Provider Retention in High Need Areas

Final Report
Sebastian Negrusa, Projesh Ghosh and John T. Warner

Prepared for: Assistant Secretary for Planning and Evaluation

Submitted by: The Lewin Group, Inc.

Acknowledgements

Sebastian Negrusa, PhD, John Warner, PhD and Projesh Ghosh, PhD contributed to this report. We benefited greatly from discussions with members of our technical expert panel on the recruiting of providers into the National Health Services Corps (NHSC) program and their retention in under-served areas. We also gratefully acknowledge the input from Caroline Taplin, the ASPE Project Officer. This study was conducted by The Lewin Group under contract number HHSP23320095639WC, task order number HHSP23337023T, with the HHS’s Office of the Assistant Secretary for Planning and Evaluation. The authors take full responsibility for the accuracy of material presented herein. The views expressed are those of the authors and should not be attributed to ASPE or HHS.

Executive Summary

Since 1972, the National Health Services Corps (NHSC) has provided scholarships and loan repayment incentives for a large number of primary care, mental health, and dental health clinicians to work in Health Professional Shortage Areas (HPSAs) across the country. The NHSC deploys annually almost 9,000 health care professionals to thousands of sites across the country. These clinicians include physicians, dentists, certified nurse practitioners, certified nurse-midwives, physician assistants, registered dental hygienists, health service psychologists, licensed clinical social workers, psychiatric nurse specialists, marriage and family therapists, and licensed professional counselors. The objective of this study was to examine short- and long-term retention in high-need areas of providers who participated in the NHSC’s Loan Repayment Program (LRP) and Scholarship Program (SP) and compare their retention with retention of non-participants. We also conducted multivariate regression analyses to provide additional insights into the individual-level factors and local area characteristics that are associated with retention of program participants in HPSAs.

For this study, we constructed two provider-level analytic datasets based on NHSC administrative datasets and other publicly available and proprietary datasets. The first analytic dataset was constructed using NHSC data, Provider360 data (a proprietary dataset from OptumInsight Corporation, including detailed information on virtually the entire population of medical providers) and Medicare provider data. Using a number of individual-level variables, we matched the NHSC administrative files with Provider360 data and then with a Medicare provider file, obtained by rolling up Medicare claims at the provider level. The resulting analytic file contains annual information on over 1 million non-NHSC providers as well as on 8,973 NHSC participants (out of the total of 22,703 participants from NHSC data). The resulting dataset allows us to track the annual location of participants and non-participants (at the zip code level) over the timeframe between 2005 and 2011. The main advantage of this analytic dataset stems from the fact that it allows us to track providers after they complete their NHSC service.

The second analytic dataset relied mainly on NHSC and Provider360 data and its main feature was that it provided information on the NHSC providers’ location in two points in time: the year of NHSC program termination and December 2013 (i.e., the time when Provider360 information was recorded). The number of NHSC participants in this dataset increases to about 18,500. Nonetheless, once we compared the HPSA retention rates of participants after program completion from the first analytic dataset with the retention rates obtained with the second analytic dataset, we found them to be virtually indistinguishable from each other for primary care and mental health providers. ¹ We concluded that the first dataset provided a sufficiently representative picture of the entire sample of NHSC participants, and in addition it provided us with the opportunity to track providers’ location yearly. As a result, most of our quantitative analyses were conducted using this dataset.

It is also important to note that we dropped from the retention analyses the NHSC participants who left service in 2013, the last year of our timeframe. Of the initial sample of 8,973 participants we identified in the first analytic dataset, we ended up using a number of 6,296 participants, while of 18,500 participants in the second analytic dataset, we ended up using a number of 11,210 participants.

To guide the interpretation of our empirical analyses, we constructed a conceptual framework in which we modeled the location choices of providers across HPSA and non-HPSA locations. This framework yielded a sharp prediction – retention of NHSC participants in HPSAs after the completion of their obligations can never be higher than the retention of providers who choose to locate in those areas without participating in the program. The most important ingredients in our theoretical model were the concepts of being ‘fit’ for the program and the individual provider’s ‘preference’ for a HPSA location (both depending on individual characteristics that are unobservable in the data available for this project). The correlation between these two variables has a direct implication for the retention of participants in HPSAs. In the limiting case where acceptance into the NHSC program is based solely on preferences for being in a HPSA location, the program selects individuals who would have served in high-need areas even in the absence of the program, and therefore retention differences between participants and non-participants are nil. As the correlation between location preferences and program fit weakens, the program tends to select - at least to some extent - individuals who would not have gone to high-need areas in the absence of the program. In this case, participant retention after program completion will tend to be lower than non-participant retention. Somewhat counterintuitively, a lower retention rate for participants is a signal of the program’s success (not failure) in attracting to high-need areas providers who would not have located there in the absence of the program. As we show in a model simulation in Chapter VIII, the number of provider-years in HPSAs the NHSC obtains from participants after program completion is highest when the correlation between fit for the program and HPSA preference is zero.

We defined four measures of retention at the provider level:

• Serving in the same HPSA and in the same county (‘same HPSA – same county’). This variable takes the value of 1 if the NHSC provider remains in the same county as the one where he or she served in the NHSC, and 0 otherwise.

• Serving in a HPSA in another county (‘HPSA – other county’). This variable takes the value of 1 if the NHSC provider remains in a HPSA that is located in a different county than the one in which he or she served while in NHSC service, and 0 otherwise.

• Serving in a non-HPSA from the same county (‘non-HPSA – same county’). This variable takes the value of 1 if the NHSC provider moves to a non-HPSA area from the county where he or she served while in the NHSC, and 0 otherwise.

• Serving in a non-HPSA in another county (‘non-HPSA – another county’). This variable takes the value of 1 if the NHSC provider moves to a non-HPSA area from another county than the county he or she served while in the NHSC, and 0 otherwise.

Combining the first two measures, we also constructed an ‘any HPSA’ measure, taking the value of 1 if the provider remains in any HPSA and 0 otherwise. To ensure comparability across all providers, we defined these measures for non-participants as well. While in the case of participants, the reference point in time was the end of the NHSC service, for non-participants we chose the first year (“start year”) that the non-participants appear in Medicare data as their reference point-in-time. Calculating these measures for each cohort, we were able to construct retention rates one year after separation from NHSC, two years after separation from the NHSC and so on. In the case of non-participants the annual retention rates were calculated as one year since start year, two years since start year and so on for each cohort.

Using the first analytic dataset, we found that about 49% of NHSC primary care participants were located in the same HPSA one year after obligation completion, and 82% were located in any HPSA location (Figure ES.1). By the 6^th year after obligation completion, 35% of participants were located in the same HPSA where they served during NHSC service, and 72% of them were in any HPSA location. Consistent with the main prediction of our theoretical model, non-participant retention in HPSAs is higher, with the difference being much bigger for retention in the same HPSA than retention in any HPSA location. Another important finding was that much of the geographic mobility of participants was from one HPSA location to another HPSA location. Also, after an initial higher mobility, participants have better retention in HPSAs than non-participants.

Figure ES. 1: Retention Rates of NHSC Participants and Non-Participants—Primary Care

Figure ES. 2: Retention Rates of NHSC Participants and Non-Participants—Mental Health

Similar to primary care HPSAs, non-participants in mental health HPSAs were much more likely to stay in the same HPSA than participants (Figure ES.2). Their retention rate declined by 3-4 percentage points each year since the start year, while the retention rate in any HPSA declined at a lower rate, about 2-3 percentage points. Also, the retention rates in any mental health HPSA was very similar across participants and non-participants, especially in the further out years.

Next, we estimated multivariate regression models at the provider level in which we modeled the ‘same HPSA’ and ‘any HPSA’ outcomes as a function of participation in NHSC programs, a set of individual-level characteristics (age, gender, provider type), Census division indicator variables and local area characteristics at the zip code level (average family income, poverty rate, percent White, percent Black, fraction of the population over 25 years of age with a high school degree and percent of the population over the age of 65).

As shown in Figure ES.3, in the first year since separation/start year, NHSC participants are 37.0% less likely to remain in the same HPSA relative to non-participants. This difference was obtained by netting out the impact of other (observable) individual socio-demographic and local area characteristics. Given that the unadjusted difference in the retention rate in the same HPSA in the first separation/start year is 42.8 percentage points (=83.5-40.7, from Figure ES.1), it follows that 86.4% (=37/42.8) of the observed difference in primary care ‘same HPSA’ retention is explained by NHSC participation. A similar fraction, 85.6% (=13.7/16.0), in the observed retention difference was explained by NHSC participation in the case of primary care ‘any HPSA’ in the first separation/start year. The other ratios between adjusted and unadjusted retention differences in retention between participants and non-participants remained similar for the other further out separation/start years, for both primary care ‘same HPSA’ and ‘any HPSA’ measures.

Figure ES. 3: Differences in the Participants’ Retention Probability Relative to Non-Participants—Primary Care

Figure ES.4 presents the regression-adjusted retention differentials by NHSC participation for mental health HPSAs. The retention differentials are lower across the board for the ‘same HPSA’ measure than in the case of primary care HPSAs. As shown in Figure ES.2, the unadjusted retention in ‘any HPSA’ was higher for non-participants in the first separation/start years than that of participants. Nonetheless, after accounting for individual-level and local area characteristics, there was no statistically significant difference between the retention of participants and non-participants in mental health HPSAs for any of the separation/start years.

Figure ES. 4: Differences in the Participants’ Retention Probability Relative to Non-Participants—Mental Health

Other findings from the regression analyses include the following. First, HPSA retention rises with age and local characteristics, but differences by gender, discipline, and Census division are small. Second, as reflected by regression estimates showing that providers have higher retention in poorer and less educated communities, providers select into HPSAs based on their preferences for serving underserved populations.

Introduction

The National Health Service Corps (NHSC) is administered by the Bureau of Health Workforce (BHW) in the Health Resources and Services Administration (HRSA). The NHSC was originally designed to address geographic maldistribution of the health care workforce by increasing the number of health care professionals in areas designated by HRSA to be Health Professional Shortage Areas (HPSAs). Under Assistant Secretary for Planning and Evaluation (ASPE) leadership in coordination with HRSA, The Lewin Group is focusing on two programs within the NHSC: the Loan Repayment Program (LRP) and the Scholarship Program (SP). While the LRP has expanded significantly since 2009, the SP is constrained by budget policy. In FY 2014, individuals participating in LRP could receive up to $30,000 for a 2-year commitment in a HPSA with scores between 0 and 13 and up to $50,000 for a 2-year commitment in a HPSA with scoresof14 or higher. However, the statute allows for up to $35,000 per year (or $70,000 over two years). The Scholarship Program offers funding for a maximum of four years for: school tuition, required fees, education costs and a monthly support stipend. In exchange, SP participants agree to provide primary health care in a HPSA for a number of years that is equal to the number of years they were scholarship recipients.

The NHSC deploys almost 9,000 health care professionals to thousands of sites across the country annually (National Health Services Corps, 2012). In the FY 2015 Department of Health and Human Services (DHHS) budget, President Obama proposes boosting the National Health Services Corps to 15,000 a year over the next five years (DHHS Budget, 2014). These clinicians include primary care physicians, primary care certified nurse practitioners, certified nurse-midwives, primary care physician assistants, dentists, registered dental hygienists, health service psychologists, licensed clinical social workers, psychiatric nurse specialists, marriage and family therapists, and licensed professional counselors. These professionals deliver critical medical, dental, and mental health services in geographic areas, facilities, and populations that have limited access to health care services (Health Resources and Services Administration, 2013). In exchange for a service commitment (typically 2-5 years) the NHSC provides students with medical education incentives through the LRP and SP programs. Upon the conclusion of an initial service obligation period, NHSC providers may apply for additional loan repayment funding in return for further service.

In September 2013, ASPE awarded The Lewin Group a contract to examine short- and long-term retention in high-need areas of providers who participated in the LRP and SP programs and compare their retention with retention of non-participants working in those areas. Important questions for ASPE and HRSA is how many providers who participate in the NHSC loan repayment and scholarship programs remain in high need areas once they have completed their contract obligations and how their retention compares with the retention of providers in high need areas who did not participate in the program. This study addresses these questions using data from the period 2000-2013. Several past studies have addressed these questions, but with now often dated data obtained mostly by small scale surveys that have focused almost exclusively on physicians. In addition to physicians, this study will examine retention of non-physician providers, including nurse practitioners, physician assistants, mental health and dental care clinicians.

We start with a detailed survey of the literature about the NHSC program, including studies of how participant retention in high-needs areas compares with non-participant retention. Next, our empirical approach encompasses various data methods to evaluate the retention of NHSC providers beyond their initial service obligation. We utilize several large-scale administrative databases to track providers over time and compare the NHSC participant retention with the retention of comparable non-NHSC providers in the same HPSAs. In this document we produce detailed statistics on the number of years providers remain in the same HPSA after completion of their initial contract, whether they locate in a different HPSA, whether they remain in the same area but outside of HPSAs, and whether and when they move to other geographical areas (HPSAs or non-HPSAs). These statistics are broken down by provider type, age and other relevant characteristics. Next, we compare these trends with the retention and migration trends of non-NHSC providers that have similar characteristics and who at some point serve in the same HPSAs as the NHSC participants.²

In this study we also specify a formal economic model of individual geographic location decisions and apply that model to the NHSC programs, with an emphasis on the LRP. The model isolates the key factors influencing geographic location decisions, and it explains why some individuals might choose to locate in areas that others avoid. The model also explains when geographic mobility will be high and when it will be low. The general model of location decisions is modified to account for the essential features of the NHSC program. In particular, we show that the retention of participants in high-need areas after they complete their obligations depends crucially on the way that NHSC selects participants into the program.

To measure this relationship between enrollment, retention and monetary value of the NHSC programs we conduct econometric analyses to estimate the effect of LRP and SP on enrollment in the NHSC workforce and retention of participants in health care shortage areas. In these analyses we attempt to control for individual socio-demographic characteristics as well as for the multiple unobservable factors that are associated with the individual provider’s decisions to: enroll in NHSC; continue with NHSC under a new contract; stay in the same location after NHSC service completion; move to another HPSA after service completion; or move to a non-HPSA after service completion. Some of these unobservable characteristics are the preference for serving in rural/underserved areas, financial constraints or factors that may make a provider more likely to prefer to work in a certain location regardless of participation in NHSC programs (e.g., being close to one’s place of birth).

Although NHSC continues to achieve gains in the recruitment and retention of clinicians, further progress may be desirable. Today, 21 percent of the nation’s population resides in 5,800 Federally-designated shortage areas, and this proportion is climbing (NCSL, 2011; HRSA, 2013). Shortages of primary care services in underserved areas are likely to be further exacerbated, with the Affordable Care Act (ACA) coverage expansion expected to affect underserved areas disproportionally. While rural communities do not necessarily equate to underserved areas, the discrepancy between clinician need and supply in rural areas is sizeable—nearly 21 percent of the US population resides in rural areas, but only 9 percent of the physician workforce practices in such settings (Health Resources and Services Administration, 2011(a)).

A detailed survey of the existing literature relating to the NHSC scholarship and loan repayment programs is contained in Appendix A. The main body of the report is organized as follows. Chapter II provides a detailed discussion of our data sources and main measures. Chapter III presents summary statistics obtained with these data sources, including measures of the retention and geographic mobility of NHSC participants and non-participants. Chapter IV develops an economic model of location choices and uses that model to derive predictions about how the retention of NHSC participants in high-need areas will compare with the retention of non-participants. Chapter V uses the data to derive estimates of retention differences between participants and non-participants that control for a host of observable factors, including the provider’s medical discipline, age and gender as well as other factors such as local area median income and percentage of the population in poverty. Finally, Chapter VI simulates the economic model constructed in Chapter IV under different assumptions about the key parameters in the model. The empirical findings are interpreted in light of these simulations. Chapter VII concludes the report.

Data Sources and Main Measures

To accomplish the goals of this study, we constructed two provider-level analytic datasets based on administrative datasets and other publicly available and proprietary datasets. We describe each data source in detail below and then we discuss the construction of the two analytic datasets.

Data Sets

NHSC Administrative File

The NHSC administrative file provided to us by HRSA represents a panel of 22,703 participants who entered the NHSC programs over the period 2000-2013. The NHSC administrative file contains information on where each participant was located each year during the participant’s program (down to the zip code level). Participants are tracked annually from their entry year until 2013, but only while they are in the program. Of the total number of providers, 10,123 are physicians, 6,850 are nurse practitioners (NP) or physician assistants (PA) and the remaining 5,730 individuals represent other providers, such as dental or behavioral health providers. This database contains information on participant demographics and award/service characteristics, including: age, gender, race/ethnicity, award year, entry year, type of award, length of initial service obligation, funds received, practice type, provider type, location, and separation year. Table II.1 provides a breakdown of the number of NHSC enrollees by entry year, provider type, gender, program type, average age on entry, and race/ethnicity.

Table II. 1: Select Socio-Demographic Characteristics of NHSC Enrollees by Entry Year

The number of new NHSC participants has increased substantially over the last few years, with most of this increase reflected in an expansion in the number of women, non-physicians and White providers. In Appendix Table B.1 we present the distribution of the NHSC workforce by provider discipline.

In Table II.2 we present the number of records on NHSC providers by entry cohort.

Table II. 2: Number of Records on NHSC Providers by Entry Cohort

The main disadvantage of the NHSC administrative data file is that it does not include information on where participants were located after program completion. Also, it does not contain the participant’s National Provider Identification (NPI) number, or other unique identifiers which would enable us to track their location after program completion. We therefore rely on additional data sources to identify the participants’ NPI and then subsequently determine where they were located after program completion. These four data sources, which are also used to identify non-participant providers, include:

• AMA Physician Master Files of various years since 2000;

• Medicare Providers File developed from Medicare claims data over the period 2005-2011; and

• Optum Corporation’s Provider 360 File, a proprietary file containing comprehensive information on most medical providers in the United States.

Finally, we employ data from HRSA with information on HPSA designations, types, disciplines, sites and scores. We use the most recent file of these data, which was compiled in December 31, 2013.

Below we discuss these files in turn and highlight their advantages and limitations for the current project.

AMA Physician Master Files

The AMA Master file was begun in 1986 and has been tracking physicians since then. Multiple years of AMA data were extracted to obtain information about provider location over multiple years. Each year of the AMA data file contains current information on the physicians’ locations and practices, along with information on their training. The AMA is the single most comprehensive data source on physicians practicing in the US. In theory, linking NHSC administrative data with the AMA data, it is possible to track the practice location and retention information of NHSC enrollees over the years. Each observation includes location data (such as zip code), medical school, specialty, graduation year, birth date, race, and gender. We make use of complete AMA Physician Master Files for the following years: 2000, 2007, 2008, 2010-2012, and partial files (containing only data on primary care providers) for 2003 and 2005. The number of records increased steadily from about 1.0 million physicians in 2000 to about 1.3 million in the 2011 file. However, the main limitations of these files are that they include only physicians and that the providers’ location is updated infrequently, at intervals that are often longer than one year.

Provider360

Provider360 (P360) is a comprehensive provider database developed and maintained by OptumInsight, the Lewin Group’s parent company. Optum first developed Provider360 in the early 2000s and since then has updated it on a monthly basis. Optum links numerous private and public databases to create P360, and it gathers information about each provider’s demographics, education and training, NPI and DEA numbers, and location. Importantly, Provider360 contains both physician and non-physician providers, including PAs, NPs, and mental health and dental care clinicians.

This dataset covers virtually all currently active providers and includes most of the socio-demographic variables that we observe in the NHSC data, like provider demographics or provider type, as well as additional useful variables, like the provider’s NPI, medical education and practice affiliation. The large number of common variables appearing both in the NHSC files and P360, like name, birthdate, gender and others, allows for a link between the two datasets with a good match rate.

The main limitation of P360 is that it was not designed to provide panel information about providers, so we can tell where providers are located now but not each year in the past. However, even without historical P360 data retention patterns of NHSC providers can be constructed by comparing their locations upon NHSC program completion with their current locations in P360. For instance, we can look at the cohort of students finishing their medical training in, say, 2006 and determine how many of them are still serving in HPSAs as of the current year. We return to this point in the next section of this chapter. Other useful variables available in P360 include the provider’s degree, specialty, medical school attended, residency institution, license, DEA number, hospital affiliation, practice type, practice NPI, and sanctions against the provider.

Medicare Providers File

In order to track participants and non-participants over time, we constructed a provider level dataset using the Medicare claims of providers who billed Medicare between 2005 and 2011. In this dataset the unit of observation is the provider-year, meaning that each Medicare provider is observed annually along with his or her geographical location at the zip code level. The provider’s location in a given year is based on the most frequent zip code associated with that provider’s claims during that year. The dataset includes the provider’s NPI, which allows us to identify NHSC participants and track their location in the years after their service completion.

The Medicare provider file contains 5,757,405 observations on 1,099,836 unique providers. The breakdown of unique providers by type is: 626,836 physicians, 123,223 NP/PA’s, 54,168 DO’s and 295,615 other providers.³ Table II.3 shows the number of providers by type in each year.

Table II. 3: Number of Providers by Type in Each Year (Medicare Provider Data)

As in the case of another provider data, available to the Lewin team known as the de-Identified Normative Health Information (dNHI) data, Medicare providers also appear to be well distributed across all states during the period between 2005 and 2011 (Appendix Table B.2).⁴

Data on HPSA Designations and HPSA Scores

HRSA also provided Lewin with a detailed file containing information on all HPSA sites as of December 2013. As of December 2013, there were 5,976 primary care HPSAs, 4,758 dental care HPSAs and 3,876 mental health HPSAs. This file includes information on each HPSA: ID, name, status, type, discipline, HPSA score and detailed geographic identifiers. In Table II.4 below we present the number of HPSA sites by HPSA disciplines (i.e., primary care, dental health and mental health) and by the main HPSA types (facility HPSAs, single county HPSAs, Census tract HPSAs and minor civil division HPSAs).⁵

Table II. 4: Distribution of HPSA Sites by Disciplines and Types

Construction of Analytic Datasets

The first analytic dataset was constructed using NHSC data, P360 data and Medicare provider data. Using the name, birthdate, gender and a number of other variables, we matched the NHSC administrative file with the P360 data. This way we were able to uniquely identify about 18,500 of the 22,703 NHSC participants in the P360.⁶ This match gives us the participants’ NPIs as well as other important information including their location’s zip code in 2013. Almost all of these matches were associated with a valid NPI (i.e., for about 17,900 NHSC participants). The matches without an NPI correspond to NHSC participants that do not have an NPI.

The next step was to link the merged P360-NHSC file with the Medicare Provider file by NPI. The analytic file contains annual information on over 1 million non-NHSC providers as well as on 8,973 NHSC participants (out of the 17,900 participants with a valid NPI). The dataset allows us to track the annual locations at the zip code level for each year individuals are in the Medicare Provider data (i.e., over the timeframe between 2005 and 2011). Figure II.1 below summarizes the main steps we took to create the first analytic dataset.

Figure II. 1: Steps to Create the First Analytic Dataset on NHSC Providers

As our chosen level of geography was the zip code, the next step was to determine whether the zip code associated with each provider’s location was part of a HPSA or not. For this purpose we used the dataset we received from HRSA containing information on all HPSAs and constructed an algorithm to determine whether a provider’s zip code was part of a single-county HPSA, Census tract HPSA, Census division HPSA, or a facility HPSA. Some Census tract and Census division HPSAs may straddle two or more zip codes, in which case we determined that if more than half of the zip codes’ population was within the bounds of a Census tract or Census minor civil division HPSA, those zip codes would be flagged as HPSAs. Also, we flagged the HPSA zip codes as primary care, mental health and dental HPSAs, using HRSA’s classification of HPSAs by type. An inherent limitation of this approach is that we only used data on HPSAs as of December 2013. To the extent various areas lose or gain HPSA status over our period of analysis, we may erroneously place providers in HPSAs (or non-HPSAs) in the years prior to 2013.

To ensure a consistent classification of NHSC providers and non-NHSC providers by their medical discipline, we mapped the detailed health care occupation (available in P360) of each non-NHSC participant into the corresponding NHSC discipline. The list of NHSC disciplines is presented in Table II.5 along with the distribution of providers by these disciplines in the overall population of NHSC participants and in the sample of participants we identified in the first analytic dataset. Table II.5 also provides the average age of providers (at entry in NHSC service) and the distribution by gender and HPSA type.

Table II. 5: Comparison between NHSC Providers in the Overall Population and in the First Analytic Dataset

We completed the construction of the first analytic dataset by eliminating the following groups of non-NHSC participant providers: (i) those who did not fall under any one of the NHSC discipline types; (ii) those who did not serve in HPSAs; (iii) those who served in HPSAs where no NHSC participant served over the 2005-2011 period; and (iv) specialists (cardiologists, dermatologists etc.). The purpose of these data restrictions was to ensure a degree of comparability between NHSC participants and non-NHSC participants who served in HPSAs. We ended up with a total of 202,999 non-participant providers serving in primary care HPSAs and a total of 19,304 non-participants serving in mental health HPSAs. One limitation of the analytic dataset is that it has very few dentists and pediatricians, since these providers are unlikely to file Medicare claims and thus appear in the Medicare data. The main characteristics of non-participants are presented in Table II.6.

Table II. 6: Non-NHSC Providers Serving in Primary Care and Mental Health HPSAs in the First Analytic Dataset

In Table B.3 of Appendix B we present the distribution of providers serving in NHSC in each state over the timeframe of the first analytic data.

Finally, in the case of providers who changed their location from one year to the next we constructed variables measuring the distance of the move by calculating the distance in miles between the zip code centroid of the initial location and the zip code centroid of the next year’s location.

The main advantage of the first analytic dataset stems from the fact that it allows us to track the location of providers after they complete their NHSC service. However, the number of NHSC providers is potentially limited. We therefore constructed a second analytic dataset relying only on NHSC data, P360 data and HPSA data. The main feature of this dataset is that provides information on the NHSC providers’ location in two points in time: the year of program termination and December 2013 (i.e., the time when P360 information is recorded).

The number of NHSC participants in this second dataset increases to about 18,500 (of which 17,983 have a valid NPI). As expected, the distribution of NHSC providers by discipline resembles the distribution from the overall NHSC population more closely (Table II.7). Also, the other characteristics of participants from the second dataset are more similar to the characteristics from overall population of participants. In the next chapter we compare the retention patterns of participants from the first dataset with the retention patterns of participants from the second dataset.

Table II. 7: Comparison between NHSC Providers in the Overall Population and in the Second Analytic Dataset

NOTE: The total number of matched NHSC providers presented in the above table is limited to providers with a valid NPI.

Figure II.2 summarizes the main steps needed to construct the second analytic dataset.

Figure II. 2: Steps to Create the Second Analytic Dataset on NHSC Providers

Retention Measures

Using the first analytic dataset we constructed four retention measures at the provider level:

• Serving in the same HPSA and in the same county. This variable takes the value of 1 if the NHSC provider remains in the same county as the one where he or she served in NHSC, and 0 otherwise.

• Serving in a HPSA in another county. This variable takes the value of 1 if the NHSC provider remains in a HPSA that is located in a different county than the one in which he or she served while in NHSC service, and 0 otherwise.

• Serving in a non-HPSA from the same county. This variable takes the value of 1 if the NHSC provider moves to a non-HPSA area from the county where he or she served while in NHSC, and 0 otherwise.

• Serving in a non-HPSA in another county. This variable takes the value of 1 if the NHSC provider moves to a non-HPSA area from another county than the county he or she served while in NHSC, and 0 otherwise.

To ensure comparability across all providers, we defined these measures for non-participants as well. While in the case of participants, the reference point in time was the end of the NHSC service, for non-participants we chose the first year (“start year”) that the non-participants appear in Medicare data as their reference point-in-time. Calculating these measures for each cohort, we were able to construct retention rates one year after separation from NHSC, two years after separation from NHSC and so on. In the case of non-participants the annual retention rates were calculated as one year since start year, two years since start year and so on for each cohort.

Geographic Mobility of Participants and Non-Participants

Retention Measures Using Longitudinal Data (First Analytic Dataset)

In Figure III.1 we start with retention rates in primary care HPSAs by years elapsed since separation from service (for participants) and by years since start year (for non-participants).

Using data from the first analytic dataset, we find that about 82% of the NHSC participants serve in primary care HPSAs one year after completion of their NHSC service. More than half of participants who are still in primary care HPSAs one year after separation are actually in the same county as the one in which they served while in service (i.e., 49% of participants). There is a fairly steep decline in the retention rate in years 2 and 3 after separation (by about 7 and 3 percentage points, respectively), followed by a leveling off thereafter.

Focusing on non-participants, we note that their retention rates in primary care HPSAs are always higher than the retention rates of participants, both in terms of retention in the same HPSA as well as in terms of retention in a different HPSA. Their retention also drops after we first observe them in the data (their ‘start’ year), but in contrast with participants the move rate out of HPSAs is relatively constant over time. One year after we first observe non-participants in HPSAs, 95% of them are still in HPSAs. Moreover, a very large fraction of them (91% of all non-participants) remain in the same HPSAs where they were first observed. The retention rates in any HPSAs decline by about 3-4 percentage points every year thereafter, while the retention rates in the same HPSA decline at a slightly faster rate (about 5-6 percentage points per year). These rates indicate that once non-participant providers serve in a HPSA, they tend to remain in those areas, and to some extent they migrate from one HPSA to another.

Figure III. 1: Retention Rates of NHSC Participants and Non-Participants—Primary Care

It is important to note in Figure V.1 that the participants’ retention appears to increase after more than 6 years, while the non-participants’ retention declines at a higher rate than in the previous years. However, these findings should be viewed with caution, because the retention rates for more than 6 years are constructed using only P360 data for 2013, as the timeframe afforded by the Medicare data is only six years (2005-2011).⁸

In Figure III.2 we present the retention rates in mental health HPSAs. The retention rates of participants are lower in mental health HPSAs than in the case of primary care HPSAs. Also, the decline in retention rates (or the move rate out of HPSAs) is much lower than for primary care providers. For instance, the fraction of participants serving in the same HPSA as during the program is about the same in the first two years since separation (37-38%), declines to 36% and then remains relatively constant at 32-33% thereafter. The retention rate of participants in any mental health HPSA is relatively constant over the years, hovering around 64% and 68%.

Similar to primary care HPSAs, non-participants in mental health HPSAs are much more likely to stay in the same HPSA than participants. Their retention rate declines by 3-4 percentage points each year since the start year, while the retention rate in any HPSA declines at a lower rate, about 2-3 percentage points. It is important to note that the retention rate in any mental health HPSA is very similar across participants and non-participants, especially in the further out years.

Figure III. 2: Retention Rates of NHSC Participants and Non-Participants—
Mental Health

It is important to note that we dropped from the retention analyses in Figures III.1 and III.2 the NHSC participants who left service in 2013, the last year of our timeframe. Of the initial sample of 8,973 participants we identified in the first analytic dataset, we ended up using a number of 6,296 participants, of which 4,995 are primary care providers and 1,301 are mental health providers.

Retention Measures Using the Second Analytic Dataset

So far, we have discussed retention statistics using only the first analytic dataset, which is based on NHSC, Medicare and P360 data. While this analytic dataset is very useful in that it tracks the participants’ and non-participants’ location over time, the number of participants identified in this dataset may be arguably viewed as small (around 9,000 providers of the total of 22,703). We address this issue by presenting in Tables III.1-III.3 the retention rates of NHSC participants using the second analytic dataset. However, despite the fact that we observe about 18,500 participants in that dataset, we can only construct retention rates as of December 2013, the time when the P360 data was recorded. As in the case of the first analytic dataset, we dropped participants who appear in our data only in 2013, and thus, of the initial 18,500 participants we matched with P360 data, we ended up using a number of 11,210 participants for the retention analyses we present below.

Table III. 1: Retention Rates of NHSC Participants as of December 2013—Primary Care

The rates in Tables III.1-III.3 indicate the fraction of participants who remain in the same HPSA or any HPSA between the time of service completion and December 2013. As a result, the fraction remaining in HPSAs for the cohort of providers exiting NHSC in 2012 represents a 1-year retention rate, for those exiting in 2011 it is a 2-year rate and so on. Although the rates in Table V.1 are not directly comparable to the rates plotted in Figure V.1, they nonetheless paint a similar picture regarding the retention of participants in HPSAs after service completion. For instance, the one-year retention rate in same HPSA same county of primary care participants who left service in 2012 was 43.9% (Table III.1), compared to 48.5% in Figure III.1. Similarly, the two-year retention rate was 41.0% (Table III.1), while the two-year retention rate was 4.07% in Figure III.1.

Given that the retention rates in Tables III.1-III.3 are inherently cohort-specific, we constructed cohort-specific retention rates using the first analytic dataset (Tables B.4 and B.5 in Appendix B). Tables A.4 and A.5 allow for a direct comparison between the retention rates from the first analytic dataset and the retention rates from the second analytic dataset (in Tables III.1 and III.2). As the retention rates from the first analytic dataset are virtually indistinguishable from the retention rates from the second analytic dataset, we conclude that the first dataset provides a comprehensive and representative picture of the entire sample of NHSC participants.

Table III. 2: Retention Rates of NHSC Participants as of December 2013—
Mental Health

Another advantage of the second analytic dataset was that it allowed us to construct retention rates for providers in Dental Health HPSAs. These rates could not have been constructed with data from the first analytic dataset, as the sample size of dental health providers billing Medicare was extremely limited. As can be noticed in Table III.3, the retention rates of dental health participants were similar to the retention rates of providers in primary care and mental health HPSAs.

Table III. 3: Retention Rates of NHSC Participants as of December 2013—
Dental Health

Finally, Tables B.6 and B.7 in Appendix B show no substantial variation in the retention rates of providers by provider type. Primary care participants are slightly more likely to stay in the same HPSA same county and less likely to stay in HPSAs in other counties than NP/PAs, while mental health physicians are in general less likely than NP/PAs to remain in HPSAs. We will return to this issue in Chapter V to determine whether differences in retention rates remain after we adjust for the provider’s age, gender and other individual level and local area characteristics.

Distances of Providers’ Moves

As described in the previous chapter, we constructed a variable measuring the distance of the provider’s move from one location to another by calculating the distance (in miles) between the zip code centroids of the initial and final location. In Table III.4, we present the mean distance of moves by the four retention metrics and by years since separation from NHSC service.

Table III. 4: Average Distance (in Miles) between NHSC Service Location and Current Location, by Retention Metrics and Years since Separation from NHSC

NHSC participants who remained in the same HPSA same county were, not surprisingly, located on average within 3 to 5 miles from their original NHSC service location, while those who moved to a non-HPSA in the same county were located on average within 5 to 8 miles from their initial NHSC location. Providers who moved to a HPSA outside the county where they served while in NHSC tended to be located further and further away as time since separation went by. For instance, primary care providers moved on average 229 miles away within one year since NHSC separation, 371 miles 4 years since separation and 401 miles 6 years since separation. This trend was less pronounced for mental health providers, who moved within shorter distances to HPSAs in other counties. Finally, all providers going to non-HPSAs in other counties had moves at greater distances than the providers who moved to HPSAs in other counties.

Retention of Participants by Place of NHSC Service

We also analyzed the retention patterns of NHSC providers by the place of service to understand whether the institution type they work in is associated with their propensity to remain in HPSAs after service completion.

Since about half of all NHSC participants work in FQHCs, we investigated the retention of providers by whether they were working in an FQHC at the time of NHSC service completion. Using the second analytic dataset, we found that about 88% of the 8,760 participants who worked in FQHCs were still in the same zip code as of December 2013. In contrast, participants who were in non-FQHCs while in service are substantially less likely to be in the same location at the end of our observation period.

Table III. 5: Comparison of Retention Trends by NHSC Place of Service (FQHC vs non-FQHC locations)

Economic Model of Location Choices with Application to NHSC Programs

The retention patterns we observed in Figures III.1 and III.2 appear to indicate that participants have lower HPSA retention than non-participants. There may be a suite of factors in addition to NHSC participation that influence the providers’ decisions to first locate in a HPSA and then move out of a HPSA. In this chapter we provide a theoretical model to provide insights into the impact of these additional factors on the observed retention patterns.

Specifically, in this model we aim to: (i) isolate the key factors influencing providers’ location decisions; (ii) explain why some providers locate in areas that others avoid; and (iii) explain when geographic mobility is high or low. We start with a general location choice model and then expand the model to incorporate the incentives offered by the NHSC programs. A technical version of this model is available in Appendix B.

An Economic Model of Location Choice

In general, in any given time period an individual calculates the value (or utility) of each possible location and chooses the location offering the highest value. The value of each location depends on three main factors:

1. The value that the individual i places on the non-pecuniary factors associated with living in location j (climate, environment, local amenities, spousal employment opportunities, etc.), which is assumed to be time-invariant (denoted by the symbol ).

2. The expected present value of money wages if the individual chooses location j in period t. This expected present value is the sum of:

• the wage available in location j in period t, (); and

• the discounted value of expected future utility if the individual chooses location j in period t (ρΕ(V^t+1)(where ρ is a one-period discount factor. Expected future utility depends on the value of all future wages in all possible locations.⁹

3. Finally, a completely random location shock that is unrelated to the individual’s preference for location j in any given period t (denoted by the symbol ). This random shock accounts for unobservable factors that might induce an individual to choose a location she might dislike in period t, or leave a location she likes in period t.

Mathematically, the utility of location j at time t can be written as

                (1)

In this model, an individual will choose location j if its utility () exceeds the utilities associated with all other possible locations. Clearly, an individual who has strong non-pecuniary preferences for a particular location is more likely to choose it over other locations. That is to say, the probability of choosing to locate in location j initially, or remaining in location j if the individual is already there, increases with . But dislike for a particular location can be overcome if wages in that location are high enough. That is to say, even if an individual does not like a particular location as given by a negative value of , she may still choose to locate there if the pecuniary advantage of locating in the area, as measured by the value of the current wage plus the expected present value of future wages, is high enough. Given the values of the pecuniary and non-pecuniary factors associated with different locations, an individual’s propensity to move from one location to another is governed by the size of the location-specific random shocks. If wages were stable and random shocks did not exist, an individual would select his or her best (i.e., utility-maximizing) location in the first period and remain there forever.

Consider now aggregate (population average) probabilities of choosing a particular location and the aggregate probabilities of remaining in that location. These average probabilities are simply weighted averages of individual probabilities of selecting a location or remaining in it. The weights on which the aggregate averages are based are the fractions of the population with different values of . For example, if there were 5 different values of in the population and each value occurred with equal frequency, each value would receive a 1/5 weight in the calculation of aggregate probabilities. In general, the aggregate probabilities depend on the frequency distribution (probability density) of preferences (the ) in the population as well as the frequency distribution of the random shocks (the ). The parameters of these distributions (means and standard deviations) affect the aggregate probabilities and their sensitivities to changes in wages. We may show that, all else constant:

• a smaller standard deviation of the random shock (denoted σ_ε reduces the probability of an individual move from location j and increases the expected number of periods an individual stays in the initial location j;

• the smaller is σ_ε, the smaller is the frequency of moves in a cohort of individuals;

• a smaller average preference for location  j (denoted ) results in a smaller fraction of individuals choosing a location or remaining in it;

• higher current or future pay in location j increases the fraction of the population choosing to locate there and remain in it;

• a larger standard deviation of in the population (denoted σ_θ) decreases the impact of pay changes.

Stated alternatively, the last proposition says that the more heterogeneous people are in their preferences for different locations, the less influence wage changes will have on their location choices. Conversely, if all individuals placed the same non-pecuniary value on each location, there exists a single set of wages across locations that would make individuals indifferent among locations. In other words, supposing location-specific random shocks are zero, wages would be the most important determinant of location choices. If wages were insufficiently high in locations with low non-pecuniaries, no one would choose those locations. Heterogeneous preferences ensure that most, if not all, locations will attract or retain some people, even when the average value preferences for those locations (i.e., their μ_j) are low or when wages are low.

Location Decisions in the Presence of the NHSC LRP

Consider now NHSC’s loan repayment program (LRP). A unique feature of this program is that an individual who applies for the LRP must have an NHSC-approved job in a HPSA and also have outstanding student debt in order to qualify for the program. Applicants are screened by the NHSC and not all applicants are accepted into the program. Importantly, in deciding whether to approve an applicant for the program, NHSC makes a determination regarding the applicant’s fit for the program and for the position the individual accepted. NHSC strives to choose the ‘most qualified’ applicants, but during the acceptance process it gives weight to an applicant’s fit for a particular position. An implication is that NHSC may select an applicant judged to be a ‘good fit’ over other applicants with better academic records. In addition to individual qualifications, the main driver of acceptance into the program is the severity of the shortage of health care providers in a particular area as measured by the HPSA score. Individuals applying for approval of a position in an area with a high HPSA score may have a better chance of approval than individuals applying in an area with a low HPSA score. Prior to 2009, approved individuals received funding only if the HPSA had a score of 14 or above. After the expansion in 2009, all approved individuals were funded regardless of HPSA score. Because of program funding constraints, not all applicants receive approval, even when applying for positions in high HPSA score areas.¹⁰

If an individual is accepted into the program and qualifies for the loan repayment amount L_j, the utility of location j is given by

       (2)

The individual prefers to participate in the program if there exists at least one HPSA location j for which the utility associated with that location is higher than the utilities associated with any other location. If the location that maximizes utility is not a HPSA, the individual of course does not apply. The attractiveness of a given HPSA depends on the loan amount L_j, so that the probability of choosing location j increases with the amount L_j. Some providers who do not participate in NHSC may still locate in a HPSA. This group will include: (1) individuals without student debt (and therefore ineligible to apply for NHSC program); (2) individuals who have student debt but did not apply (perhaps due to a low expected probability of approval, or lack of knowledge of the program); and (3) individuals who did apply and were not accepted.

^10 In fiscal years 2012 and 2013 the admission rate into NHSC programs was around one third of applicants in those years.

LRP Participation

LRP participation is a joint outcome of application and acceptance. But this process is unobservable, and we only observe the outcomes of participation and non-participation. We denote the unobservable factors related to admission in the population of applicants with γ_i, where i represents the individual program participant. The term γ_i can be viewed as the individual’s ‘fit’ for the program. Fit for the program depends in part on observable factors such as academic background. Fit may also be related to the strength of an individual’s preference for a particular NHSC location (i.e., θ_j). We assume that NHSC ranks applicants in order of fit for the program and then fills all available spaces in the program. Rejection occurs when there are more applicants than spaces.

Exactly who is selected into the program depends on the weight NHSC places on factors other than preferences (e.g., academic background) during the selection process and the weight it gives to preferences. If fit for the program were based only on preferences, preferences would receive all of the weight and the rank-order of fit for the program would be identical to the rank-order of individuals’ location preferences. In this case, preferences and fit would be perfectly correlated and the program would select the same individuals who would have the strongest preferences for service in high-need areas in the absence of the program. At the other extreme, if preferences received no weight in the selection process and other factors such as academic ability received all of the weight, preferences and fit would have no correlation. In this case, the program would tend to attract the highest number of individuals who would be unlikely to serve in an underserved area in the absence of the program.

Retention

For simplicity, assume there are only two location types: HPSAs and non-HPSAs, and each individual has a certain preference for locating in HPSAs, θ_i. The correlation between the unobservable factors that affect admission to the LRP program and choice of geographic location (i.e., the correlation between γ_i and θ_i) has implications for observed retention patterns of NHSC participants and non-participants.

Under normal circumstances, there will be a number of participants with negative values for θ_i who will locate in HPSAs only because of the program. As a result of these providers going to HPSAs, the average preference of participants is lower than the average preference of non-participants and therefore the retention of participants in HPSAs will always be lower than the retention of non-participants. The difference between the retention rates is larger when the correlation between γ_i and θ_i is lower. As discussed above, the limiting case is when the correlation is zero, meaning that preferences for HPSAs play no role in the acceptance process. In this case, the number of NHSC participants with negative preferences for HPSAs will be the highest and therefore, the average retention rate in the population of participants will be at its lowest point relative to any other scenario when the correlation between γ_i and θ_i is strictly positive.¹¹

In the unlikely case when there are more individuals with positive values for θ_i than the number of available program positions and the correlation between γ_i and θ_i is positive, all program participants will be individuals who have a preference for locating in HPSAs. Such individuals will have a higher average preference for HPSAs than program non-participants who were attracted to HPSAs anyway. Because they have a higher average preference for HPSAs than non-participants, program participants will tend to have higher retention in HPSAs after completion of their service obligations than non-participants. However, given that not all candidates are accepted into NHSC, this case cannot be encountered in reality.

Finally, the magnitude of the program effect on HPSA provider supply also depends on how sensitive location choices are to the presence of the program and the LRP amount. As in the general case, the impact depends on the mean preferences for various locations (the μ_j), the standard deviation of preferences (σ_θ), and the standard deviation of the random shocks (σ_ε). Choices are more sensitive the more homogeneous location preferences are (i.e., the smaller is σ_θ) and the smaller the role of random shocks to location decisions (i.e., the smaller isσ_ε).

^11 A case of negative correlation between and would mean that the NHSC purposefully selects into the program providers who do not want to there.

Summary

The theoretical model gives a sharp prediction – retention of NHSC participants in HPSAs after the completion of their obligations can never be higher than the retention of providers who choose to locate in those areas without participating in the program. In the limiting case where selection into the program is based solely on preferences, the program selects individuals who would have tended to serve in high-need areas in the absence of the program and retention differences between participants and non-participants are nil. As the correlation between location preferences and program fit weakens, the program tends to select individuals who would not have gone to high-need areas in the absence of the program. In this case, participant retention after program completion will tend to be lower than non-participant retention. Somewhat counterintuitively, a lower retention rate for non-participants is a signal of the program’s success in attracting to high-need areas providers who would not have located there in the absence of the program.

Determinants of Provider Retention in HPSAs

One of our main goals in this study was to estimate the impact of the NHSC programs on HPSA retention.  In this chapter we present the main findings from our empirical analysis of providers’ retention in HPSAs.

Basic Econometric Model

For this purpose, we used the data on NHSC participants and the data on non-participants from the first analytic dataset to estimate regression models in which we control for observable characteristics on each individual (Χᵢ), local area characteristics (Ζᵢ) and an indicator for program participation (Prog i)

                     (1)

We estimated model (1) by using the ‘same HPSA’ and the ‘any HPSA’ indicator variables as the dependent variable, respectively.  The coefficient of interest, α, indicates the impact of NHSC program participation on the number of years served in a HPSA. The X vector included individual-level characteristics like age, gender and provider type, while the Z vector included Census division indicator variables and local area characteristics (at the zip code level), such as: the family income, poverty rate, percent White, percent Black, fraction of the population over 25 years of age with a high school degree and percent of the population over the age of 65.  These variables helped control for factors that retain or induce providers to leave from their initial place of service.

Main Results

We estimated model (1) using a logit regression specification, separately for primary care and mental health HPSAs, and by using the ‘same HPSA’ and the ‘any HPSA’ indicators as dependent variables.  The coefficient estimates are shown in Tables V.1-V.4.  In each of these tables we present 7 models, each estimated by the number of years elapsed since separation from service (for participants) and by the number of years since start year (for non-participants).

In all models of Table V.1 we estimated a lower probability of remaining in the same primary care HPSA for participants relative to non-participants.  Female providers were slightly less likely to remain in HPSAs, and older providers were more likely to remain in the same HPSA over time.  We detected virtually no differences in retention by provider type and, with the exception of the first two years since separation/start year, found no differences in retention by Census divisions.  Providers leaving service or appearing for the first time in the data in the earlier years (2005 or 2006) had a higher probability of remaining in the same HPSA relative to those whose separation/start year was beyond 2008.

It is important to note that the coefficient estimates on the local area characteristics are in many cases statistically significant, indicating that providers are more likely to remain in HPSAs where the poverty rate is higher, the fraction of older population is higher and the ratio of individuals over the age of 25 with a high school degree is lower.  Estimating the models separately on the population of participant providers and then on the population of non-participants, we found that the size of these effects are somewhat larger for non-participants than for participants.

Variable

1 Year

2 Years

3 Years

4 Years

5 Years

6 Years

> 6 Years

NHSC Participant

-2.253***

-1.531***

-1.326***

-1.212***

-1.080***

-1.083***

-0.868***

(0.081)

(0.071)

(0.077)

(0.081)

(0.091)

(0.108)

(0.093)

Female

0.010

-0.044***

-0.039**

-0.076***

-0.066***

-0.081***

-0.048**

(0.021)

(0.017)

(0.016)

(0.018)

(0.018)

(0.018)

(0.019)

Age 36 to 45

0.265***

0.259***

0.290***

0.323***

0.291***

0.249***

0.082

(0.023)

(0.023)

(0.022)

(0.024)

(0.033)

(0.044)

(0.064)

Age 46 to 55

0.581***

0.608***

0.680***

0.762***

0.749***

0.727***

0.447***

(0.033)

(0.030)

(0.030)

(0.030)

(0.037)

(0.048)

(0.064)

Age 56 to 65

0.771***

0.782***

0.898***

0.977***

0.977***

0.981***

0.705***

(0.044)

(0.038)

(0.038)

(0.036)

(0.043)

(0.051)

(0.064)

Age Over 65

1.051***

1.045***

1.189***

1.254***

1.265***

1.279***

0.991***

(0.069)

(0.059)

(0.056)

(0.051)

(0.056)

(0.055)

(0.070)

Medical Doctor

-0.199

-0.202

-0.241*

-0.241

-0.116

-0.331**

0.061

(0.176)

(0.134)

(0.139)

(0.147)

(0.128)

(0.150)

(0.177)

Nurse Practitioner

-0.032

-0.017

-0.043

-0.036

0.027

-0.175

0.253

(0.182)

(0.135)

(0.142)

(0.152)

(0.141)

(0.154)

(0.179)

Physician Assistant

-0.202

-0.180

-0.222

-0.221

-0.075

-0.355**

0.105

(0.188)

(0.139)

(0.147)

(0.156)

(0.151)

(0.159)

(0.196)

Start Year 2005

0.360***

1.038***

0.920***

0.852***

0.811***

0.530***

0.177***

(0.039)

(0.032)

(0.040)

(0.040)

(0.036)

(0.028)

(0.025)

Start Year 2006

0.157***

0.558***

0.520***

0.473***

0.404***

(0.044)

(0.040)

(0.044)

(0.044)

(0.039)

Start Year 2007

-0.091**

0.472***

0.512***

0.464***

(0.042)

(0.035)

(0.046)

(0.040)

Start Year 2008

-0.034

0.392***

0.369***

(0.041)

(0.040)

(0.049)

Middle Atlantic

-0.249

-0.217

0.074

0.068

0.049

0.020

-0.142

(0.225)

(0.225)

(0.244)

(0.258)

(0.258)

(0.253)

(0.234)

East North Central

-0.586***

-0.514***

-0.137

-0.147

-0.095

-0.139

-0.279

(0.145)

(0.141)

(0.190)

(0.193)

(0.191)

(0.190)

(0.184)

West North Central

-0.321*

-0.264

0.105

0.091

-0.004

-0.022

-0.210

(0.166)

(0.171)

(0.212)

(0.218)

(0.224)

(0.223)

(0.216)

South Atlantic

-0.406***

-0.386***

-0.030

-0.026

-0.008

-0.042

-0.057

(0.146)

(0.146)

(0.189)

(0.194)

(0.190)

(0.187)

(0.174)

East South Central

-0.770***

-0.674***

-0.273

-0.249

-0.180

-0.209

-0.252

(0.171)

(0.161)

(0.207)

(0.212)

(0.211)

(0.206)

(0.190)

West South Central

-0.818***

-0.688***

-0.266

-0.266

-0.223

-0.253

-0.216

(0.229)

(0.235)

(0.284)

(0.287)

(0.292)

(0.288)

(0.283)

Mountain

-0.131

-0.129

0.214

0.173

0.174

0.114

0.031

(0.141)

(0.138)

(0.187)

(0.193)

(0.193)

(0.187)

(0.181)

Pacific

-0.269*

-0.247

0.051

0.056

0.050

0.088

-0.032

(0.156)

(0.163)

(0.207)

(0.213)

(0.210)

(0.202)

(0.197)

Log Family Income

-0.095

-0.151

-0.192

-0.235

-0.265

-0.304

-0.648***

(0.220)

(0.224)

(0.227)

(0.235)

(0.231)

(0.225)

(0.219)

Poverty Rate

0.035***

0.033***

0.032***

0.031***

0.030***

0.030***

0.026***

(0.006)

(0.006)

(0.006)

(0.007)

(0.006)

(0.006)

(0.006)

Percent White

-0.299

-0.376

-0.621*

-0.679*

-0.633*

-0.450

0.054

(0.377)

(0.354)

(0.371)

(0.371)

(0.367)

(0.367)

(0.361)

Percent Black

0.559

0.500

0.489

0.435

0.460

0.555

0.887**

(0.385)

(0.357)

(0.374)

(0.368)

(0.370)

(0.368)

(0.369)

Pct HS Grads Over 25 Yrs

-3.886***

-4.140***

-3.899***

-3.401***

-3.850***

-3.744***

-4.858***

(0.935)

(0.958)

(0.994)

(0.982)

(1.007)

(0.981)

(1.037)

Pct Population Over 65

3.218**

3.047**

4.128***

3.914***

3.980***

3.676**

3.353**

(1.280)

(1.279)

(1.438)

(1.513)

(1.540)

(1.432)

(1.387)

Intercept

4.201*

3.706

3.355

3.340

3.675

4.333*

8.132***

(2.375)

(2.407)

(2.419)

(2.554)

(2.499)

(2.396)

(2.390)

Observations

195,189

191,713

178,925

165,474

149,059

141,902

147,725

NOTE: Robust standard errors, clustered by HPSA, in parentheses.  *, ** and ***: significant at 10%, 5% and 1%.  The base group is defined as providers who are: male, age 26-35, non-physician and non-NP/PA, having a start year of 2009 or later, and serving in a location in the Northeast.  The coefficient on any of the dummy variables included in these models shows the estimated difference relative to the corresponding excluded category.

Estimating model (1) with the ‘any HPSA’ indicator as the dependent variable, we found very similar patterns in the retention of providers in primary care HPSAs (Table V.2).   The main difference is that the coefficients on family income variable become statistically significant and negative, indicating that providers are less likely to stay in HPSAs where the average family income is increasing.  Along with the estimates on the other local area characteristics (that are directionally similar to the estimates from Table V.1), these findings are in line with our hypothesis that providers serving in HPSAs have a preference to serve underserved populations.

Table V. 2: Logit Models of Retention in Any HPSA by Years since Separation—Primary Care

NOTE: Robust standard errors, clustered by HPSA, in parentheses. *, ** and ***: significant at 10%, 5% and 1%. The base group is defined as providers who are: male, age 26-35, non-physician and non-NP/PA, having a start year of 2009 or later, and serving in a location in the Northeast.  The coefficient on any of the dummy variables included in these models shows the estimated difference relative to the corresponding excluded category.

In the case of mental health HPSAs, male and female providers did not differ in their retention probability in the same HPSA (Table V.3).  Similar to primary care providers, the older the providers the more likely they were to remain in the same mental health HPSA.   Also, providers with a separation/start year at the beginning of our timeframe were in general more likely to remain in the same HPSA.  In terms of differences in retention relative to primary care HPSAs, we estimated a lower probability to remain in the same mental health HPSA for medical doctors and nurse practitioners relative to other mental health workers, with the effect being stronger in the further out separation/start years.  We also estimated more variation in same HPSA retention by Census divisions.  Providers in the East South Central and South Atlantic divisions were least likely to remain in the same mental health HPSA.  Finally, although the estimates on poverty rate and percent of people over 25 who are high school graduates are directionally the same as in the case of primary care models, we found that a higher proportion of Blacks at the local level increases the retention probability of mental health providers.

Table V. 3: Logit Models of Retention in the Same County-Same HPSA by Years since Separation—Mental Health

NOTE: Robust standard errors, clustered by HPSA, in parentheses. *, ** and ***: significant at 10%, 5% and 1%.  The base group is defined as providers who are: male, age 26-35, non-physician and non-NP/PA, having a start year of 2009 or later, and serving in a location in the Northeast.  The coefficient on any of the dummy variables included in these models shows the estimated difference relative to the corresponding excluded category.

As shown in Table V.4, the difference in retention between participants and non-participants disappeared in the case of mental health ‘same HPSA’.  Other important features in Table V.4 are that the differences in retention by age group, provider type, separation/start year and Census division were less pronounced or virtually non-existent, while the differences by local area characteristics are similar to those from Table V.3.

Table V. 4: Logit Models of Retention in Any HPSA by Years since Separation—Mental Health

NOTE: Robust standard errors, clustered by HPSA, in parentheses. *, ** and ***: significant at 10%, 5% and 1%.  The base group is defined as providers who are: male, age 26-35, non-physician and non-NP/PA, having a start year of 2009 or later, and serving in a location in the Northeast.  The coefficient on any of the dummy variables included in these models shows the estimated difference relative to the corresponding excluded category.