Local Area Physician Workforce Planning Model Pilot

Physician workforce planning tools are needed to address the adequacy of the nation’s physician workforce to meet its health needs. Since health care is an inherently local phenomenon, physician supply and demand need to be understood in the local area context. Therefore, we seek to build a Local Area Physician Workforce Planning Model that emphasizes a local perspective. To that end, we are developing a process to improve physician workforce planning to better align workforce capacity with community needs. The cornerstone of this effort is the extensive use of an iterative series of group model building processes that engage local system stakeholders in creating a system dynamics-based model of physician workforce capacity and need that facilitates planning and the evaluation of potential policy changes. We are approaching the conclusion of our work in our pilot site, the U.S. metropolitan area of Cleveland, Ohio. The model and development process have generated important lessons, including: the importance of advance preparation for group model building sessions; how a combination of group model building and data analysis can verify shared mental models and highlight key policy levers; the difficulties associated with modeling complex supply chains; and the limitations imposed by a dearth of suitable data. Overview We are in the early stages of a long term system dynamics modeling project that seeks to address the question of how the physician workforce can be configured to best meet a local population’s health needs. Complicating this task is a dynamic environment where payment and care delivery models are evolving, significant changes are occurring in federal and state level policies, and physician work patterns are shifting dramatically. Our goal is to create a model-based tool to support policy decision-making. This requires an enhanced understanding of how supply and demand forces dynamically shape how well an area’s physician workforce capacity meets its health care needs. Past physician workforce projections that we published were based on linear models which could not be applied at the local area level, where most physician workforce decisions are made and where physician services are provided, in a way that reflects that area’s unique circumstances. We have therefore moved to a community-based group model building approach, grounded in local stakeholder input and centered on a systems-based workforce projection model and policy evaluation tool designed explicitly to help localities design effective policy interventions. Local decision-makers have a number of policy and programmatic “levers” available to them as they try to adjust the health workforce to meet the population’s health care needs. These levers include the size and mix of local training programs, incentives to retain people completing training and those who might migrate in from other areas, and the numbers and kinds of positions offered and compensation and other conditions that go with those positions. The problem faced by local decision makers is that there is no framework for understanding how their decisions affect each other and the size and makeup of the health care workforce. Without such a framework, it is difficult for them to come together and create a set of workforce programs and policies in a coherent manner. This model is designed to meet that need by integrating the various aspects of physician workforce supply and demand. We are approaching the conclusion of our work in our pilot site, the U.S. metropolitan area of Cleveland, Ohio. The model and model development process we have established thus far have generated some important lessons for us as model builders, including: the importance of advance preparation for group model building sessions; how a combination of group model building and data analysis can verify shared mental models and highlight key policy levers; the difficulties associated with modeling inherently complex supply chains; and the limitations imposed upon modeling by a dearth of suitable data. This paper will describe the process of model development including group model building, the structure of the pilot model and some simulations it produced, data dilemmas we faced, and plans for future development. Background Physician workforce planning tools are needed to address the adequacy of the nation’s physician workforce to meet its health needs, including alarm over insufficient absolute numbers and specialty mix, capacity to adapt to new models of care and payment, and chronic geographic maldistribution (IHS, 2015; Dill & Salsberg, 2008; Colwill et al, 2008; Peterson et al, 2012; Phillips et al, 2009). Key issues in this context include: a rise in overall utilization of health care services (CDC, 2014); the increasing need for management of multiple chronic conditions (IOM, 2001); changes in physician work hours, efficiency and retirement patterns; an aging population and workforce (IHS, 2015; Dill & Salsberg, 2008); and pressures to change the way both Medicaid and Medicare pay for health care, such as the Bundled Payments for Care Improvement (BPCI) Initiative (CMS, 2014). In 2008, the AAMC released The Complexities of Physician Supply and Demand: Projections Through 2025; and in 2015, IHS prepared The Complexities of Physician Supply and Demand: Projections from 2013 to 2025 for AAMC (IHS, 2015). These reports – major resources for policy advocacy, research, and analysis on the physician workforce – contain series of scenariobased physician workforce projections for the U.S. However, the models used could not be applied at local area level, where a great deal of variation can occur. Indeed, while national numbers are important, health care is an inherently local phenomenon, needed, sought, and provided at the local area level. Relative supply, demand, and access levels vary greatly by geographic area. Thus, physician supply and demand need to be understood in the local area context in order to inform both local and national level physician workforce planning discussions. We therefore seek to illuminate the nuances of physician workforce planning that require insights into the local dynamics of supply and demand, specialty mix, socioeconomic challenges with seeking and accessing health care, and challenges posed by state and local policy environments. Even more significantly for our purposes, physician workforce planning happens largely within local health systems, where most physicians are employed and provide their services. We set out to build a Local Area Physician Workforce Planning Model that emphasized a local perspective. We are developing a process for helping localities and their health systems to improve their physician workforce planning to better align their workforce capacity with community health care needs. The cornerstone of this effort is a system dynamics-based computer simulation model of physician workforce capacity and need that facilitates workforce planning and the evaluation of potential policy changes, especially those related to physician workforce recruitment and training. Methodology Models should be based on our best understanding of how the underlying system works and evolves. A systems-based methodology is most appropriate for this modeling work, as it tends to suit the dynamic complexity of problems within health systems (Homer & Hirsch, 2006). Moreover, a system dynamics model focuses on endogenous explanations for changes over time in key variables (Richardson, 2011). This endogenous focus enables users to identify leverage points for achieving change, making system dynamics models particularly well-suited to local area workforce planning needs, as they provide stakeholders options for action, rather than merely reaction. The endogenous focus is also crucial for capturing the effects of interactions between workforce supply and demand. The modeling work described in this paper also draws on extensive experience in applying system dynamics methods to modeling health care systems, care delivery, and population health, both for general populations and people facing particular health problems such as chronic illnesses (Homer et al, 2004; Milstein et al, 2010; Hirsch et al, 2012; Hirsch et al, 2014). This work has been important for identifying the structures that affect the health of populations and the nature of the health care that they receive. We are drawing upon the direct involvement of local system stakeholders in model conceptualization and development, employing an iterative group model building process. Building the model in this fashion has led to a better understanding of local health care systems and the effects that policy changes are likely to have on them, because the model is informed and validated by those with firsthand knowledge of those systems. Thus, the group model building process lies at the heart of our modeling strategy, bringing together stakeholders from across the health care landscape as a type of learning collaborative with a shared interest in a common problem. Well established in the field of system dynamics (e.g., Richardson & Andersen, 1995), group model building allows stakeholders with diverse perspectives, and even diverse problem definitions, to share their views and critically examine them in a collaborative environment to collectively create a better overall understanding of the problem (Vennix, 1996). The strength of group modeling building lies in the engagement of stakeholders, the explicit sharing of mental models, the use of simulation to test hypotheses, and ultimately moving participants toward a shared confidence in the new mental models that emerge from their collaboration (Richardson & Anderson in Kilgour & Eden, 2010). The process is conducted through facilitated face-to-face meetings with stakeholders to elicit model structure in an inductive fashion and engage participants directly in the process of model conceptualization and formulation. Facilitators employ a set of “scripts” intended to elicit the desired types of input from participants (Andersen et al, 2007; Luna-Reyes et al, 2006). Groups unfamiliar with technical modeling methods are introduced to their language and symbols through a set of small concept models constructed specifically for this purpose, using simple pictures labeled in the group’s lexicon to draw them into this approach (Richardson, 2006). Inspired by work done on the ReThink Health project (Rippel Foundation, 2015), we opted for a strategy of working with a series of local area sites in order to develop a general Local Area Physician Workforce Planning Model, and a process for adapting and implementing the model. We will be working with one site after another, modifying and improving our model and our process for working with local area stakeholders at each site. We will keep calibrating the model and fine-tuning the process in each specific place, until we have a model and process that are polished enough that they can be implemented in any area with engaged stakeholders. Not having done anything like this before, we needed a pilot site, where we could begin with a deep dive into the local health care landscape, learning its contours: the pressing health, health care, and physician workforce issues; and getting to know the key stakeholders. Process Given its focus on developing a local area health workforce that meets local community needs, and its potential for generating initial local area contacts, we opted to select one of the sites involved in another AAMC cooperative endeavor, Urban Universities for HEALTH (Urban Universities for HEALTH, 2015). After a preliminary review of basic demographic and health systems data on four of the five main Urban Universities for HEALTH sites (Cincinnati, Cleveland, Kansas City and Albuquerque – having discounted Brooklyn as inappropriate for a pilot site), we decided to focus on Cleveland. It is a largely urban area, leaving the complexity of rural-urban disparities for later sites; it is the site of some significant innovation initiatives, such as Aligning Forces for Quality (Robert Wood Johnson Foundation, 2015) and Centers for Medicare and Medicaid Innovation initiatives (CMS, 2015); there are medical schools and other health professions schools present; and it appeared to offer good relevant data at the state, if not metropolitan area, level. Beginning with our initial local contacts, who represented Urban Universities for HEALTH partners at Northeast Ohio Medical University and Cleveland State University, we eventually established comprehensive Cleveland area stakeholder contacts through extensive networking, including site visits to Cleveland, as well as phone calls and e-mails to establish, develop, and maintain connections. Our current stakeholder network now comprises representatives of all the area’s major health systems, local and state government, area medical schools, practitioners, consumer advocacy groups, non-profit health services research organizations, Federally-qualified Health Centers, private insurance, the local Veterans Administration medical center, the city school district, and the state boards of medicine and nursing. What we are trying to do appears to resonate with actors across Cleveland’s health care landscape as a necessary and highly valued endeavor. Bringing the group together After a series of telephone and e-mail communications, as well as a site visit and tour of the Cleveland area that focused on the geospatial arrangement of its neighborhoods and major health care facilities, we began to plan our first group model building session. We wanted to ensure that participants in our group model building process were of a sufficient seniority level to have a system-wide perspective, but still connected enough to daily operations to have a good sense of the specific issues faced by practitioners and their patients. We also needed to achieve an appropriate mix of system leaders, practitioners, trainers, trainees, analysts, and patient advocates. We were able to draw on our network of contacts to recruit them directly, or those they recommended based on who would be best suited to the intensive, technical, and creative work we planned to undertake. GMB I The first group model building session was held on June 11, 2014, at Cleveland State University. We chose a neutral site – one with no affiliation with the major health systems – in order to put everyone on an equal footing and thwart potential competitive posturing which we had been cautioned about in the course of our information gathering. For our first group model building session, participants included executive suite leaders in the two largest health systems, physicians, experts in healthcare economics, innovators in health care education, and a leading patient advocate. All participants had been primed on our work through prior communications that included structured and semi-structured interviews, but none had a background in system dynamics modeling. Prior to the first group model building session, AAMC contracted with Gary Hirsch, a leading system dynamics modeler in the field of health care who helped to develop the HealthBound and the ReThink Health models (Rippel Foundation, 2015; Hirsch et al, 2012; Milstein, Homer and Hirsch, 2010; CDC, 2008), to lead the technical model building, and with David Andersen and George Richardson to facilitate the group model building session itself. Drs. Andersen and Richardson developed a detailed agenda and accompanying scripts, a list of which plus one example can be found in Appendix A. These scripts were used to guide participants through a primer on the technical aspects of model building, as well as a series of group activities aimed at illuminating and elaborating on mental models of the way that the local health system works and the problems it faces, and to begin the process of mapping out the model and the stories it should tell (Richardson and Andersen, 2010). The first session involved 9 participants (see Appendix C) in an eight hour day that started with a brief description of the purpose of the gathering, with AAMC as convener and listener, and an overview of the day. Next, we began to use the scripts, as participants described their hopes and fears for the day, for their community, and for the model and modeling process. This activity was particularly beneficial as there was considerable consensus in both the hopes and fears described. Cleveland health care is a particularly competitive environment, and this activity helped break down preconceived notions of the big issues concerning the various players. A “small but wrong” model was then used to introduce modeling terminology and processes. Participants were next asked to graph key variables the model should display, and the behavior of those variables over time. This was followed by a discussion of the policy levers that might influence change in those variables. The variables focused primarily on factors related to the of training physicians, public health service utilization, and overall health, while the policy levers emphasized provider education and organizational policies, expanded services, and patient education campaigns. Participants honed in on key policy gaps in primary care and public health, specifically service gaps faced by the area’s vulnerable poor and minority populations in the inner city, as well as the training needed to meet these needs. We asked participants to develop “system policy stories” (another script) to show what causes and impacts these issues. Three groups developed the following ideal scenarios:  “Access for All”, which focused on providing appropriate care and encouraging service in underserved areas.  “Public Health as Though the Public Matters”, which focused on building a healthy community infrastructure and connecting patients with community resources.  “Enhancing Connectivity between the Health Care System and Local Community Health”, which focused on building bridges between resources and providers. During the development of system stories, participants identified the most important stakeholders in implementing the policies that were an integral part of those stories. These stakeholders were plotted on a “power and interest” grid, shown in Appendix B, allowing us to identify several key constituencies who were missing from the group of stakeholders we had brought together for the group model building and broader project: trainees (medical students and residents); specialty care providers; community health centers; and the other prominent health systems in the area. The day concluded with Mr. Hirsch showing the participants a simple model that he had developed during the session to demonstrate how their day’s work translated into the icons and language of a system dynamics model. This effectively demonstrated that the key variables and policy levers they had identified translated effectively into a system dynamics model that could be used to tell the types of system stories they wanted told. Initial model building This initiated an iterative process where each succeeding session provided feedback on the modeling team’s interpretation of the last session and shifted into next steps for further model improvement. Thus, we used the model from the conclusion of our first group model building session as the basis of the actual system dynamics model we subsequently built. In the months following that initial group model building session, the modeling team built and refined model structures identified and defined by the group. We then identified various data that could be used in the model, employing a series of “fall back” positions. Where we could find data specific to the actual study area (the Cleveland-Elyria-Mentor Ohio Metropolitan Statistical Area), we used those. If the parameters we sought to populate could not rely on local area data, then we looked for data at the state level (Ohio). Failing that, we sought out national level data. Where no data were available, we relied on the literature for estimates. GMB II The second group model building session was held in the same location on November 20, 2014, for another eight hours. This session focused on eliciting stakeholder feedback on the model building that had occurred as a result of the first session. We were able to secure a panel of 17 participants (see Appendix C), with the only absence being an administrator for the countyowned safety net hospital, though practitioners from that system were represented. After a brief catch-up on methodology for new participants, we reviewed the model and the outputs it could display, talking through the points where we knew there were gaps. The participants divided into groups to go through each portion of the model, discussing what needed to be changed, added, and removed, and where each participant might help us fill in data gaps. This took the majority of the day, as participants discussed supply for both primary care and specialty care, and demand for the same. The group benefited from the inclusion of a health economist who understood the labor market flows, and the leaders of a pipeline program who could illuminate key issues with the way we had designed training structures. In the concluding activity, we asked participants to prioritize the variables that should be manipulated and the outputs they want to see, as they are the ultimate end users of the model. We received fascinating feedback, particularly on when workforce recruiting and planning begin (with high school students at the least and elementary school students at the ideal), the complexity of the path through medical training, and the degree to which the health system relies upon Advanced Practice Registered Nurses (APRNs) and Physician Assistants (PAs) despite sometimes restrictive scope of practice regulations. These items constituted the primary focus for the next revision of the model. Revising the model In the months following the second group model building session, we attempted to make the requested revisions to the model. Primary among them was building out the training pipeline. Because of the complex process for entering and completing training, as well as the variety of ways that people can change their path, this proved particularly challenging. The number of possible education pathways a physician can follow, especially if we extend the pipeline back to where they attended high school, numbers in the thousands. Moreover, because the medical school through residency education timeline is so long, typically taking from 8 to 14 years depending on the specialty, and this pipeline can be exited and re-entered multiple times, the data available to trace the pathways cannot always be assumed to be complete. Data collection for the earliest point in which our stakeholders are interested, location of high school, only began in 2002. Given the three-year lag in updating the primary database on practicing physicians, the American Medical Association’s Masterfile, we encounter a data window of 9 years: less than that required for most physicians to complete their training. Nonetheless, we were able to perform some important analyses on data related to physicians’ pathways through the educational pipeline. In particular, group model building participants had emphasized policy levers related to “growing their own”, i.e., recruiting and training future physicians from within the local population. Originally highlighted during a group model building script that focused specifically on policy levers available and important to our stakeholder participants, this concept lent itself to quantitative investigation in two different ways: creation and analysis of original survey data on physicians practicing in the area; and analysis of existing data on where practicing physicians attended high school. In conjunction with our pilot modeling work in Cleveland, we fielded a survey of physicians practicing in Cuyahoga County, the core county within the Cleveland MSA (2,500 surveys mailed, with 40% response rate prior to third mailing which is currently in field). Based on group model building participants’ expressed interest in “growing their own”, we included a question on that survey about where physicians attended high school. Our results showed that more than a third of practicing physicians in that County had attended high school in the Cleveland metropolitan area, and half within the state of Ohio, where Cleveland is located. Moreover, for the subset of practicing physicians nationwide for whom AAMC has data on high school location, and who matriculated to medical school during the 2002-2005 period (the first years we collected high school location data), we found that those who attended high school in Cleveland were 26 times as likely to practice there when compared with those who attended high school anywhere else in the U.S. We hope to follow up on that finding with further analyses. GMB III The third all-day group model building session was held on March 18, 2015, at Cleveland State University. This session had 7 attendees (see Appendix C), though many who could not attend expressed interest in continuing to participate. We are currently completing a series of web-based conferences with them in order to elicit their input, as well. The content for this session focused on key model structures which we had not been able to develop satisfactorily, and validating data we had collected and analyzed. The group had particular interest in the physician pipeline and the use of APRNs and PAs. We were able to flesh out some key policy levers for training decisions made by programs and students, and where the decision points fit in the system. We also devoted considerable time to reviewing some detailed data we had developed for the pipeline parameters, paying particular attention to concerns we had over their veracity. In compiling them for the most recent version of the model, we had discovered some crucial problems with core data on physician supply and retention rates at different stages of the educational pipeline. Stakeholders were able to help us understand what we were seeing in the data, confirm our suspicions that some of the data were in error (though they were also able to help us validate many of the data points we shared), and even provide us with more accurate data through their own access to the sources of those data points. Informed by our participants’ feedback and data, we are now conducting analyses to improve the pipeline retention rates; and we have initiated a longer term project to develop a method for improving our physician supply estimates. We are have not yet learned all the lessons and built all the structures necessary to call the model complete. We do need to move on to a new site in order to expand key aspects of the model, such as the need for a rural component, and to improve the process through replication. However, we established a plan for maintaining our contact with pilot site participants as we move forward. The group showed interest in participating in ongoing conversations, both with the modeling team and with other sites, through telephone and web conference. The main focus of these sessions will be to discuss potentially major changes to the model structure and identify data needs that they may be able to resolve for their own site. We will follow up with our Cleveland contacts at the conclusion of the multi-site local model development project to deliver a final model that can be used for the purpose of ongoing, locality-wide workforce planning. The model The model is still in development, as we are still just completing work at the pilot site. Nonetheless, the basic structure of the model is fairly well developed (Figure 1). It includes supply chains for physicians and the complementary professions of PAs and APRNs, whose participation in care delivery directly affects the capacity of care physicians can provide, and demand driven by population health, age, and income. Figure 1. Overview of Pilot Local Area Physician Workforce Model Figure 2 highlights the feedback loops inherent in the system. Loops A and B adjust the local area population based on births and deaths. Loops C and D are balancing loops in which greater needs for care ideally result in the greater demand for care, utilization of more care, improving health, and limiting future needs for care. However, the delivery of care is constrained by the capacity to provide both primary and specialty care. Capacity depends on the number of physicians and mid-levels employed for primary and specialty care plus residents and fellows who can also provide care. Loops E and F adjust employment to accommodate utilization plus provide some percentage of slack capacity by increasing or decreasing positions available. Employment also depends on the willingness of people to fill vacant positions. Loops G and H embody the response of physicians, PAs and APRNs who may migrate from other parts of the country to fill vacancies and loops I and J affect the willingness of people trained in local programs to remain in the area depending on the job opportunities available. Population Distributed by Age, Income, and Health Status Demand for Primary Care Demand for Specialty Care Visits, Procedures Primary Care Visits Delivered Specialty Visits and Procedures Delivered Fraction with Insurance Coverage and Types of Coverage Capacity to Deliver Primary Care Visits, by Setting (Office, CHC, OPD) Capacity to Deliver Specialty Visits and Procedures, by Setting (Office, CHC, OPD) Primary Care Physicians by Age, Gender, and Setting Specialty Physicians by Age, Gender, and Setting Fractions of Facilities Accepting Medicaid, by Setting Average Patient Care Hours by Age and Gender Residents in Local Primary Care Training Programs Residents in Local Specialty Training Programs Fractions Completing Local Programs Retained in Area Physicians Coming from and Going Elsewhere Nurse Practitioners by Age and Setting Physician Assistants by Age and Setting Medical Students in Local Programs Nurse Practitioners in Local Training