Addressing patient’s unmet social needs: disparities in access to social services in the United States from 1990 to 2014, a national times series study

We identified and compiled locations of social service providers across the continental, lower 48 U.S. states annually from 1990 to 2014 from the National Establishment Times Series (NETS). Since 1990, the NETS provides an annual, January census of U.S. business, nonprofit, institutional and government establishments that have a Data Universal Number System (DUNS) identifying number [37]. DUNS numbers are unique identification numbers issued to entities by Dun & Bradstreet (D&B), these numbers are issued and administered as part of D&B’s business of developing predictive credit ratings and scores for banking and insurance purposes [38]. The U.S. Federal Government requires all applicants for federal grants and cooperative agreements to have a DUNS number and requires DUNS numbers to be physical location specific [39, 40]. The Internal Revenue Service and the U.S. Small Business Administration state that each physical location of an entity should have their own unique DUNS number [39, 40]. DUNS numbers are required for the running businesses and not-for-profits, including establishing credit worthiness, setting contracts with suppliers, banking, applying for business certificates and getting business insurance [38]. Because State and Federal funded social service programs are commonly delivered through for-profit and not-for profit entities, and additionally because social service agencies commonly need business services, it is expected that the vast majority of social service agencies have DUNS numbers.

Through a partnership with D&B, Walls and Associates developed the longitudinal NETS product to capture dynamics of the U.S. economy [41, 42]. NETS is considered one of the most comprehensive establishment data sources available, effectively serving as a census of American businesses, organizations and institutions for each year since 1990 [42]. NETS contains over 300 variables per business for each year including company name, location (city, state, zip code, latitude and longitude), relocation history, eight digit modified Standard Industrial Classification (SIC) codes, sales volume, and number of employees. While NETS has primarily been used for research in economics to study job growth, and business formation, dissolution, and relocation [42‐47], researchers have used NETS in health research to characterize neighborhood access to health-relevant businesses and institutions [41, 48‐50].

All establishment addresses in NETS were geocoded to longitude and latitude coordinates using Esri ArcGIS software and a national street network reference data from NAVTEQ 2014. Addresses for each year from 1990 to 2014, were linked to 2010 US Census tract geometries. From NETS, we selected all establishments with SIC codes between 83,220,000 and 83,229,999, “Individual and Family Social Services Establishments”, which include 58 sub-classifications of service activity for entities “…primarily engaged in providing nonresidential social assistance to children and youth, the elderly, persons with disabilities, and all other individuals and families” [51]. The data include sole proprietorships, such as solo-practitioner clinicians who provides mental health services and large multi-location charities and not-for profits such as Catholic Charities. Counts of these establishments per 2010 Census tract geometries were calculated in each year from 1990 to 2014.

The Longitudinal Tract Database (LTDB) was used to estimate population compositional characteristics of 2010 Census tract geometry areas in 1990 and 2010. The LTDB provides Census data harmonized from the 1990 to 2010 Decennial Censuses and aggregated to 2010 Census tract boundaries [52]. For these analyses, we used data on total population, count of people living below the federal poverty line, percent of population living below the poverty line (poverty rate), and inflation-adjusted median household income. Using data on the land area of 2010 Census tract geometries from the 2016 TIGER/Line files, we calculated population density of each tract. Density of social service providers was calculated both per tract area (land area in Km²) and as providers per 100 people living in poverty for each tract in 1990, 2000 and 2010.

As described previously, retail, commercial and institutional establishments that support the living needs of residents and people who work nearby, collectively referred to as “destinations that support living needs,” were identified using NETS [53]. We calculated the density of destinations that support living needs (per Km² land area) for each Census tract in the U.S. for 1990. Also as described previously, we used data from The Center for Transit-Oriented Development to identify all rail transit stops in 2010 in the U.S., with internet research used to identify Census tracts with a rail transit stop in 1990 [53].

Analyses were restricted to tracts that fell within a core-based statistical area (CBSA) in the continental U.S. and had a nonzero population in 1990. We stratified tracts into tertiles of population density and into quartiles of poverty rate using the respective distribution of both variables in CBSA in 1990. For tracts in the highest quartile of poverty rate we calculated the difference in inflation-adjusted median household income from 1990 to 2010 and classified tracts into quartiles of change in median household income. Separate quartile cut points for change in median household income were calculated for tracts in each tertile of population density.

We plotted graphs of mean annual density of social service providers per Km² for tracts in each quartile of 1990 poverty rate, in each tertile of 1990 population density. Graphs were also plotted of mean annual count of social service providers per Km² for tracts in the highest quartile of poverty rate in 1990 by quartile of change in inflation-adjusted median household income from 1990 to 2010. For tracts in the highest quartile of poverty rate in 1990, graphs were also plotted of count of social service providers per 100 persons living in poverty for 1990, 2000 and 2010, stratified by quartile of change in inflation-adjusted median household income from 1990 to 2010.

Generalized linear regression models, with a Gamma outcome distribution and identity link function, were fit with the density of social services providers per tract as the outcome variable; generalized estimating equations (GEE) were used to estimate model parameters. To avoid having zero values for the outcome variable, a violation of Gamma models, a value of 0.000005 was added to all observations, which is equal to one tenth of the smallest non-zero density of social services in the data set. Thus, the dependent variable is the count of social services agencies per Km² within each tract, plus 0.000005, in each year of observation. The GEE accounted for correlation within Census tracts and used an autoregressive 1 structure to account for repeated observations within tracts across years. The regression models predicted annual tract-level social service agencies/Km² using quartiles of 1990 poverty rate, with the lowest quartile of poverty rate set as the referent group, year of observation, and year*quartile of 1990 poverty rate as independent variables. We performed separate analyses for tracts falling in the 1^st, 2^nd and 3^rd tertiles of population density. To assess the effect of quartiles of change in median household income from 1990 to 2010 on the density of social services in tracts in the highest quartile of poverty in 1990, we implemented GEE models with count of social services agencies per Km² within each tract plus 0.000005 in each year of observation as the dependent variable and used variables for year of observation, quartile of change in median household income and year*quartile of change in median household income as independent variables.

We also conducted cross-sectional analyses of associations between quartile of change in median household income from 1990 to 2020 in high-poverty tracts and 1990 median household income and 1990 built environment features. We analyzed associations between quartile of change in median household income from 1990 to 2020 (predictor) and social services per Km² (dependent variable) and between quartile of change in median household income from 1990 to 2020 (predictor) and destinations that support living needs per Km² (dependent variable) using a generalized linear model with a gamma distribution. We cross-sectionally analyzed the association between quartile of change in median household income from 1990 to 2020 (predictor) and 1990 median household income (dependent variable) using linear regression. We also cross-sectionally analyzed the association between quartile of change in median household income from 1990 to 2020 (predictor) and the presence of rail transit stops in 1990, coded and present vs absent as the dependent variable, using logistic regression.

Researchers examining the nonprofit sector note that such organizations tend to cluster [25] because of “cost reducing agglomeration effects” [14]. Agglomeration economics posits that co-location increases the availability of resources necessary for an organization and provides other advantages, such as allowing for specialization of the services an organization provides, better matching between employers and workers, and diffusion of innovation through tighter professional and social networks [54]. For social service providers, spatial clustering of services may also allow for the more efficient provision of specialized services to families with multiple needs. Studies of Los Angeles County investigating agglomeration phenomena for services to families with children and to the homeless, provide some evidence that rather than being solely determined by proximity to potential users, social service location decisions are influenced by functional linkages with similar providers and a willingness of communities to host service facilities [55, 56]. They found that children’s service organizations in LA County were concentrated in middleclass areas, where other related services were located. Volunteers and donations were also more abundant than in poorer areas, and the political culture was more open to controversial human service facilities than in White-dominant suburban communities prone to “not in my backyard” (NIMBY)ism [57].

Our analyses suggest that in 1990, among high-poverty tracts, social service providers were clustered in tracts with amenities that “support living needs,” such as public transit, which facilitates easier commute for both workers and clients, and a more robust retail and commercial environment with shops, restaurants, and banks as well as vendors and suppliers that support the running of the providers. This co-location of providers and retail may, in part, be driven by zoning regulations that commonly establish some areas as being solely for residential uses. Unsurprisingly, our analyses also indicate that tracts that experienced increases in median household income from 1990 to 2010, were richer in amenities that support living needs in 1990.

Contrary to an expectation that the availability of social services would expand in high-poverty tracts that experienced further economic decline, providers per Km² and access per 100 individuals living in poverty increased most among high-poverty tracts that experienced increases in household income through time. These seemingly unexpected trends can be explained by the observation that welfare has evolved to become increasingly marketized through time, in a way that would favor provision of services in high-poverty tracts that experienced income growth. The “Marketization of Welfare,” describes the shifting balance of welfare services towards those who can pay, consequently decreasing access for the disadvantaged [26]. The delivery of social welfare programs in the U.S. involves extensive and complex partnership between government and private nonprofit and for-profit providers, with private institutions funded through government contracts handling the delivery of services. During the 1970 and 1980s, the proportion of income that non-profit social welfare providers generated through fees and commercial income increased and there was a concurrent rise in the for-profit sector in delivering social welfare programs [26]. Subsequent research through the early 2000s showed continued growth of commercial revenue among human services nonprofits [58]. Thus, welfare agency budgets often rely on a combination of government funds or fees and commercial income from paying clients, or reimbursement from client’s insurance providers or employee assistance programs. Such financial models suggest an explanation for the trends in spatial access observed in the data; high-poverty tracts on an upward economic trajectory provided an attractive mix of residents, those in need of government funded services and increasingly those with the means or insurance coverage to utilize fee-for-service products. Conversely, high-poverty tracts on downward economic trajectories were expected to be the least attractive areas for providers with the prevailing budget models. The apparent preference for social service providers to locate in high-poverty tracts that were on an upward economic trajectory may reflect a form of “client creaming” by fee-motivated providers [59].

Our big data approach to tracking changes in spatial access to social services has several strengths including multiple decades of data, a national scope, a uniform approach to identifying social service providers, and the incorporation of both Census and urban form data. However, this approach lacks the nuance of other studies that focused in depth on single cities, specific types of services, had administrative, organizational and/or funding data for providers or that could focus on agency service areas. The number of social services outlets in a particular location is not necessarily indicative of the level of services available in that location due to differential funding, both governmental and private, available to different providers of varying sizes and capacity [16]. Our data did not capture the size of the organizations or the funding they commanded. While the Federal Government directs entities to have a separate DUNS number for each physical location, it is possible that some agencies’ satellite locations did not have a unique DUNS number and the NETS undercounts social service locations. The use of spatial units such as Census units (Blocks Groups and Tracts), ZIP Codes and counties as a unit of analysis is common in urban health and spatial access research, for instance see [13, 18, 60‐62]. However, the use of such spatial units can cause boundary issues in analyses and interpretation [62]. It is clearly possible that residents of a tract with no or relatively few services could access services in a neighboring or nearby tract. However, tract level analyses do allow for the linking of Census and other spatial data to the NETS data and provide a national picture of how the location of social services has evolved as population characteristics have changed over a 25-year period.