCS. a Diagnosis categories defined by ICD-9 diagnosis and procedure codes to reflect types of care most likely provided by a physician specialty.
Source: IHS Markit © 2020 IHS Markit
IHS Markit | HWMM Documentation
Confidential. © 2020 IHS Markit. All rights reserved. 17 May 2020
Exhibit 9 Illustration of probability of emergency department consultation
Demand for medications is the workload driver to model demand for pharmacy-related health occupations. The
NAMCS and NHAMCS indicate prescription medications ordered by a health provider, though this is used as a
proxy for number of prescriptions filled (under the assumption that the ratio of prescribed-to-filled remains
relatively constant over time). Patients who visit a cardiologist in an office setting average 6.11 prescriptions per
visit, for example, while for primary care visits the average is 3.82 prescriptions per visit (Exhibit 10). To model
projected growth in demand for pharmacy-related occupations, under the status quo scenario, provider demand is
tied to projected growth in number of prescriptions.
IHS Markit | HWMM Documentation
Confidential. © 2020 IHS Markit. All rights reserved. 18 May 2020
Exhibit 10 Average prescriptions per healthcare visit
Average prescriptions per healthcare visit
Physician Specialty Office Outpatient Emergency
Nephrology - 5.43 3.58
Cardiology 6.11 4.20 2.76
Vascular Surgery - 3.02 2.98
Endocrinology - 4.03 2.75
Thoracic Surgery - 3.18 2.01
Pulmonology - 2.95 2.65
Neurology 3.72 2.90 2.59
Gastroenterology - 2.94 2.66
Hematology & Oncology - 3.58 2.67
Psychiatry 2.30 2.16 1.62
Rheumatology - 2.66 1.76
Urology 3.36 2.42 3.01
Orthopedic Surgery 2.46 2.49 2.07
Allergy & Immunology - 2.70 1.98
Dermatology 2.38 2.64 2.23
Plastic Surgery - 1.79 2.28
Ophthalmology 2.80 1.78 1.68
Otolaryngology 2.75 2.19 2.12
Primary Care 3.82 - -
General Surgery 2.22 1.91 1.76
OBGYN 1.80 1.83 1.96
Neurological Surgery - 1.67 1.81
Neonatal-perinatal - 1.15 1.04
Other Med Spec 3.78 1.77 1.45
Note: Average prescriptions per visit based on analysis of 2013-2015 combined NAMCS and 2011-2015 combined NHAMCS files.
Source: IHS Markit © 2020 IHS Markit
To model demand for oral health services we analyzed the MEPS Dental Visits File for the period 2012-2016. The
combined file was used to model annual visits to dental hygienists, and annual visits to each type of dentist
including general or pediatric dentist, endodontist, orthodontist, periodontist and other type of dentist. The
regressions were estimated separately for adults and children. MEPS does not identify pediatric dentists as a
unique specialty, and so using MEPS we cannot indicate whether dental services provided to children were by a
pediatric dentist or a general dentist. Information from ADA’s survey of dental practices allowed us to model the
proportion of dental visits by children and adolescents that likely were to general dentists and pediatric
dentists.26,43
These regressions quantify the relationship between patient characteristics and annual oral health visits similar to
the regression output summarized in Exhibit 6. The regression results show that use of oral health services is
highly correlated with insurance status (where medical insurance is used as a proxy for dental insurance),
household income, living in a metropolitan area, patient age, and race/ethnicity.
MEPS is a representative sample of the non-institutionalized population, and although the healthcare use
prediction equations are applied to a representative sample of the entire U.S. population, parts of the model
require calibration to ensure that at the national level the predicted healthcare use equals actual use. Applying the
prediction equations to the population for 2016 through 2017 creates predicted values of healthcare use in those
years (e.g., total hospitalizations, inpatient days, and ED visits by specialty category, and total office visits by
physician specialty). For model calibration, we compared predicted national totals to estimates of national total
hospitalizations and inpatient days, by diagnosis category, derived from the 2017 NIS. Comparative national
estimates of ED visits and office visits came from the 2016 NHAMCS and 2016 NAMCS, respectively.
IHS Markit | HWMM Documentation
Confidential. © 2020 IHS Markit. All rights reserved. 19 May 2020
Multiplicative scalars were then created by dividing national estimates by predicted estimates. For example, if the
model under-predicted ED visits for a particular diagnosis category by 10% then a scalar of 1.1 was added to the
prediction equation for that diagnosis category.
Applying this approach to diagnosis/specialty categories, the model’s predicted healthcare use was consistent with
national totals for most settings (see Exhibit 11 for calibration scalars for physician office visits). Setting/category
combinations where the model predicted less accurately (and therefore required larger scalars) tended to cluster
around diagnosis categories in the ED characterized by lower frequency of visits likely due to a combination of
small sample size in both MEPS and NAMCS.
Exhibit 11 HDMM calibration: physician office visits
HDMM calibration: physician office visits
Specialty
NAMCS Visits (in thousands), 2016 a
HDMM Initial Visits Pre-Scalar (in thousands), 2018 Scalar
Family Medicine 202,494 411,955 0.492
Pediatrics 136,119 81,775 1.665
Internal Medicine 81,701 72,292 1.130
Obstetrics & Gynecology 73,198 80,804 0.906
Orthopedic Surgery 30,114 124,001 0.243
Ophthalmology 46,289 127,436 0.363
Dermatology 49,947 90,870 0.550
Psychiatry 29,993 110,045 0.273
Cardiovascular Diseases 27,783 32,945 0.843
Otolaryngology 28,965 27,495 1.053
Urology 26,153 35,925 0.728
General Surgery 15,685 16,282 0.963
Neurology 14,407 29,811 0.483
All other specialties 120,875 96,173 1.257
Note: a https://www.cdc.gov/nchs/data/ahcd/namcs_summary/2016_namcs_web_tables.pdf
Source: IHS Markit © 2020 IHS Markit
Health workforce staffing patterns
Demand for healthcare workers is derived from the demand for healthcare services. The status quo scenario in
HDMM extrapolates current staffing levels as reflected by national healthcare use-to-provider ratios. For example,
demand for registered nurses (RNs) under the status quo is modeled based on the current national ratio of
inpatient days-to-RNs to model RNs in hospital inpatient settings, the national ratio of ED visits-to-RNs to model
demand for RNs in emergency departments, the national ratio of office visits-to-RNs to model demand for RNs in
office settings, etc.
The national number of health workers comes from many different sources, as described in the chapter describing
supply modeling, including associations’ Master Files (e.g., AMA Master File for physicians, ADA Master File
for dentists), the Health Resources and Services Administration’s (HRSA’s) National Sample Survey of
Registered Nurses for RNs and advanced practice registered nurses (APRNs), association publications such as
NCCPA reports for number of licensed physician assistants (PAs), and ACS and Occupational Employment
Statistics (OES) survey data collected from employers by the Bureau of Labor Statistics for select health
occupations.
The distribution of health workers across care delivery settings comes from multiple sources—including
published data collected by specialty associations via surveys of their members (e.g., NCCPA data on physician
assistants); specialty surveys (e.g., HRSA’s National Sample Survey of Registered Nurses); and OES data from
IHS Markit | HWMM Documentation
Confidential. © 2020 IHS Markit. All rights reserved. 20 May 2020
employer surveys reported by detailed health occupation, industry sector, and state. Limitations of OES data
include (1) it counts job positions, which may produce overcounting in occupations that have a high proportion of
part time workers, and (2) the data are for employed individuals, which can undercount the workforce in
occupations with a high proportion of self-employed individuals such as dentists or physicians.
For many occupations, demand is tied to one workload measure—e.g., demand for dentists is tied to demand for
dental visits (excluding dental cleaning visits), and demand for dental hygienists is tied to demand for dental
cleanings. For nurses, physicians, APRNs, PAs, and health occupations that work in multiple care delivery
settings there are multiple workload measures specific to each occupation and employment setting. The use of
multiple workload measures reflects that demand in each setting will grow at different rates.
In addition to using current staffing ratios to model a status quo scenario, HDMM was designed to model possible
changes in staffing patterns to reflect emerging care delivery models as informed by the literature. These scenarios
are discussed in more detail later and are also areas of ongoing research. Population health risk factors affect the
demand for healthcare services, but HDMM staffing currently does not account for variation across geographic
areas or over time in average patient acuity level for those who seek care. This is also an area of ongoing research.
Scenarios
The capabilities of HDMM to model alternative demand scenarios continue to evolve, and scenarios previously
modeled continue to be refined as new information becomes available. Many of these scenarios have been
described and the demand implications summarized in previous publications.25,44
• Status quo. This scenario models the implications of changing demographics as the population grows, ages,
and becomes more racially and ethnically diverse. Under this scenario healthcare use and delivery patterns are
modeled as remaining consistent with current patterns (i.e., observed during the 2013-2017 as reflected in the
MEPs and the 2017 NIS). Prevalence of disease and other health risk factors (e.g., smoking and obesity)
remain constant within each demographic group, but do change in the aggregate level as population
demographics change. For example, prevalence of diabetes and heart disease will rise as the population ages
but do not change independent of changing demographics. This scenario models the future demand for health
workers to provide a level of care consist with current levels.
• Increased medical insurance coverage. Earlier workforce studies modeled the implications of expanded
medical insurance coverage under the Affordable Care Act (ACA), but because recent patterns of healthcare
use and delivery largely have incorporated the effects of ACA this scenario is no longer modeled. However,
HDMM has been used to model hypothetical scenarios of insuring the uninsured to estimate the potential
impact of goals to improve access to care. This scenario assumes that a person who gains insurance will have
healthcare use patterns similar to his or her commercially insured counterpart with the same demographics and
risk factors. Although there may be an initial uptick in care sought, the scenario captures what happens when
the care sought by the newly insured settle into patterns of the currently insured. In HDMM this is essentially
done by switching the insurance status of a person from uninsured to insured and holding all other patient
characteristics constant.
• Reducing barriers to accessing care. This scenario builds on the increased medical insurance coverage
scenario to model the impact on health workforce demand if historically underserved populations had
improved access to care. Populations identified as underserved include minority populations and people living
in non-metropolitan areas—as well as people without medical insurance.45–48 When modeling this scenario for
oral health, lower household income is also identified as a barrier to receiving care (whereas for most other
healthcare services household income has only a small correlation with use of healthcare services controlling
IHS Markit | HWMM Documentation
Confidential. © 2020 IHS Markit. All rights reserved. 21 May 2020
for insurance status). In some studies this scenario has been referred to as a “health care utilization equity”
scenario.25
• Increased use of managed care principles. A variety of integrated care delivery models are being
implemented for both publicly and privately insured populations in an effort to both control rising medical
expenditures and improve delivery of care. Risk-bearing entities such as accountable care organizations
(ACOs) and Health Maintenance Organizations (HMOs) incorporate financial incentives for patients and
providers to better manage utilization by creating incentives for providers to collaborate in providing and
coordinating patient care across settings. ACOs have been promoted under ACA, but because they are a
relatively new care delivery model there is still limited data on their impact on patient use of services, how
care is delivered, and the demand implications for the health professions. Looking historically at the effect of
HMOs and other risk-bearing delivery models on use of services provides insights on what might happen if
ACOs gain greater prominence. One aspect of managed care is promotion of primary care and preventive care
to reduce need for expensive, hospital-based care and need for specialist care. One of the explanatory
variables in HDMM is the MEPS variable of whether the person is in an HMO-type managed care plan. By
changing people’s status from non-HMO to HMO, while holding all other characteristics constant, we model
the demand implications of increasing the proportion of the population in managed care plans. In general,
scenario findings are an increase in demand for primary care services and providers with a decrease in demand
for many types of specialist services and their providers.
• Expanded use of retail clinics. Retail clinics provide a convenient, cost-effective option for patients with
minor acute conditions. The number of retail clinics has grown rapidly over the past decade and is projected to
reach about 5,600 clinics by 2022.49–51 Retail clinics appear to be servicing demand for some types of services
historically provided in other settings, and also appear to be creating a net increase in healthcare utilization for
services provided to populations historically underserved and who would not otherwise receive care.52,53 For
example, an estimated 39% of visits to retail clinics replace physician visits, 3% replace emergency
department visits, and 58% are new visits that would not otherwise have occurred.52 This scenario explores
the demand implications of shifting care from primary care physician offices to retail clinics for 10 conditions
typically treated at retail clinics: upper respiratory infection, sinusitis, bronchitis, otitis media (middle ear
infection) and otitis externa (external ear infection), pharyngitis, conjunctivitis, urinary tract infection,
immunization, blood pressure check or lab test, and other preventive visit.51,53
In this scenario, patient visits to specialist physician are unaffected, and patients with modeled chronic
conditions in HDMM (i.e., cardiovascular, diabetes, asthma, hypertension or history of stroke) will continue
to be seen by their regular primary care provider even for non-complex health issues that could be treated in a
retail clinic. The scenario models a shift in demand from primary care physician offices to retail clinics,
incorporating into the workforce demand implications that 83% of visits to a pediatrician’s office are handled
primarily by a physician (reflecting that between NPs and physicians, 83% of the pediatric workforce are
physicians) and 71% of adult primary care office visits will be handled primarily by a physician. Care in retail
clinics is provided mostly by nurse practitioners.
• Increased use of APRNs and PAs. Studies conducted for the Association of American Medical Colleges
(AAMC) have modeled the implications on demand for physicians of the rapid growth in supply of APRNs
and PAs. This scenario, described elsewhere, uses different assumptions of the degree to which demand for
physicians might decrease as a result of growing supply of APRNs and PAs.25 The scenario assumes that a
portion of the increased supply of APRNs and PAs will replace some physician demand, a portion will expand
overall patient access to care but not replace physician demand, and a portion will increase the
comprehensiveness of care provided to patients but not replace physician demand. A 2012 study, for example,
estimated that patients receiving care from primary care physicians working alone received only 55% of
IHS Markit | HWMM Documentation
Confidential. © 2020 IHS Markit. All rights reserved. 22 May 2020
recommended chronic and preventive services.54 The gap between services recommended and services
provided was attributed to physicians being overworked with unmanageable patient panel sizes, but that much
of the care for recommended services could be delegated to other team members such as NPs and PAs, thus
delivering more comprehensive care while also reducing provider burnout.
• Achieving select population health goals: This scenario models the healthcare services and workforce
demand implications of achieving select population health goals. Consistent with Healthy People goals and
objectives, many programs and interventions target modifiable lifestyle behaviors and health risk factors
known to contribute to chronic disease—including efforts to reduce excess body weight, hypertension,
dyslipidemia, hyperglycemia, and smoking.55–58 This scenario is modeled using the Disease Prevention
Microsimulation Model (DPMM), discussed later, to simulate the healthcare demand implications of (a) a
modest 5% sustained reduction in excess body weight among adults who are overweight or obese; (b)
reductions in blood pressure, cholesterol, and blood glucose levels among adults with elevated levels with the
magnitude of reductions reflecting what can be achieved through appropriate medication and counseling as
reported in published clinical trials; and (c) 25% of smokers quit smoking—though with high recidivism.59–65
The mechanisms by which this hypothetical scenario could be achieved included increased use of medical
homes, value-based insurance design, and increased emphasis on preventive care to provide patients with
testing and counseling to improve patient adherence to treatment regimens.66–72 Research shows that people
who stop smoking can lower their risk for various cancers, diabetes, cardiovascular disease and other
morbidity and also reduce mortality.73–75
This scenario produces three impacts on provider demand:
(1) improved population health delays or prevents onset of adverse patient conditions thereby reducing
demand for some types of healthcare services and providers;
(2) shifts between different types of providers and care delivery settings—e.g., lower demand for
endocrinologists but higher demand for geriatricians, and shifts from hospital to ambulatory-based care; and
(3) people living to an older age due to met population health goals with chronic conditions will require more
healthcare services.
• Evolving care delivery system. While each of the above scenarios are modeled in isolation to quantify their
individual effect on demand, the healthcare system continues to evolve along multiple fronts—often with the
above scenarios overlapping. For example, to achieve the modeled population health goals requires more
comprehensive preventive services and improved access to care, as well as team-based care that involves
greater use of NPs, PAs, and other providers. Improved access to counseling and medications to achieve these
modeled population health goals is consistent with managed care principles, patient centered medical home
(PCMH) models of care delivery, and value-based insurance design (VBID). Managed care principles are also
consistent with interventions to divert costly hospital-based care to appropriate ambulatory settings, and
improve integration of care delivery. Policy initiatives try to advance national goals of increasing equity in
health outcomes and improving access to high quality, affordable care. This scenario, therefore, combines
several of the above scenarios with attention paid to not double counting the effects that overlapping scenarios
might have on demand for healthcare services and providers.
Input summary
HDMM uses data from a variety of public data sources, which are summarized in Exhibit 12. The model
undergoes a major update in November of each year—reflecting that many of the government sponsored annual
surveys and data sources used in the model are often released to the public approximately July – October each
year.
IHS Markit | HWMM Documentation
Confidential. © 2020 IHS Markit. All rights reserved. 23 May 2020
Exhibit 12 Demand model input data summary
Demand model input data summary
Data Source Use Latest Data
Used Last Updated
Population File
American Community Survey (ACS) Create representative sample of population in each state (demographics, insurance, household income)
2018 November 2019
U.S. Census Bureau Population estimates by County
Create representative sample of population in each county (demographics)
2018 November 2019
Behavioral Risk Factor Surveillance System (BRFSS)
Create representative sample of community-based population in each state (heath risk factors and disease prevalence)
2018 November 2019
CMS Nursing Home Minimum Data Set (MDS)
Create representative sample of nursing home population in each state (heath risk factors and disease prevalence)
2017 November 2019
CMS Medicare Beneficiary Survey (MCBS)
Create representative sample of population in residential care facilities in each state (heath risk factors and disease prevalence)
2017 November 2019
U.S. Census Population Projections National population projections 2016 November 2017
State Population Projections Individual state population projections Various November 2019
County Population Projections Individual county population projections and IHS Markit’s regional forecast
Various November 2019
Healthcare Use
Medical Expenditure Panel Survey (MEPS)
Estimate health seeking behavior by care delivery setting and provider type
2017 November 2019
National Inpatient Sample (NIS) Estimate hospital length of stay; model calibration for annual hospital visits
2017 November 2019
National Ambulatory Medical Care Survey (NAMCS)
Model use of non-physician services during office visits; model calibration for annual office visits
2016 November 2019
National Hospital Ambulatory Medical Care Survey (NHAMCS)
Model use of non-physician services and physician consults during ED visits; model calibration for annual ED visits
2016 November 2019
Healthcare Provider Staffing
Bureau of Labor Statistics, Occupational Employment Statistics
Estimate provider staffing ratios by health occupation and delivery setting (excluding physicians)
2018 November 2019
Individual profession association surveys
Estimating staffing across care delivery settings Various November 2019
Source: IHS Markit © 2020 IHS Markit
Health Workforce Supply Model HWSM is designed to project future supply of health professionals under alternative forecasting scenarios using a
microsimulation approach. Supply projections account for characteristics of the current and projected future
workforce and other external factors (e.g., training capacity, demand for services) that might affect career choices
of health professionals. Below, we describe the logic, data, methods, and assumptions for modeling health
workforce supply, as well as the major components of the model and the scenarios that can be modeled.
Starting supply input files
The microsimulation model projects future supply by simulating likely workforce decisions of individual, de-
identified healthcare providers. This approach requires developing a starting supply file of all providers
(preferred approach) or a representative sample of providers from survey data. When modeling supply for
individual states and at the sub-state level the primary data source of de-identified, individual-level provider data
is state licensure files. These files typically contain the providers’ occupation/specialty, active/inactive status,
geographic area where working, and demographics. Age is the most important demographic information used to
IHS Markit | HWMM Documentation
Confidential. © 2020 IHS Markit. All rights reserved. 24 May 2020
model workforce decisions as hours worked patterns and retirement probabilities are highly correlated with age.
Workforce decisions (especially hours worked patterns) also vary systematically by sex. Race/ethnicity is added
for some occupations based on availability, but it is a less significant predictor of workforce decisions than age
and sex. State licensure files sometimes contain information collected via survey at time of re-licensure—such as
weekly patient care hours worked, employment setting, and retirement intentions (as discussed later).
Other data sources that have been used to develop a file for starting supply when state licensure data is
unavailable include national surveys, national certification data, and association membership and registration
databases:
• National databases (licensure, membership, or registration)
o American Medical Association (AMA) Master File: continuously updated with a record for each
physician who has been licensed in the U.S.
o American Dental Association (ADA) Master File: continuously updated with a record for each dentist
who has been licensed in the U.S.
o American Academy of Physician Assistants (AAPA) Master File: includes a record for physician
assistants by specialty, we combine this with NCCPA publications on total number of PAs
o Membership files created by individual professional associations
o National Plan and Provider Enumeration System (NPPES), continuously updated to provide a unique
identifier for providers who bill CMS for services
• Surveys
o American Community Survey (ACS), updated annually by the U.S. Census Bureau, contains a stratified
random sample of the population in each state and lists occupation and employment status
o Occupational Employment Statistics (OES), updated by the U.S. Bureau of Labor Statistics, collects data
on employed individuals via an employer-based survey
o Occupation/specialty surveys
▪ HRSA National Sample Survey of Registered Nurses (NSSRN), last updated in 2018, includes an
oversample of APRNs
Each of the data sources contains different types of data and comprehensiveness of health workers—ranging from
licensure files that contain a complete census of providers in the geographic area of interest, to association
membership files that contain data on members and limited data on non-members, to employer or population-
based surveys that use sample weights to scale to the population in the geographic area. State licensure files are
usually the most accurate source of data to create the starting supply files, and some of the above data sources are
derived from and updated periodically using state licensure data.
New entrants
When modeling at the national level the new entrants are those individuals entering the workforce after
completing appropriate training and licensure. When modeling at the state or sub-state level the new entrants
reflect both those individuals newly entering the workforce for the first time, as well as individuals who might be
migrating mid-career from one geographic area to another.
Each year new entrants are added to the supply file via creation of a “synthetic” population based on the number
and characteristics of new entrants to the workforce. For example, if 100 new providers in a given occupation or
specialty entered the workforce in a particular year then the model creates 100 new records—one for each person.
The age and sex of each new person is generated based on the estimated distribution from recent entrants to the
workforce. If, for example, 90% of new entrants to the RN workforce were female then the model generates a
random number for each new person using a uniform (0, 1) distribution. The person is designated as male if the
random number for that person is less than or equal to 0.1, and otherwise designated as female. A similar process
IHS Markit | HWMM Documentation
Confidential. © 2020 IHS Markit. All rights reserved. 25 May 2020
is used to designate each new person’s age and race/ethnicity (for those occupations were this dimension has been
added to the supply model).
For state-level analyses, licensure files are the most useful source of information on the number and
characteristics of providers entering the workforce. Analyzing several years’ data helps provide a sufficient
sample size to estimate the annual number and demographics of new entrants. In addition to state licensure files,
additional national data sources for information on the number and characteristics of newly trained health
providers entering the workforce are listed in Exhibit 13.
Data limitations regarding new entrants presents challenges for modeling future supply of some health
occupations. This includes some aide/assistant/paraprofessional occupations where new entrants might enter the
workforce through formal or on-the-job training, or where there is no formal licensure process.
Exhibit 13 Data sources for number and characteristics of new entrants
Data sources for number and characteristics of new entrants
Profession Number and Characteristics of New Entrants
All licensed professions State licensure files (where available)
Registered nurses NCLEX; National League for Nursing, http://www.nln.org/researchgrants/slides/topic_nursing_stud_demographics.htm
Licensed practical nurses National Council Licensure Examination (NCLEX)
Dentists American Dental Association Master File
Dental hygienists Integrated Postsecondary Education Data System (IPEDS)
Physicians American Medical Association (publications76 and Master File)
Advanced practice nurses American Association of Colleges of Nursing (AACN)
Physician assistants National Commission on Certification of Physician Assistants; Physician Assistant Education Association
Therapeutic service providers IPEDS
Rehabilitation service providers IPEDS
Respiratory care providers IPEDS
Vision and hearing care providers IPEDS
Dietitians & nutritionists IPEDS
Pharmacy professions IPEDS
Non-physician behavioral health providers IPEDS
Diagnostic laboratory providers IPEDS
Source: IHS Markit © 2020 IHS Markit
Hours worked patterns
The model simulates weekly hours worked for each health worker and captures changing demographics of the
workforce over time. Hours worked patterns vary by occupation/specialty, provider age and sex, and for some
occupations by economic conditions and geographic location. Hours worked is converted to FTE levels by
dividing the hours worked for each provider by current average hours worked in the profession. Patterns of hours
worked were calculated differently by occupation based on data availability. Where possible, we used regression
analysis (with Ordinary Least Squares regression) to estimate the effect of workforce determinants on weekly
hours worked.
IHS Markit | HWMM Documentation
Confidential. © 2020 IHS Markit. All rights reserved. 26 May 2020
Physicians
For physicians, the prediction equation for weekly hours worked in initial workforce studies was based on OLS
regression using Florida’s 2012-2013 bi-annual Physician Licensure Workforce Survey (n=17,782), restricted to
physicians who reported working at least 8 hours per week in professional activities, where patient care hours was
the dependent variable and explanatory variables consisted of the following class variables: specialty, age group,
sex, and age-by-sex interaction terms. Subsequently, we added similar data from South Carolina (2014, n= 8,924),
New York (2014, n = 44,181), and Maryland (2016-2017, n=24,668) for a total of 95,555 total physician
responses. This increase in sample size allowed for OLS regressions by individual specialty.
Exhibit 14 illustrates regression results and presents information for general internal medicine. Starting with 46.29
hours/week for the reference group (male, age<35, working in Florida), we find that females work 3.53 fewer
patient care hours/week, on average, relative to their male colleagues with older females working even fewer
hours than their male colleagues; hours/week drops slightly around age 60 and then continues to decline more
rapidly for older physicians; and patient care hours/week are lower in New York and Maryland as compared to
Florida and South Carolina.
Exhibit 14 OLS regression example: weekly patient care hours for general internal medicine
OLS regression example: weekly patient care hours for general internal medicine
Parameter Patient hours
Intercept 46.29
Age 35 to 44 0.52
Age 45 to 54 0.30
Age 55 to 59 0.17
Age 60 to 64 -2.04 **
Age 65 to 69 -5.65 **
Age 70-74 -8.37 **
Age 75+ -14.91 **
Female -3.53 **
Age 25 to 34 Female 0.75
Age 35 to 44 Female -2.20 **
Age 45 to 54 Female -2.03 **
Maryland -4.01 **
New York -7.13 **
South Carolina 0.50
Note: Notes: Statistically significant at the 1% (**) or 5% (*) level. Comparison groups are age <35, male, Florida. R2=0.09.
Source: IHS Markit © 2020 IHS Markit
Our goal is to use a nationally representative data source to create the OLS regressions; to that end, IHS Markit
teamed with AAMC to estimate physician hours worked patterns based on the AAMC National Sample Survey of
Physicians (2019) for the 2020 update of AAMC’s physician workforce projections report. Where we have
conducted workforce studies for individual specialties or other professional occupations, we use survey data
collected by the associations sponsoring such studies (e.g., American Academy of Neurology [2016 AAN Career
Satisfaction Survey, n=910], American Academy of Pediatric Dentistry [2017 AAPD Survey of Pediatric
Dentists, n=2,546], Association of Academic Physiatrists [2019 AAP Survey of Physiatrists, n=616]).
IHS Markit | HWMM Documentation
Confidential. © 2020 IHS Markit. All rights reserved. 27 May 2020
Other health occupations
The hours worked regressions for other health occupations modeled analyze ACS data for employed clinicians
who reported at least 8 hours worked per week. Dependent variables include clinician characteristics such as age,
sex, and race. We also include the year the clinician responded to the ACS to control for changes in hours worked
over time, but that variable is not included in the supply simulation. Multiple years of ACS data were combined to
increase sample size.
Exhibit 15 summarizes regression output for select occupations using ACS data from 2013-2017. For all
occupations, weekly hours worked decline rapidly from age 65 onward. Using RNs as an example, we find that on
average, male RNs work 2.85 more hours than their female counterparts, Hispanic RNs work 1.25 hours more
than non-Hispanic white RNs, and RNs in 2017 worked 0.39 more hours than in 2013. The low R2 values for
occupation-specific regressions suggest that demographics alone explain only a small portion of the variation
across providers in weekly hours worked. Other factors that might explain variation in weekly hours worked are
household characteristics (e.g., number and age of children, health status, marital status and earnings potential of
spouse or significant other). However, these variables are unavailable to be included in the microsimulation model
for supply. Still, at the occupation level the regression results do find statistically significant and substantial
variation in weekly hours worked that can be explained by provider demographics
Exhibit 15 OLS regression coefficients predicting weekly hours worked for select occupations
OLS regression coefficients predicting weekly hours worked for select occupations
Parameter RN LPN Dental hygienist Physical therapist Pharmacist
Intercept 39.30 ** 39.05 ** 38.08 ** 43.39 ** 45.67 **
Female -2.85 ** -2.13 ** -6.31 ** -5.97 ** -5.28 **
Hispanic 1.25 ** 0.85 * 0.77 -1.05 1.87
Non-Hispanic black 2.00 ** 0.49 ** 4.43 ** 2.32 ** -1.18
Non-Hispanic other 1.12 ** 0.17 1.51 ** 0.74 ** -0.25
Age 35 to 44 0.32 ** 0.84 ** -1.37 ** -2.59 ** -1.07
Age 45 to 54 1.35 ** 1.09 ** -1.40 ** -1.80 ** -0.22
Age 55 to 59 1.14 ** 0.77 ** -2.37 ** -1.07 ** -1.09
Age 60 to 64 0.30 ** -0.18 -2.82 ** -2.20 ** -2.00 *
Age 65 to 69 -3.20 ** -3.93 ** -5.12 ** -4.44 ** -6.81 **
Age 70+ -7.10 ** -7.43 ** -10.49 ** -11.93 ** -10.29 **
Year 2014 0.07 -0.03 0.59 0.35 -0.46
Year 2015 0.30 ** 0.35 * 0.21 0.61 * 0.07
Year 2016 0.40 ** 0.56 ** 0.75 * 0.45 -0.33
Year 2017 0.39 ** 0.72 ** 0.92 ** 0.53 -0.42
Sample size 171,635 44,641 9,752 12,573 16,035
R-squared 0.03 0.02 0.05 0.09 0.05
Notes: Analysis of American Community Survey. Statistically significant at the 0.01 (**) or 0.05 (*) level. Comparison groups are age <35, male, non-Hispanic white, ACS year=2013.
Source: IHS Markit © 2020 IHS Markit
One limitation of ACS as a data source for modeling hours worked is that ACS does not collect data on specialty
area or clinical setting. Therefore, to model hours worked for physicians by individual specialty category we used
data collected by states or professional associations. Similarly, for ongoing work with HRSA the prediction
equations for primary care PAs is based on data from the 2019 AAPA Salary Survey. Prediction equations for
RNs by delivery setting and for APRNs by specialty are based on HRSA’s National Sample Survey of Registered
Nurses (2018).
IHS Markit | HWMM Documentation
Confidential. © 2020 IHS Markit. All rights reserved. 28 May 2020
Labor force participation
Labor force participation decisions encompasses whether a provider joins the workforce as well as his/her level of
participation. Providers might temporarily leave the labor force due to familial, educational, economic or other
considerations. Labor force participation status for some occupations is modeled using prediction equations
derived from ACS data, with multiple years of data combined to increase sample size. This analysis focuses on
clinicians under age 50 (as the HWSM switches to permanent retirement as the activity status changes for
clinicians age 50 and over). The dependent variable was whether the nurse was employed or not employed) with
explanatory variables listed in Exhibit 16. The estimated odds of being employed for three occupations by
clinician demographics—in particular age. Registered nurses and licensed practical nurses are less likely to be
active in the 30-34 and 35-39 age groups compared to RNs under age 30, but pharmacists are much more likely to
be active in their thirties and forties. Female RNs, LPNs, and pharmacists are all less likely to be active than males
in their occupation. Race has an effect on labor force participation that differs by the three occupations illustrated
here.
Exhibit 16 Odds ratios predicting probability active
Odds ratios predicting probability active
Parameter RN (n=90,696)
Odds ratio and CI LPN (n=22,836)
Odds ratio and CI Pharmacist (n=9,773)
Odds ratio and CI
Female b 0.77 0.69 0.86 0.66 0.55 0.78 0.62 0.49 0.79
Non-Hispanic black b 1.45 1.27 1.65 1.72 1.50 1.97 1.00 0.62 1.61
Non-Hispanic other b 1.22 1.10 1.34 1.01 0.87 1.17 0.84 0.66 1.06
Hispanic b 0.90 0.66 1.23 0.68 0.49 0.94 0.56 0.24 1.33
Age 30-34 0.81 0.73 0.89 0.93 0.80 1.08 1.37 1.04 1.82
Age 35-39 0.91 0.83 1.01 0.97 0.84 1.13 1.49 1.10 2.03
Age 40 to 44 1.08 0.98 1.20 1.02 0.87 1.18 1.66 1.20 2.29
Age 45 to 49 1.15 1.04 1.28 0.97 0.84 1.13 1.71 1.22 2.40
Year 2014 b 1.01 0.92 1.12 0.89 0.76 1.03 1.07 0.75 1.53
Year 2015 b 1.08 0.98 1.19 0.89 0.76 1.04 0.94 0.67 1.33
Year 2016 b 1.07 0.97 1.18 0.94 0.80 1.10 0.78 0.57 1.08
Year 2017 b 1.09 0.99 1.20 0.97 0.83 1.14 0.72 0.52 1.00
Notes: Odds ratios and 95% confidence interval (CI) from logistic regression. Comparison groups are male, non-Hispanic white, age <30, and ACS year=2013. Labor force participation regressions are based only on clinicians under age 50.
Source: IHS Markit © 2020 IHS Markit
Retirement
In addition to temporary departures from the workforce, clinicians can also leave permanently. The approach to
modeling retirement differs by occupation depending on data availability. The supply model assigns each person
an attrition probability based on age, sex, and occupation/specialty. However, surveys asking about retirement
intention are rare and often have few retirements, so in many cases the responses from male and female clinicians
are combined before creating the retirement pattern. Alternatively, some specialties or occupations may be
combined into groups such as all primary care physicians or all counselors and therapists. Once calculated, this
probability is then compared with a random number between 0 and 1 (using a uniform distribution) generated for
each observation in the supply input file to simulate whether the person leaves the workforce each year. For
example, if an active clinician age 66 has a 20% probability of retiring by age 67, then if the random number is
below 0.2 the person will be removed from the simulated workforce. Otherwise, that person is considered still
active at age 67 and the simulation will consider them for retirement at age 68 during the next iteration. Each
occupation has a maximum age at which the retirement probability is set to 1 and all providers are removed upon
reaching that age. This is necessary because the retirement patterns are calculated based on age and at a
IHS Markit | HWMM Documentation
Confidential. © 2020 IHS Markit. All rights reserved. 29 May 2020
sufficiently advanced age there is not enough data to predict retirement at that age. In general, this maximum age
is 75 for most occupations and 90 for physicians and dentists.
Physician attrition patterns
Historically there has been a paucity of information on physician retirement patterns. Few surveys collect
information on retirement intentions or retirement age, and state licensure files often have insufficient sample
sizes for older physicians in individual specialties to adequately estimate retirement patterns. National surveys like
ACS do not indicate physician specialty.
Many of the retirement rates for individual specialties used in HWSM were estimated using survey data from the
Florida bi-annual physician survey (2012-2013 data) that asks about intention to retire in the upcoming five years.
Derived retirement patterns from this survey are similar to estimates derived from analysis of the AAMC’s 2006
Survey of Physicians over Age 50 (which collected information on actual retirement age of retired physicians, or
age those physicians still active were expecting to retire). For select physician specialties and other health
professionals the retirement patterns were estimated using survey data collected by the sponsoring associations
(e.g., American Academy of Neurology [2016 AAN Career Satisfaction Survey, n=910], American Academy of
Pediatric Dentistry [2017 AAPD Survey of Pediatric Dentists, n=2,546], and Association of Academic
Physiatrists [2019 AAP Survey of Physiatrists, n=616]).
IHS Markit has worked with AAMC to create physician retirement patterns based on the AAMC National Sample
Survey of Physicians (2019), and those retirement rates are used in HWSM to produce the projections in AAMC’s
2020 report on the physician workforce.
Based on data from the Florida survey, female physicians intend to retire slightly earlier than their male
colleagues (Exhibit 17). Overall, among 100 physicians active in the workforce at age 50, by age 60
approximately 80 will still be active. By age 70 approximately 30 will still be active. When taking into
consideration that average hours worked declines with age, the number of FTE physicians above age 70 is lower
than indicated by retirement patterns alone.
IHS Markit | HWMM Documentation
Confidential. © 2020 IHS Markit. All rights reserved. 30 May 2020
Exhibit 17 Physician retirement patterns by age and sex
Exhibit 18 shows estimated overall attrition patterns for male physicians by specialty, with some specialties such
as emergency medicine experiencing earlier attrition relative to other specialties. For example, by age 65
approximately 65% of allergists & immunologists are still active, while only 50% of emergency physicians are
still active.
IHS Markit | HWMM Documentation
Confidential. © 2020 IHS Markit. All rights reserved. 31 May 2020
Exhibit 18 Probability male physician is still active by specialty and age
Nurse retirement patterns
HRSA’s 2018 National Sample Survey of Registered Nurses (released in late 2019), which collected information
on age respondents intended to retire, now is used to model retirement patterns for RNs and for APRNs by
specialty.
Different approaches were explored and used to estimate nurse retirement patterns in prior studies of the nurse
workforce. One approach used ACS data and state licensure data to estimate attrition by comparing the number of
nurses in each age cohort across years. For example, the number of active nurses age 60 in a particular year (Y)
are compared to those still active at age 61 in the subsequent year (Y+1) to estimate retirement during age 60. The
advantages of this approach are that a cohort comparison estimates net attrition with some people leaving the
workforce and others re-entering. Disadvantages of this approach are (a) data sources such as ACS survey
different groups of people each year so the number of people in a particular age group and occupation might differ
from year-to-year due to sampling issues, and (b) in both national surveys and state licensure files the number of
people of a specific age and occupation might be small—especially when sub-setting the data to estimate separate
IHS Markit | HWMM Documentation
Confidential. © 2020 IHS Markit. All rights reserved. 32 May 2020
retirement patterns for men versus women. Still, estimated retirement patterns for RNs and for LPNs were similar
using this approach with ACS and state licensure data.
Retirement patterns of other health providers
For other health occupations, HWSM uses retirement patterns estimated from the most recent ACS 5-year public
use microdata sample (PUMS) data. Using the 5-year file instead of the 1-year file allows for a larger sample size
of recent retirees. The ACS includes questions that asks respondents both if they are currently working and if they
were working one year previously. The retirement pattern is created using the assumption that respondents who
are not working currently, were working one year ago, and are 50 or more years old have retired in the last year.
Thus, these retirement patterns are created from observed retirements instead of survey responses about intention
to retire.
Geographic migration
Migration patterns of clinicians across states is an ongoing area of research for HWSM. Cross-state migration can
happen at the start of one’s career upon completion of training or can occur mid-career. The probability of cross-
state migration and the factors influencing such migration vary by occupation, specialty and by state. Higher-
paying occupations like physicians are more likely to be in a national labor market relative to lower-paying health
occupations (from which recruiters might look locally). However, providers in occupations with high rates of self-
employment (e.g., dentists or physicians) are probably less likely to move mid-career--after establishing a
practice—relative to those in occupations with higher proportions of employees who tend to be more mobile.
One scenario models that areas of the country experiencing faster growth in demand for healthcare services will
also experience faster growth in provider supply relative to areas of the country experiencing slower growth in
demand for services. That is, health workers will migrate to those geographic areas where there is greater demand
for their services. This approach has been applied when modeling demand for physicians, dentists, and RNs. The
approach produces the following for the occupation or medical specialty of interest:
1. Projected growth in demand in each state over the forecast time horizon;
2. Projected retirements in each state over the same time horizon;
3. Add projected growth and projected retirements to estimate total new workers required to meet future
demand for services;
4. Sums of total new requirements across states and estimates of each state’s share of total requirements; and
5. How new workers will be distributed across states using this distribution of requirements as a proxy.
Each new entrant to the workforce is assigned a state using this calculated distribution under the assumption that
new graduates will migrate to those geographic locations where growth in demand or retirements creates
opportunities for employment (but allowing current mal-distribution of health professionals to persist). For
example, faster growing states are anticipated to attract a growing proportion of the nation’s new health
professionals while slower growing states are likely to attract a smaller proportion than historical patterns. This
topic is an area for continued research.
Migration patterns for select occupations for which workforce studies have been conducted have used de-
identified records from association membership files to assess change-in-address information for modeling cross-
state migration patterns.
IHS Markit | HWMM Documentation
Confidential. © 2020 IHS Markit. All rights reserved. 33 May 2020
Scenarios
HWSM is designed to model a status quo scenario as well as alternative scenarios based on changes in supply
drivers—namely, number of new entrants to the workforce, changes in labor force participation or hours worked
patterns, and changes in retirement patterns.
• Status quo. This scenario uses the most recent data on number and characteristics of health workers being
trained and entering the workforce, and current data on labor force participation, hours worked, and retirement
patterns.
• Change in number of new graduates. The status quo supply projections model the current annual number of
workers trained, or in the case of occupations with rapid growth model the increase in training capacity as
announced new programs start graduating new workers. High growth scenarios might model, for example, the
implications of training 10% more providers. Low growth scenarios might model the implications of training fewer
providers. In situations where specific data on future changes to the training pipeline is available, a scenario can be
created that compares the potential future change to the status quo. For example, HWSM has been used to evaluate
the workforce implications of proposed state or national legislation to modify the number of physicians being
trained.
• Delayed and Early Retirement. A common concern in the field of health workforce analysis is that retirement
patterns can evolve over time. In particular is concern that rising levels of provider burnout could contribute to
earlier retirement.77–83 Scenarios simulating a one- or two-year shift in retirement patterns can make it easier to
understand the effect this may have on the overall supply of a health profession.
• Hours Worked Cohort Effects. Work-life balance expectations and hours work patterns for health workers
newly entering the workforce could be systematically different from current patterns. That is, in 10 years the
typical physician currently 30-year old might not work the same number of hours as a typical physician currently
40 years old. Analysis of ACS data indicates that average hours worked by physicians has declined over the past
two decades across all age groups, though in recent years the hours worked patterns appear to be stabilizing.25
Likewise, declines in the number of clinicians who are self-employed and changes in reimbursement schemes could
contribute to physicians working fewer hours.84 A scenario which models shifts in hours worked patterns explores
the potential effects of declining hours worked.
Workforce implications of strategies to prevent or manage chronic disease The Disease Prevention Microsimulation Model (DPMM) is designed to model the health and economic
implications of interventions to improve population health. Population health management plays an important role
in modeling future demand for healthcare services and providers—using lifestyle indicators and health-related
behavioral variables regarding smoking, diet, physical activity, and other activities (e.g., preventative screenings,
vaccinations, and early treatment) linked to patient health. Improved lifestyle choices and other preventative care
can help prevent, delay onset, or reduce severity of many chronic conditions such as asthma, diabetes, heart
disease, and cancer.
DPMM has been used to model the implications of lifestyle counseling among overweight and obese adults with
risk factors for cardiovascular disease and diabetes; improved control of blood pressure, cholesterol, and blood
glucose levels through medication; tobacco cessation; and screening and early treatment for select preventable
conditions.25,35–39 Detailed documentation of DPMM is available elsewhere.1
An interdependent relationship exists between the health workforce and prevention efforts to improve health.
IHS Markit | HWMM Documentation
Confidential. © 2020 IHS Markit. All rights reserved. 34 May 2020
• Many prevention interventions are provided by health workers (e.g., screening, counseling, and vaccinations) thus
increasing demand for the occupations that provide such services.
• Reducing prevalence or severity of chronic conditions and adverse medical events through prevention reduces
demand for clinicians who provide those services (and can shift demand to lower-acuity care delivery settings).
• Preventing or delaying onset of chronic disease can increase life expectancy thus increasing patient use of other
healthcare services over the lifespan.
DPMM uses a Markov Chain Monte Carlo simulation approach to model likelihood and timing of disease onset
for each person in a representative sample of the population of interest. Earlier we described creation of a
representative sample of the population in each U.S. county for modeling with HDMM. The population file for
simulation using DPMM requires additional information not needed for HDMM modeling which we obtained
from the National Health and Nutrition Examination Survey (NHANES). In addition to the variables used in the
HDMM population files, DPMM requires body mass index, systolic blood pressure, cholesterol levels, blood
glucose levels, and the presence of other diseases. We constructed a nationally representative sample of 50,000
adults combining multiple years of NHANES data for running the DPMM simulations. Outcomes from DPMM
where then extrapolated to the population employed in HDMM using propensity matching for the demographics,
health behaviors, socioeconomic characteristics, and disease presence information available in both the HDMM
and DPMM population files.
Exhibit 19 provides an overview of DPMM. In a particular year (y), a person’s health risk factors and biometric
readings can affect how biometric levels change over the year as the person ages (to year y+1). Changing
biometrics (as well as the other risk factors) are linked to the probability of various health states (e.g., onset of
diabetes or heart disease). The health states are also linked to each other—e.g., diabetes is an independent risk
factor for heart disease in addition to sharing common risk factors such as obesity and smoking. The presence and
severity of chronic disease affect patient mortality, medical expenses, and other economic and quality of life
outcomes.
IHS Markit | HWMM Documentation
Confidential. © 2020 IHS Markit. All rights reserved. 35 May 2020
Exhibit 19 Overview diagram of the Disease Prevention Microsimulation Model
Similarly, Exhibit 20 illustrates how a biometric variable like BMI is linked to various cancers and endocrine,
cardiovascular, respiratory, and other medical conditions. Many of these medical conditions have independent
effects on disease onset risk for other medical conditions. Each arrow represents a prediction equation in DPMM
that connects disease states.
Risk FactorsBiometrics and Health Inputs
Health States Outcomes
Demographics
Disease history
Biometrics
Smoking
Alcohol misuse
BMI
A1c
SBP
DBP
HDL-C
Total Cholesterol
LVH
Atrial fibrillation
Diabetes & Sequelae
Cardiovascular
Cancers• Breast• Cervical• Colorectal• Endometrial• Esophageal• Gallbladder• Kidney• Leukemia
• Liver• Ovarian• Pancreatic• Prostate• Stomach• Thyroid• Lung• Non-Hodgkin's
Mental & Cognitive
• Diabetes• Prediabetes
• Amputation• Retinopathy
• CHF• IHD• Hypertension
• Dyslipidemia• Stroke• MI
• Depression• Alzheimer’s
• Bipolar • Schizophrenia
Pulmonary• Pneumonia• Asthma• COPD
• Pulmonary embolism
Others• Osteoporosis• Chronic back pain
• GERD• NAFLD
Year y
Year y+1
Mortality
Employment
Absenteeism
Social Security Cost
QALY
Medical Expenses
Personal Income
Long Term Care
Note: Connecting lines show the items in the model that are linked
Abbreviations: BMI=body mass index, CHF=congestive heart failure, CKD=chronic kidney disease, DBP=diastolic blood pressure, HbA1c=hemoglobin A1c, HDL=high-density lipoprotein, IHD=ischemic heart disease, LVH=left ventricular hypertrophy, PVD=peripheral vascular disease, SBP=systolic blood pressure.
Source: IHS Markit © 2020 IHS Markit
Overview diagram of the Disease Prevention Microsimulation Model
IHS Markit | HWMM Documentation
Confidential. © 2020 IHS Markit. All rights reserved. 36 May 2020
Exhibit 20 Overview diagram of body weight component in DPMM
The patient-level output from DPMM can then be run through HDMM to simulate how the presence of chronic
conditions affects patient use of healthcare services and demand for providers.
Model validation, strengths, and limitations The models described in this report continue to be updated and refined to incorporate new data, changes in trends
or care delivery, and improvements to the underlying methods and assumptions. In this chapter we describe model
validation activities, and strengths and limitations of these models.
Endocrine
Diabetes (HbA1c)
Prediabetes (HbA1c)
Cardiovascular
LVH
Hypertension (SBP, DBP)
Dyslipidemia (HDL, Total cholesterol)
IHD
CHF
Direct Effect Disease States Indirect Effect Disease States
Atrial fibrillation
Amputation
PVD
Renal failure
CKD
Stroke
Myocardial infarction
Blindness
Body weight(BMI)
Respiratory
PneumoniaPulmonary embolism
Other
Chronic back pain
Osteoarthritis
Gallstones & gallbladder
GERDMajor depression
NAFLDOSA
CancersBreast
Cervical
Endometrial
Esophageal
Gallbladder
Kidney
Leukemia
Liver
NHL
Multiple Myeloma
Ovarian
Pancreatic
Prostate
Stomach
Thyroid
Colorectal
Note: Connecting lines show the items in the model that are linked
Abbreviations: BMI=body mass index, CHF=congestive heart failure, CKD=chronic kidney disease, DBP=diastolic blood pressure, GERD= gastroesophageal reflux disease, HbA1c=hemoglobin A1c, HDL=high-density lipoprotein, IHD=ischemic heart disease, LVH=left ventricular hypertrophy, NAFLD=non-alcoholic fatty liver disease, OSA=obstructive sleep apnea, PVD=peripheral vascular disease, SBP=systolic blood pressure.
Overview diagram of body weight component in DPMM
Source: IHS Markit © 2020 IHS Markit
IHS Markit | HWMM Documentation
Confidential. © 2020 IHS Markit. All rights reserved. 37 May 2020
Validation activities
Validation activities continue on an ongoing basis during model development and refinement as a long-term
process of evaluating the accuracy of the model and making refinements as needed. For each of four primary
types of validation deployed, key short term and long-term activities include the following:
• Conceptual validation: Through reports, presentations at professional conferences and submission of peer-
reviewed manuscripts the three models described here (HDMM, HWSM, and DPMM) continue to undergo a peer-
review evaluation of their theoretical framework. Contributors to these models include health economists,
statisticians and others with substantial modeling experience; physicians, nurses, behavioral health providers and
other clinicians; health policy experts; and professionals in management positions within health systems.
Conceptual validation requires transparency of the data and methods to allow health workforce researchers and
modelers to critique the model. This technical documentation is an attempt to increase the transparency of these
complex workforce projection models and facilitate improvements to the theoretical underpinnings, methods,
assumptions, and other model inputs. Additional technical documentation for various health occupations is
published elsewhere.24,85,86
• Internal validation: The models run using R, which is open source software. As new capabilities are added to the
models and data sources updated, substantial effort is made to ensure the integrity of the programming code.
Internal validation activities include: (1) generating results for comparison to published statistics used to generate
the models to ensure that population statistics for the input files are consistent with published statistics, (2)
checking for consistency with earlier versions of the models, and (3) stress-testing the models for comparison
against a priori expectations.
• External validation: Presenting findings to subject matter experts for their critique is one approach to externally
validate the model. Intermediate outputs from the model also can be validated. For example, HDMM has been used
to project demand for healthcare services for comparison to external sources not used to generate model inputs.
Results of such comparisons across geographic areas indicate that more geographic variation in use of healthcare
services occurs than is reflected in geographic variation in demographics, presence of chronic disease, and health
risk factors such as obesity and smoking.
• Data validation: Extensive analyses and quality review have been conducted to ensure data accuracy as model data
inputs were prepared. Most of the model inputs come from publicly available sources (e.g., MEPS, BRFSS, ACS,
or published studies)—with the exception that licensure data used in the model is often proprietary to each state
licensure board and purchased data from the American Medical Association and other groups has sometimes been
used for certain studies.
Model strengths
The main strengths of the three models are (1) use of recent data sources and efforts to continuously update and
maintain the models; (2) use of a sophisticated microsimulation approach that has substantial flexibility for
modeling changes in care use and delivery by individuals or by the healthcare system, as well as flexibility to
model a wide range of scenarios; and (3) its development and use across a broad range of different health
occupations, specialties, and stakeholder groups.
Compared to population-based modeling approaches used historically, these microsimulation models incorporate
more detailed information on population characteristics and health risk factors when making demand projections
across geographic areas and over time. For example, rates of disease prevalence and health related risk factors and
household income vary substantially by geographic area. Such additional population data can provide more
IHS Markit | HWMM Documentation
Confidential. © 2020 IHS Markit. All rights reserved. 38 May 2020
precise estimates of service demand at State and county levels compared to models that assume all people within a
demographic group have the same risk factors or use the same level of services.
HDMM simulates care use patterns by delivery setting. Certain populations have disproportionately high use of
specific care delivery settings (e.g., emergency care) and lower use of other settings. Setting-specific information
on patient characteristics and use rates provides insights for informing policies that influence the way care is
delivered. Because the microsimulation approach uses individuals as the unit of analysis, HDMM can simulate
demand for healthcare services and providers of care for sub-populations of particular interest such as low income
populations, populations in select underserved areas, or populations with certain chronic conditions. Additionally,
using individuals as the unit of analysis creates flexibility for incorporating evidence-based research on the
implications of changes in technology and care delivery models that disproportionately affect subsets of the
population with certain chronic conditions or health-related behaviors and risk factors. This information leads to
presumably more accurate projections at state and local levels.
DPMM models the implications for patient health and mortality of changes in health risk factors or health-related
behavior. Combining DPMM with HDMM creates the ability to model scenarios of how changes in population
health can affect demand for healthcare services and providers capturing factors that decrease demand (e.g.,
improved health) and factors that increase demand (e.g., extending longevity so there are more people still living).
Model limitations
Many limitations of these microsimulation models stem from data constraints, as well as uncertainties associated
with an evolving care delivery system and medical technology. Conceptual limitations of the model include how
to define whether the size and mix of the health workforce is adequate, distinguishing between demand versus
need for healthcare services and providers, and quantifying the overlapping scope of services provided by
clinicians in different occupations and specialties.
• Data limitations: These include small sample size associated with some surveys, time lags between when data are
collected via surveys or medical claims files and when information becomes available to researchers, and data not
collected or nor reported in ways that are ideal for modeling.
o Supply data: Supply data for many health professions comes from proprietary data sources such
as association master files (e.g., American Medical Association Master File), state licensure files,
and association-based surveys collecting information on their member practices. Other supply data
comes from the U.S. Census Bureau’s American Community Survey, the Department of Labor’s
Occupational Employment Statistics, data collected and reported by CMS, and other published
sources such as association publications. Each of these data sources has limitations. For example,
surveys might have small sample sizes for certain geographic areas and subsets of the workforce
(e.g., older workers, which data are needed to model retirement patterns), or surveys will differ in
the wording of questions they ask regarding workforce decisions such as retirement intentions.
OES data are collected on filled positions that do not distinguish between part-time and fulltime
employment, and under-report employment in occupations where people are self-employed.
Estimates of workforce supply can differ by source—e.g., estimates from the AMA Master File
data can differ from numbers in state licensure files.
o Demand data: One of the major sources of data on healthcare use patterns is MEPS, which
although is a rich source of data has several limitations. The MEPS sample contains approximately
30,700 to 35,100 individuals per year with the sample size varying from year to year. While for
most healthcare use this sample is large, for less frequent events such as hospitalizations and
emergency visits by diagnosis area the sample is relatively small. Therefore, we combine the latest
IHS Markit | HWMM Documentation
Confidential. © 2020 IHS Markit. All rights reserved. 39 May 2020
five years of survey data to increase sample size for analyses that estimate the correlation between
patient characteristics and annual use of healthcare services. Furthermore, base year utilization
estimates based on MEPS are calibrated to estimates from other sources such as NIS and NAMCS
which have large sample size and/or are weighted to more accurately reflect national totals for
specific types of healthcare use.
A limitation of BRFSS as a data source for disease prevalence and health-related behavior among
the population is that as a telephone-based survey it tends to exclude people in institutionalized
settings who typically do not own telephones. Hence, when creating the population files that
underlie the demand projections BRFSS data is supplemented with CMS data containing
information on a representative sample of people in residential care facilities and nursing homes.
• Evolving care delivery system uncertainties: The healthcare system continues to evolve, with
substantial uncertainty in how advances in medicine and technology might affect demand for services.
Some advances might cure disease or reduce the time to perform certain procedures or patient recovery
time (e.g., laparoscopic surgery) thus reducing demand for healthcare providers; however, these and other
advances can make some services more accessible and improve patient outcomes thereby increasing
demand for these services and providers. To address uncertainty in evolving care delivery, we use
sensitivity analysis and model a variety of scenarios involving potential changes in care delivery.
• Conceptual limitations: Conceptual limitations with the workforce models occur, in part, due to lack of
data to clearly define certain aspects of the modeling and thus the reliance on assumptions.
o Supply adequacy: Most workforce studies start with the assumption that at the national level
supply is adequate to meet demand for services unless there is clear evidence of a supply shortfall
or surplus. This assumption of base year national equilibrium essentially presents future adequacy
relative to current levels, and geographic differences in adequacy compared to the national
average. To the extent that there are current gaps between national supply and demand, then such
base year imbalances persist into the projections of future supply adequacy. One approach used to
quantify current national imbalances is to use estimates from HRSA of the number of additional
providers required to remove Health Professional Shortage Area (HPSA) designations—for
primary care, dental care, and mental health.25,44 Another approach is to survey practices on their
“busyness” to determine if they have excess capacity and would prefer to have more work, or if
they are running at or above capacity and would prefer to have less work or turn some new
patients away because of capacity constraints.26,87
o Demand versus need: HDMM models use of healthcare services, with demand modeled as use
patterns under current prices, economic conditions, care delivery patterns and policies, and current
social expectations. The concept of “need” for services implies a clinical judgement, with need for
health providers also based on a specified care delivery model. Need for services does not take
into consideration economic realities or social expectations (e.g., stigma associated with seeking
behavioral health services) that might cause some people with a clinical need to not seek services.
Likewise, demand for services includes any inefficiencies in care delivery—such as provider-
induced demand, the practice of defensive medicine, or patients seeking care that is not needed.
While the model does not capture demand based on need, some of the modeled scenarios address
the topic of inequity in access to receiving services by modeling the demand implications if
historically underserved populations used care at the same levels as populations that historically
have had fewer barriers to accessing care.
IHS Markit | HWMM Documentation
Confidential. © 2020 IHS Markit. All rights reserved. 40 May 2020
o Overlapping scope of practice: The U.S. healthcare system has substantial overlap in capabilities
across health occupations and medical specialties. For example, much of primary care could be
provided by a physician (primary care or specialist physician), NP or PA. The goal of optimizing
the health workforce is to have providers practicing at the top of their abilities in a cost-effective
manner. While analysis of medical claims files might indicate what types of services are being
provided by different types of providers, there is insufficient information in such claims files to
know if the patients’ needs were effectively being met by the provider seen.
These workforce models were developed using a microsimulation approach in part with the goal of reflecting
evolving standards of care, newly enacted policies, and changing economic factors. To date, data limitations have
limited the ability to model some emerging care delivery models. However, increasingly data is becoming
available to model trends in care use and delivery. This “research in progress” is part of ongoing efforts to
continue to refine and improve the microsimulation models.
IHS Markit | HWMM Documentation
Confidential. © 2020 IHS Markit. All rights reserved. 41 May 2020
References 1. Chen F, Dall T, Siego C, Jones K. IHS Markit Disease Prevention Microsimulation Model Technical Report.
IHS Markit; 2020. https://ihsmarkit.com/products/healthcare-modeling.html
2. Zedlewski S, Giannarelli L. TRIM3: Transfer Income Model, Version 3. Urban Institute; 2015.
http://www.urban.org/sites/default/files/publication/82921/2000878-TRIM-A-Tool-for-Social-Policy-
Analysis.pdf
3. Vreman RA, Goodell AJ, Rodriguez LA, Porco TC, Lustig RH, Kahn JG. Health and economic benefits of
reducing sugar intake in the USA, including effects via non-alcoholic fatty liver disease: a microsimulation
model. BMJ open. 2017;7(8):e013543.
4. Bongers ML, De Ruysscher D, Oberije C, Lambin P, Uyl–de Groot CA, Coupe VMH. Multistate statistical
modeling: a tool to build a lung cancer microsimulation model that includes parameter uncertainty and
patient heterogeneity. Medical Decision Making. 2016;36(1):86–100.
5. Çağlayan Ç, Terawaki H, Chen Q, Rai A, Ayer T, Flowers CR. Microsimulation Modeling in Oncology.
JCO Clinical Cancer Informatics. 2018;(2):1-11. doi:10.1200/CCI.17.00029
6. Harland K, Birkin M. Geocomputation: A Practical Primer. In: First Edition. SAGE Publications, Inc.; 2015.
doi:10.4135/9781473916432
7. Reinhardt UE. The GMENAC forecast: an alternative view. Am J Public Health. 1981;71(10):1149-1157.
doi:10.2105/AJPH.71.10.1149
8. Weiner JP. Forecasting the effects of health reform on US physician workforce requirement. Evidence from
HMO staffing patterns. JAMA. 1994;272(3):222-230.
9. Weiner JP. Prepaid group practice staffing and U.S. physician supply: lessons for workforce policy. Health
Aff (Millwood). 2004;Suppl Web Exclusives:W4-43-59. doi:10.1377/hlthaff.w4.43
10. Yett DE, Drabek L, Intriligator MD, Kimbell LJ. Health Manpower Planning: An Econometric Approach.
Health Services Research. Published online 1972:134-147.
11. Cooper RA, Getzen TE, McKee HJ, Laud P. Economic and demographic trends signal an impending
physician shortage. Health Aff (Millwood). 2002;21(1):140-154. doi:10.1377/hlthaff.21.1.140
12. Hogan PF, Bouchery E. Workforce Study of Rheumatologists. The American College of Rheumatology,
prepared by The Lewin Group; 2006. Accessed May 4, 2020.
https://www.rheumatology.org/Portals/0/Files/LewinReport.pdf
13. Yett DE, Intriligator MD, Kimbell LJ. A Microeconomic Model of the Health Care System in the United
States. Annals of Economic and Social Measurement. Published online 1975:407-433.
14. Health Resources and Services Administration. Health Workforce Projections. HRSA Health Workforce.
Published 2018. Accessed December 16, 2018. https://bhw.hrsa.gov/health-workforce-
analysis/research/projections
IHS Markit | HWMM Documentation
Confidential. © 2020 IHS Markit. All rights reserved. 42 May 2020
15. Streeter RA, Zangaro GA, Chattopadhyay A. Perspectives: Using Results from HRSA’s Health Workforce
Simulation Model to Examine the Geography of Primary Care. Health Services Research. 2017;52:481-507.
doi:10.1111/1475-6773.12663
16. IHS Markit. The Primary Care Workforce in Arkansas: Current and Future Supply and Demand (Appendix
B). Prepared for the Arkansas Center for Health Improvement; 2012.
http://www.achi.net/Content/Documents/ResourceRenderer.ashx?ID=206
17. IHS Markit. Maryland Health Workforce Study. Maryland Health Care Commission; 2013.
http://mhcc.maryland.gov/mhcc/pages/plr/plr_workforce/plr_workforce.aspx
18. IHS Markit. Florida Statewide and Regional Physician Workforce Analysis: Estimating Current and
Forecasting Future Supply and Demand. Prepared for Safety Net Hospital Alliance of Florida; 2015.
https://ahca.myflorida.com/medicaid/Finance/finance/LIP-
DSH/GME/docs/FINAL_Florida_Statewide_and_Regional_Physician_Workforce_Analysis.pdf
19. Texas Center for Nursing Workforce Studies. Nurse Supply and Demand Projections, 2015-2030.; 2016.
file:///C:/Users/lum36407/Downloads/SupplyDemandReport_092916%20(1).pdf
20. IHS Markit. Current and Projected Future Health Care Workforce Demand in Vermont. Prepared for the
State of Vermont Agency of Administration; 2017.
http://healthcareinnovation.vermont.gov/sites/vhcip/files/documents/Vermont%20Health%20Care%20Dema
nd%20Modeling%20Final%20Report%206-16-17%20FINAL.pdf
21. Texas Department of State Health Services. Texas Projections of Supply and Demand for Primary Care
Physicians and Psychiatrists, 2017 - 2030. Texas Health and Human Services; 2018. https://dshs.texas.gov ›
SB-18-Physicians-Workforce-Report-Final
22. Texas Department of State Health Services. Texas Projections of Supply and Demand Dental Projections,
2018 - 2030. Texas Health and Human Services; 2019. https://www.dshs.texas.gov/chs/hprc/
23. Dall T, Storm M, Chakrabarti R, et al. Supply and demand analysis of the current and future US neurology
workforce. Neurology. 2013;81(5):470-478. doi:10.1212/WNL.0b013e318294b1cf
24. IHS Markit. Technical Report for Psychologist Workforce Projections for 2015-2030: Addressing Supply
and Demand.; 2018. https://www.apa.org/workforce/publications/supply-demand/technical-report.pdf
25. IHS Markit. The Complexities of Physician Supply and Demand: Projections from 2017 to 2032. Report
prepared for the Association of American Medical Colleges. Published 2019. https://aamc-
black.global.ssl.fastly.net/production/media/filer_public/31/13/3113ee5c-a038-4c16-89af-
294a69826650/2019_update_-_the_complexities_of_physician_supply_and_demand_-
_projections_from_2017-2032.pdf
26. Surdu S, Dall TM, Langelier M, Forte GJ, Chakrabarti R, Reynolds RL. The pediatric dental workforce in
2016 and beyond. The Journal of the American Dental Association. 2019;150(7):609-617.e5.
doi:10.1016/j.adaj.2019.02.025
27. BDO and IHS Markit. Target Workforce State Report for OneCity Health Partners for Health NYC:
Delivery System Reform Incentive Payment Program.; 2016.
https://www.health.ny.gov/health_care/medicaid/redesign/dsrip/workforce_docs/ms1/nychhc.pdf
IHS Markit | HWMM Documentation
Confidential. © 2020 IHS Markit. All rights reserved. 43 May 2020
28. BDO, IHS Markit. Target Workforce State Report for Albany Medical Center Hospital Performing Provider
System: Delivery System Reform Incentive Payment Program Workforce Strategy Deliverable. Albany
Medical Center Hospital Performing Provider System; 2016.
https://www.health.ny.gov/health_care/medicaid/redesign/dsrip/workforce_docs/ms1/albany.pdf
29. West T, Semilla AP, Dall TM. Target Workforce State Report for Adirondack Health Institute PPS.
Adirondack Health Institute PPS; 2016.
https://www.health.ny.gov/health_care/medicaid/redesign/dsrip/workforce_docs/ms1/adirondack.pdf
30. BDO, IHS Markit. Workforce Gap Analysis Report for Nassau Queens Performing Provider System:
Delivery System Reform Incentive Payment Program (DSRIP) Workforce Strategy Deliverable. Nassau
Queens Performing Provider System; 2016.
https://www.health.ny.gov/health_care/medicaid/redesign/dsrip/workforce_docs/ms3/nassau.pdf
31. Center for Health Workforce Studies, IHS Markit. Advocate Community Partners Target Workforce State.
Advocate Community Partners; 2016.
https://www.health.ny.gov/health_care/medicaid/redesign/dsrip/workforce_docs/ms1/advocate.pdf
32. BDO, IHS Markit. Target Workforce State Report for Community Care of Brooklyn. Community Care of
Brooklyn; 2016.
https://www.health.ny.gov/health_care/medicaid/redesign/dsrip/workforce_docs/ms1/maimonides.pdf
33. Dall TM, Gallo PD, Chakrabarti R, West T, Semilla AP, Storm MV. An aging population and growing
disease burden will require a large and specialized health care workforce by 2025. Health Aff(Millwood).
2013;32(1544-5208 (Electronic)):2013-2020. doi:10.1377/hlthaff.2013.0714
34. Dall TM, Chakrabarti R, Storm MV, Elwell EC, Rayburn WF. Estimated Demand for Women’s Health
Services by 2020. Journal of Women’s Health. 2013;22(7):643-648. doi:10.1089/jwh.2012.4119
35. Dall TM, Storm MV, Semilla AP, Wintfeld N, O’Grady M, Narayan KM. Value of lifestyle intervention to
prevent diabetes and sequelae. AmJPrevMed. 2015;48(1873-2607 (Electronic)):271-280.
36. Su W, Huang J, Chen F, et al. Modeling the clinical and economic implications of obesity using
microsimulation. JMedEcon. 2015;18(1941-837X (Electronic)):886-897.
37. Chen F, Su W, Becker SH, et al. Clinical and Economic Impact of a Digital, Remotely-Delivered Intensive
Behavioral Counseling Program on Medicare Beneficiaries at Risk for Diabetes and Cardiovascular Disease.
PLoSOne. 2016;11(1932-6203 (Electronic)):e0163627.
38. Su W, Chen F, Dall TM, Iacobucci W, Perreault L. Return on Investment for Digital Behavioral Counseling
in Patients With Prediabetes and Cardiovascular Disease. Preventing Chronic Disease. 2016;13.
doi:10.5888/pcd13.150357
39. Semilla AP, Chen F, Dall TM. Reductions in mortality among Medicare beneficiaries following the
implementation of Medicare Part D. AmJManagCare. 2015;21(1936-2692 (Electronic)):s165-s171.
40. IHS Markit. What Is the Impact of Chronic Disease on America? Partnership to Fight Chronic Disease;
2017.
https://www.fightchronicdisease.org/sites/default/files/pfcd_blocks/PFCD_US.FactSheet_FINAL1%20%28
2%29.pdf
IHS Markit | HWMM Documentation
Confidential. © 2020 IHS Markit. All rights reserved. 44 May 2020
41. Citro C, Hanushek E. Improving Information for Social Policy Decisions: The Uses of Microsimulation
Modeling – Volume I: Review and Recommendations. National Academy Press; 1991:360.
https://www.nap.edu/catalog/1835/improving-information-for-social-policy-decisions-the-uses-of-
microsimulation-modeling
42. Centers for Disease Control and Prevention. 2013 NCHS Urban-Rural Classification Scheme for Counties.
https://www.cdc.gov/nchs/data_access/urban_rural.htm
43. American Dental Association. 2017-2018 Surveys of Dental Education. Published 2019. Accessed July 18,
2019. https://www.ada.org/en/science-research/health-policy-institute/data-center/dental-education
44. Health Resources and Services Administration. Health Workforce Projections. HRSA Health Workforce.
Published 2020. Accessed May 14, 2020. https://bhw.hrsa.gov/health-workforce-
analysis/research/projections
45. U.S. Congress Joint Economic Committee. Understanding Economic Challenges in Rural America:
Minority Staff Report.; 2017. https://www.jec.senate.gov/public/_cache/files/fe1f3a2f-f89c-4628-ac6a-
0e3b5fb8fafc/rural-economy-report-final.pdf
46. Chan KS, Parikh MA, Thorpe RJ, Gaskin DJ. Health Care Disparities in Race-Ethnic Minority Communities
and Populations: Does the Availability of Health Care Providers Play a Role? J Racial and Ethnic Health
Disparities. Published online December 16, 2019. doi:10.1007/s40615-019-00682-w
47. Crowley R, Daniel H, Cooney TG, Engel LS, for the Health and Public Policy Committee of the American
College of Physicians. Envisioning a Better U.S. Health Care System for All: Coverage and Cost of Care.
Ann Intern Med. 2020;172(2_Supplement):S7. doi:10.7326/M19-2415
48. Butkus R, Rapp K, Cooney TG, Engel LS, for the Health and Public Policy Committee of the American
College of Physicians. Envisioning a Better U.S. Health Care System for All: Reducing Barriers to Care and
Addressing Social Determinants of Health. Ann Intern Med. 2020;172(2_Supplement):S50.
doi:10.7326/M19-2410
49. Bachrach D, Frohlich J, Garcimonde A, Nevitt K. Building a Culture of Health: The Value Proposition of
Retail Clincs. Robert Wood Johnson Foundation; 2015.
http://www.manatt.com/uploadedFiles/Content/5_Insights/White_Papers/Retail_Clinic_RWJF.pdf
50. Jones Lange LaSalle IP, Inc. Retail and the New Healthcare Consumer.; 2019.
https://www.us.jll.com/content/dam/jll-com/documents/pdf/research/americas/us/am-research-retail-and-the-
new-healthcare-consumer.pdf
51. BlueCross BlueShield. Retail Clinic Visits Increase Despite Use Lagging Among Individually Insured
Americans. BlueCross BlueShield; 2017. https://www.bcbs.com/the-health-of-america/reports/rising-costs-
patented-drugs-drive-growth-pharmaceutical-spending-us
52. Ashwood JS, Gaynor M, Setodji CM, Reid RO, Weber E, Mehrotra A. Retail Clinic Visits For Low-Acuity
Conditions Increase Utilization And Spending. Health Affairs. 2016;35(3):449-455. Accessed January 24,
2019. http://www.healthaffairs.org/doi/10.1377/hlthaff.2015.0995
IHS Markit | HWMM Documentation
Confidential. © 2020 IHS Markit. All rights reserved. 45 May 2020
53. Mehrotra A, Wang MC, Lave JR, Adams JL, McGlynn EA. Retail Clinics, Primary Care Physicians, And
Emergency Departments: A Comparison Of Patients’ Visits. Health Affairs. 2008;27(5):1272-1282.
Accessed January 24, 2019. http://www.healthaffairs.org/doi/10.1377/hlthaff.27.5.1272
54. Altschuler J, Margolius D, Bodenheimer T, Grumbach K. Estimating a Reasonable Patient Panel Size for
Primary Care Physicians With Team-Based Task Delegation. The Annals of Family Medicine.
2012;10(5):396-400. doi:10.1370/afm.1400
55. Patel SA, Winkel M, Ali MK, Narayan KMV, Mehta NK. Cardiovascular Mortality Associated With 5
Leading Risk Factors: National and State Preventable Fractions Estimated From Survey Data. Annals of
Internal Medicine. 2015;163(4):245. Accessed January 29, 2019.
http://annals.org/article.aspx?doi=10.7326/M14-1753
56. Mehta NK, Patel SA, Ali MK, Venkat Narayan KM. Preventing Disability: The Influence Of Modifiable
Risk Factors On State And National Disability Prevalence. Health Affairs. 2017;36(4):626-635. Accessed
January 29, 2019. http://www.healthaffairs.org/doi/10.1377/hlthaff.2016.1281
57. Song M, Giovannucci E. Preventable Incidence and Mortality of Carcinoma Associated With Lifestyle
Factors Among White Adults in the United States. JAMA Oncology. 2016;2(9):1154. Accessed January 29,
2019. http://oncology.jamanetwork.com/article.aspx?doi=10.1001/jamaoncol.2016.0843
58. Office of Disease Prevention and Health Promotion. Healthy People: 2020 Topics & Objectives. U.S.
Department of Health and Human Services; 2020. Accessed May 14, 2020.
https://www.healthypeople.gov/2020/topics-objectives
59. Taylor F, Huffman MD, Macedo AF, et al. Statins for the primary prevention of cardiovascular disease.
CochraneDatabaseSystRev. 2013;(1469-493X (Electronic)):CD004816.
60. Baguet JP, Legallicier B, Auquier P, Robitail S. Updated meta-analytical approach to the efficacy of
antihypertensive drugs in reducing blood pressure. ClinDrug Investig. 2007;27(1173-2563 (Print)):735-753.
61. Sherifali D, Nerenberg K, Pullenayegum E, Cheng JE, Gerstein HC. The effect of oral antidiabetic agents on
A1C levels: a systematic review and meta-analysis. Diabetes Care. 2010;33(1935-5548 (Electronic)):1859-
1864.
62. Yeboah P, Hsu F-C, Bertoni AG, Yeboah J. Body Mass Index, Change in Weight, Body Weight Variability
and Outcomes in Type 2 Diabetes Mellitus (from the ACCORD Trial). The American Journal of
Cardiology. 2019;123(4):576-581. doi:10.1016/j.amjcard.2018.11.016
63. Shao H, Fonseca V, Stoecker C, Liu S, Shi L. Novel Risk Engine for Diabetes Progression and Mortality in
USA: Building, Relating, Assessing, and Validating Outcomes (BRAVO). PharmacoEconomics.
2018;36(9):1125-1134. doi:10.1007/s40273-018-0662-1
64. Hayes AJ, Leal J, Gray AM, Holman RR, Clarke PM. UKPDS Outcomes Model 2: a new version of a model
to simulate lifetime health outcomes of patients with type 2 diabetes mellitus using data from the 30 year
United Kingdom Prospective Diabetes Study: UKPDS 82. Diabetologia. 2013;56(9):1925-1933.
doi:10.1007/s00125-013-2940-y
IHS Markit | HWMM Documentation
Confidential. © 2020 IHS Markit. All rights reserved. 46 May 2020
65. Steele CB, Thomas CC, Henley SJ, et al. Vital Signs: Trends in Incidence of Cancers Associated with
Overweight and Obesity — United States, 2005–2014. MMWR Morb Mortal Wkly Rep. 2017;66(39):1052-
1058. doi:10.15585/mmwr.mm6639e1
66. Proia KK, Thota AB, Njie GJ, et al. Team-Based Care and Improved Blood Pressure Control. American
Journal of Preventive Medicine. 2014;47(1):86-99. Accessed March 1, 2018.
http://linkinghub.elsevier.com/retrieve/pii/S0749379714001226
67. Costello T, Dorrell M, Kellams T, Kraska K. Review of Pharmacologic Weight Loss Medications in a
Patient-Centered Medical Home. Journal of Pharmacy Technology. 2016;32(1):37-41. Accessed August 13,
2018. http://journals.sagepub.com/doi/10.1177/8755122515604858
68. Hibbard JH, Greene J, Sacks RM, Overton V, Parrotta C. Improving population health management
strategies: identifying patients who are more likely to be users of avoidable costly care and those more likely
to develop a new chronic disease. Health Services Research. 2017;52(4):1297–1309.
69. Lauffenburger JC, Shrank WH, Bitton A, et al. Association Between Patient-Centered Medical Homes and
Adherence to Chronic Disease Medications: A Cohort Study. Annals of Internal Medicine. 2017;166(2):81.
Accessed July 10, 2018. http://annals.org/article.aspx?doi=10.7326/M15-2659
70. Agarwal R, Gupta A, Fendrick AM. Value-Based Insurance Design Improves Medication Adherence
Without An Increase In Total Health Care Spending. Health Affairs. 2018;37(7):1057-1064. Accessed
August 3, 2018. http://www.healthaffairs.org/doi/10.1377/hlthaff.2017.1633
71. Look KA. Value-based insurance design and medication adherence: opportunities and challenges. Am J
Manag Care. 2015;21(1):e78–90.
72. Morgado MP, Morgado SR, Mendes LC, Pereira LJ, Castelo-Branco M. Pharmacist interventions to enhance
blood pressure control and adherence to antihypertensive therapy: Review and meta-analysis. American
Journal of Health-System Pharmacy. 2011;68(3):241-253. Accessed March 1, 2018.
http://www.ajhp.org/cgi/doi/10.2146/ajhp090656
73. Yang W, Dall TM, Zhang Y, et al. Simulation of quitting smoking in the military shows higher lifetime
medical spending more than offset by productivity gains. Health Aff(Millwood). 2012;31(1544-5208
(Electronic)):2717-2726.
74. Jha P, Ramasundarahettige C, Landsman V, et al. 21st-Century Hazards of Smoking and Benefits of
Cessation in the United States. N Engl J Med. 2013;368(4):341-350. Accessed January 29, 2019.
https://doi.org/10.1056/NEJMsa1211128
75. Centers for Disease Control and Prevention. Smoking & Tobacco Use: Health Effects. Published 2020.
https://www.cdc.gov/tobacco/basic_information/health_effects/index.htm
76. Brotherton SE, Etzel SI. Graduate Medical Education, 2018-2019. JAMA. 2019;322(10):996.
doi:10.1001/jama.2019.10155
77. Abraham CM, Zheng K, Poghosyan L. Predictors and Outcomes of Burnout Among Primary Care Providers
in the United States: A Systematic Review. Med Care Res Rev. Published online November 18,
2019:107755871988842. doi:10.1177/1077558719888427
IHS Markit | HWMM Documentation
Confidential. © 2020 IHS Markit. All rights reserved. 47 May 2020
78. Busis NA, Shanafelt TD, Keran CM, et al. Burnout, career satisfaction, and well-being among US
neurologists in 2016. Neurology. 2017;88(8):797-808. Accessed June 29, 2018.
http://www.neurology.org/lookup/doi/10.1212/WNL.0000000000003640
79. Cimbak N, Stolarski A, Moseley J, O’Neal P, Whang E, Kristo G. Burnout Leads to Premature Surgeon
Retirement: A Nationwide Survey. J Surg Res. 2019;2(3):159–169.
80. Crowe RP. An Assessment of Burnout among Nationally-Certified Emergency Medical Services
Professionals. Published online 2016.
81. El-Ibiary SY, Yam L, Lee KC. Assessment of Burnout and Associated Risk Factors Among Pharmacy
Practice Faculty in the United States. Published June 8, 2017. Accessed August 20, 2018.
https://www.ajpe.org/doi/abs/10.5688/ajpe81475
82. Eliot KA, Kolasa KM, Cuff PA. Stress and Burnout in Nutrition and Dietetics: Strengthening
Interprofessional Ties. Nutrition Today. 2018;53(2):63-67. doi:10.1097/NT.0000000000000269
83. Essary AC, Bernard KS, Coplan B, et al. Burnout and Job and Career Satisfaction in the Physician Assistant
Profession: A Review of the Literature. NAM Perspectives. Published online December 3, 2018.
doi:10.31478/201812b
84. Mahajan A, Skinner L, Auerbach DI, Buerhaus PI, Staiger DO. Association between the growth of
accountable care organizations and physician work hours and self-employment. JAMA Network Open.
2018;1(3):e180876. http://dx.doi.org/10.1001/jamanetworkopen.2018.0876
85. Health Resources and Services Administration. Technical Documentation for Health Resources Services
Administration’s Health Workforce Simulation Model. U.S. Department of Health and Human Services;
2020. Accessed May 15, 2020. https://bhw.hrsa.gov/sites/default/files/bhw/health-workforce-
analysis/research/projections/technical-documentation-health-workforce-simulation-model.pdf
86. IHS Markit, Center for Health Workforce Studies. 2013 U.S. Veterinary Workforce Study: Modeling
Capacity Utilization. American Veterinary Medical Association; 2013.
https://www.avma.org/sites/default/files/resources/Veterinarian-Workforce-Final-Report.pdf
87. Dall TM, Forte GJ, Storm MV, et al. Executive summary of the 2013 US Veterinary Workforce Study.
Journal of the American Veterinary Medical Association. 2013;242(11):1507-1514.
doi:10.2460/javma.242.11.1507
IHS Markit | HWMM Documentation
Confidential. © 2020 IHS Markit. All rights reserved. 48 May 2020
Tim Dall Executive Director [email protected]
Ryan Reynolds Senior Consultant [email protected]
Ritashree Chakrabarti Senior Consultant [email protected]
Will Iacobucci Senior Consultant [email protected]
Kari Jones Associate Director [email protected]
Contacts
Disclaimer
The information contained in this report is confidential. Any unauthorized use, disclosure, reproduction, or dissemination, in full or in part, in any media or by any means, without the prior written permission of IHS Markit or any of its affiliates ("IHS Markit") is strictly prohibited. IHS Markit owns all IHS Markit logos and trade names contained in this report that are subject to license. Opinions, statements, estimates, and projections in this report (including other media) are solely those of the individual author(s) at the time of writing and do not necessarily reflect the opinions of IHS Markit. Neither IHS Markit nor the author(s) has any obligation to update this report in the event that any content, opinion, statement, estimate, or projection (collectively, "information") changes or subsequently becomes inaccurate. IHS Markit makes no warranty, expressed or implied, as to the accuracy, completeness, or timeliness of any information in this report, and shall not in any way be liable to any recipient for any inaccuracies or omissions. Without limiting the foregoing, IHS Markit shall have no liability whatsoever to any recipient, whether in contract, in tort (including negligence), under warranty, under statute or otherwise, in respect of any loss or damage suffered by any recipient as a result of or in connection with any information provided, or any course of action determined, by it or any third party, whether or not based on any information provided. The inclusion of a link to an external website by IHS Markit should not be understood to be an endorsement of that website or the site's owners (or their products/services). IHS Markit is not responsible for either the content or output of external websites. Copyright © 2020, IHS Markit®. All rights reserved and all intellectual property rights are retained by IHS Markit.
IHS Markit Customer Care:
Americas: +1 800 IHS CARE (+1 800 447 2273)
Europe, Middle East, and Africa: +44 (0) 1344 328 300
Asia and the Pacific Rim: +604 291 3600