Deconstructing rurality to better “place” health data

Abstract

Rural-urban classification schemes are frequently used in ecological studies of population health. However, the algorithms used to produce these classifications as well as their underlying assumptions may not match their intended use in health research. Here, we focus on the spatial distribution of features of the physical environment that are related to health – such as healthcare – to examine the extent to which eight classification schemes capture the heterogeneous context of rural places. We further explore how well rural-urban classifications distinguish between different types of rural places by comparing rural Tribal reservations with other rural areas in the American southwest. Because health services and infrastructure are often distributed through state and federal programs to underserved populations in rural areas, this approach speaks to the broader political implications in how rural communities are defined and represented. Results indicate that rural-urban classifications do not adequately reflect heterogeneous contexts within and across rural places. We advocate for more appropriate population health models that explain contextual differences in the relationship between health and place.

Introduction

Estimates of rural populations and classifications of rural areas depend on what is being measured (Tickamyer, Sherman, and Warlick 2017; Bennett et al. 2019; Zahnd, Mueller-Luckey, et al. 2019; Lin and McAreavey 2021). Approximately 20–21% of the U.S. population lives in sparsely populated rural regions, which account for anywhere from 75–97% of the nation’s total land area (Ratcliffe et al. 2016; U.S. Census Bureau 2020). Rural-urban classifications often rely on GIS matching algorithms and the accuracy or relevance of the geospatial datasets to which they are aggregated. However, many standardized measures of rurality commonly used in population health research often assume homogeneity in the contextual factors that shape health (Zahnd, Mueller-Luckey, et al. 2019; Long, Delamater, and Holmes 2021).

Rural places are host to many place-specific health outcomes and disparities, which may stem from lack of access to resources, higher levels of poverty, and aging populations – among other factors – suggesting that rural populations tend to shoulder a higher burden than their urban counterparts in meeting their basic needs (Singh et al. 2017; Tickamyer, Sherman, and Warlick 2017; Jensen et al. 2020). It should come as no surprise that rurality is considered a social determinant of health (SDH) (Singh et al. 2017; Bennett et al. 2019; Zahnd, Mueller-Luckey, et al. 2019). However, spatially aggregated rural-urban classifications may not necessarily capture the physical context vis-à-vis the built environment of rural places that shape population-level health outcomes (Zahnd, Mueller-Luckey, et al. 2019).

What is rurality?

Contradictions in dominant definitions of rurality remain unresolved because rurality may encompass a range of variables that individually or collectively contribute to health outcomes. That level of refinement cannot be adequately achieved when rurality is assessed as a uniform construct (Lin and McAreavey 2021). Likewise, based on current rural studies, it remains unclear whether what we are referring to as ‘rural’ is truly useful or whether we are glossing over a deeper analysis of the relationship between people, place, and health. While rurality can be a useful predictor of population-level health outcomes, the experience of rurality manifests differently for diverse populations (Hawthorne and Kwan 2013). Simply put, there is more than one type of ‘rural.’ Despite this, rural places are frequently codified and aggregated using simplistic metrics (Bennett et al. 2019).

Many current definitions and measures of rurality do not fully represent the diversity of rural areas and places. We broadly attend to four types of diversity that may not be captured by common rural-urban measures. First, spatial diversity or local variation is smoothed when rurality is defined as a uniform construct across large geographic areas. Second, attribute diversity means that rurality can be defined by many different variables. If we only use a limited set of variables to define rurality, then rurality may not be comprehensively represented. Third, diversity of place means that different people and populations may have different understandings and experiences of rurality. If definitions of rurality do not or cannot account for these human factors, they may fail to reflect the overall effect of rurality on people. Fourth, we also recognize that rurality is dynamic, and that temporal diversity is largely missing from rural classification schemes. However, we focus our current analysis on rurality at a single point in time due to space and computational limitations.

Rurality is a complex arrangement of people, infrastructure, and services and is in some ways the counterpart to urbanity. Flows of resource production, consumption, and waste intertwine rural and urban economies (Kelly-Reif and Wing 2016; Miller 2022). These processes leave rural spaces prone to competing claims over land use, development, resource allocation, infrastructure, recreation activities, and services (Dedina 1995; Sharma-Wallace 2016). As an example, environmental histories of colonial-Indigenous relations in rural regions set the stage for contemporary asymmetrical political power relations between urban and rural communities (Salisbury, de Melo, and Tipula 2012). Historically, colonial claims over Tribal lands (all lands within the U.S.) as sites of resource and labor extraction for urban consumption helped mobilize the creation of Tribal reservations through the myth of “empty” spaces, false promises of scientific advancement, and the production of so-called rural landscapes (Voyles 2015).

Today, Tribal reservation boundaries reproduce physical and social barriers across which bodies, material, and information pass with varying ease or cost, for example through access to transportation or telecommunications. Many – but not all – lands held in federal trust for Tribal reservations are rural, and crossing reservation borders is often a prerequisite for obtaining services like healthcare or retail food. The act of mapping and classifying rurality thus has profound material and social implications:

The state creates marginality by including remote peoples and landscapes within state classifications of space and society, but then often ignores them or creates policies based on imaginative geographies of backwardness and remoteness… borderlands thus provide a landscape of highly uneven power relations (Salisbury, de Melo, and Tipula 2012, 108–109).

By exploring the rural-urban interface as an array of complex and heterogeneous places dependent on flows of capital, bodies, material, and cultures (Watts 1983; Sharma-Wallace 2016; Curley 2021), health researchers can challenge assumptions about the composition of human settlement that limit the type of environmental health and justice analyses needed to address stubborn environmental health and wellness disparities. The specific ways in which rurality may be defined have pronounced impacts on health, environmental, and development policies (Lin and McAreavey 2021). Yet, environmental health literature – and indeed policy – tends to dichotomize rural and urban spaces or outright favor urban ones (Bullard et al. 2008; Kelly-Reif and Wing 2016). Sharma-Wallace (2016) argues that “the [rural-urban] interface deserves more – and more systematic – attention than it has yet received within environmental justice research across sectors, issues, and places” (2016, 175).

The idea of rurality carries multiple meanings about the relationship between place and health. We assert that one way these meanings are maintained is through data produced and disseminated by the state. This has profound policy implications that, when left unexamined, obscure the plural ways that rural health disparities manifest through gaps in knowledge production about, and the material production of, environmental risks (Arce-Nazario 2018), as well as assumptions about how (or whether) the composition or context of rural places relates to health. In this regard, compositional factors relate to individual characteristics within a population, such as race/ethnicity, gender identity, sex, or age, while contextual factors are physical elements (e.g., the built environment) and social conditions (e.g., poverty or segregation) (Cummins et al. 2007; Guthman and Mansfield 2013; Zahnd, Mueller-Luckey, et al. 2019).

Rural-urban classification schemes

The U.S. Census Bureau defines rural America as low-density, sparsely populated areas far from urban centers (U.S. Census Bureau 2017), or more problematically as ‘what is left (Ratcliffe et al. 2016)’ after urban areas are defined. There are more than two dozen federal definitions of rurality (McCormack and Meendering 2016). The U.S. Department of Agriculture (USDA) produces the four most commonly used: Rural-Urban Commuter Area (RUCA) codes, Rural-Urban Continuum Codes (RUCC), Urban Influence Codes (UIC), and the Frontier and Remote Area (FAR) codes. RUCA codes assign categorical values to census tracts and are based on population density and daily commuter flows between metropolitan and micropolitan areas, small towns, and rural communities (USDA 2020). Flows are estimated from American Community Survey (ACS) samples, which have larger margins of error and degrees of uncertainty in less populated areas due to smaller sample sizes (Onega et al. 2020). The RUCC is also based on an urban-rural distinction and assigns categorical values to counties (USDA 2013a). Urban and nonurban county designations are based on the Office of Management and Budget (OMB) Metropolitan and Micropolitan Statistical Areas (MSA). The OMB cautions against using these statistical designations in delineating rural and urban areas, writing that the MSA ‘do not produce an urban-rural classification (OMB 2010).’ The UIC is based entirely on MSA designations (USDA 2013b). Similarly, the National Center for Health Statistics (NCHS) produces the Urban-Rural Classification Scheme (URC) for counties based on the MSA to explain differences in population-level health across the urban-rural continuum (NCHS 2013). Finally, FAR codes designate rural areas by county and ZIP codes across four thresholds of population density ascertained from the U.S. Census with the assumption that all rural areas within each of the four classification levels experience the same degree of diminished access to goods and services. The FAR documentation states that ‘job creation, population retention, and provision of services (such as groceries, health care, clothing, household appliances, and other consumer items) require increased efforts in very rural, remote communities, all things being equal (USDA 2012, 1).’ Of course, all things are not equal.

In the next section, we challenge the assumption that rurality classifications are truly representative of the built environment in rural or urban communities (Zahnd, Mueller-Luckey, et al. 2019) and argue that the uncritical use of state-produced datasets may inadvertently reinforce the power relations that produce disproportionate burdens of rurality.

Critical approaches to geospatial data

The present analysis hinges on the assertion that rural-urban classification codes are geospatial datasets, so we evaluate any assumptions about rural health that they may confer through a geospatial data lens (Xie et al. 2017). As such, we frame our exploration through critical approaches to geospatial data. The production and compilation of data about the world continues apace, but those data are never value neutral. We hold that the value neutrality thesis of science and technology (see Nagel 1961; Leach 1968) is unattainable and assert that all scientific activities in which data are collected or analyzed are situated social practices influenced by larger power structures like the academy or funding bodies (Haraway 1988; boyd and Crawford 2012; Tadaki, Brierley, and Cullum 2014). The framing of scientific “facts” by experts through data shapes public understandings of the world (King and Tadaki 2018). So, despite the ease of access to geospatial datasets, interpreting them as reliable and accurate representations of real-world processes presents a formidable challenge.

As stated, we are chiefly interested in the broad applicability of rural-urban classifications as geospatial datasets and their ability to capture complex processes beyond their intended purpose. Accurately classifying rural places – which rural-urban classification schemes do appropriately to varying degrees – is not the same as directly identifying and characterizing drivers of health outcome disparities. Each rural-urban classification scheme relies upon algorithms that assign levels of rurality with varying degrees of accuracy, depending on the metrics used. However, if these schemes are applied to research beyond their intended scope, researchers risk essentializing rurality as the compilation of uniform characteristics and processes within strata. Results from studies that stratify by rural and urban thus need to be carefully interpreted, and researchers need to consider the underlying assumptions and algorithms that were used to produce those classifications in the first place.

These challenges may be addressed by research focusing on critiques of data science through critical data studies (boyd and Crawford 2012; Dalton, Taylor, and Thatcher (alphabetical) 2016; Iliadis and Russo 2016; Neff et al. 2017; Kitchin and Lauriault 2018). This body of literature is situated at the nexus of science and technology studies, social sciences, policy and legal domains, and the humanities. Critical data studies aim to scrutinize various forms of data science, investigating the production and management of data, claims of political neutrality in datasets, and the influence data have on diverse aspects of life.

We also couch our approach to geospatial data in critical GIS, which emerged in the early 1990s as a response to the rapid proliferation of geospatial data and software (Crampton 2010). The modern paradigm of big geospatial data analytics necessitates renewed interest in the materials and methods researchers and practitioners use and are developing to navigate the changing geospatial data ecosystem. Critical GIS is both critical of GIS and leverages GIS technologies to engage in decolonial, antiracist, feminist, and other critical geographies (O’Sullivan 2006; Pavlovskaya 2018).

Relevance to Population Health

Geospatial data enhances researchers’ ability to uncover highly local and granular patterns of human and environmental health, however we seek ways to question if the assumptions that underpin geospatial datasets adequately address the complex needs of health-based studies. This interrogation helps address the ways in which the proliferation of data-driven algorithms as inputs in decision-making for environmental health governance produces certain hegemonies around environmental knowledge, and by extension, material consequences for governed bodies (Machen and Nost 2021). Along with the social turn in health and medical geography (Kearns 1993; McLafferty 2020), a critical approach to geospatial data research can help guide spatial health research through exploration of the ways that geospatial data are used to inform science and policy.

Although rural classification schemes are frequently used throughout health research as either confounding variables or as exposures unto themselves (Zahnd, Askelson, et al. 2019; Zahnd, Mueller-Luckey, et al. 2019), we hypothesize that they poorly predict SDH in the built environment (e.g. access to healthcare) - an assertion that will be tested in this study. We base our hypothesis in part on the fact that as the size of areal units increases the likelihood of obtaining good model fit in multivariate models increases because aggregating data over larger spatial units tends to smooth out local variations within each unit, making observations within the aggregated unit more similar to each other. This similarity introduces a stronger correlation between variables at higher levels of aggregation (Fotheringham and Wong 1991). Similarly, we expect that as the number of factor levels in rural classification codes decreases, the likelihood of good model fit also increases. Poor model fit is not an indictment of the quality of either classification schemes or the explanatory variables, but rather a statement about their ability to capture the same geographic processes. These arguments are an important reassertion that commonly used datasets are subject to multiple biases and irreconcilable statistical errors. These forms of error propagation are likely not equitable, and good model fit may benefit certain places and populations by overlooking the particularities of heterogeneous rural experiences.

A component of our ongoing work is to characterize the relationship between environmental pathways of exposure and SDH among Indigenous communities in the Southwest US, which include, among other determinants, access to resources like healthcare, clean water, reliable energy, and healthy food. Because these types of services are often distributed through state and federal programs to underserved populations in remote and rural areas, there are political implications regarding how these rural communities are defined and represented. Rural Indigenous communities across the Southwest may be disproportionately underrepresented through those policies. Therefore, to address this gap we are concerned with how rurality and rural places are defined and operationalized in population health research and ask the following questions:

1. To what extent do the rural-urban classification codes described above capture the context of rural places in relation to SDH, and

2. How well do these classifications distinguish between different types of rural places?

As such, the aim of the current study is to measure agreement between rural-urban classification schemes and spatially heterogeneous measures of place-based SDH.

Data and Methods

Study Area

This study encompasses New Mexico, Arizona, Colorado, and Utah in the Southwestern US (Figure 1). These four states are emblematic of much of the American West in that there are densely populated urban centers surrounded by large portions of sparsely populated or uninhabited areas. The study area is also home to a large Indigenous population living both on and off reservation land held in federal trust. Many Tribal reservations across the West are situated in rural areas often remote from population centers. Ongoing and future research by the authors centers on the plural contexts of environmental health across the Navajo Nation, which is situated at the confluence of all four states on the Colorado Plateau. The study area is thus sufficiently large to maximize the efficacy of the models discussed throughout and to reduce edge effects and other computational errors.

Figure 1.

Study area map

Rurality Variables

We developed multiple SDH variables in GIS to represent diverse aspects of place and health in the built environment (Table 1). Our focus on the built environment is driven by data accessibility and spatial precision. The input data are publicly available or otherwise not restricted by privacy protections. Unlike census-based measures of poverty or housing, they are not aggregated to polygon units like census tracts. This allows for more precise analyses free from administrative boundary constraints. Each of the variables that directly incorporate population estimates relies on the LandScan dataset, a roughly 1-km resolution estimate of 24-hour ambient population worldwide (Oak Ridge National Laboratory 2020), downsampled to 10-km grid cells to reduce computational burden. The LandScan data product is derived using remotely sensed imagery and sub-national census data worldwide that are combined through dasymetric mapping approaches.

Table 1.

Rurality variables and data sources.

Layer	Input Data	Data Producer	Data Availability
Hospital accessibility	Hospital bed locations	Definitive Healthcare	Subscription** (ArcGIS Online Feature Layer)
Primary care accessibility	Location of primary care physicians (PCP)	National Provider Inventory (NPI) †	Free
Behavioral healthcare accessibility	Behavioral healthcare providers	NPI †	Free
Pediatrics accessibility	Pediatric healthcare providers	NPI †	Free
Access to retail food	Food retailers	Data Axle U.S. Business Database	Subscription*
Telecommunications coverage	4G LTE Coverage (AT&T Mobility, T-Mobile, UScellular, Verizon)	Federal Communications Commission (FCC)	Free
Land use mix	National Land Cover Database (NLCD)	USGS	Free
Population sparsity	LandScan	Oak Ridge National Laboratories	Free

^*Available to research team via access to institutional library services

^**Available to research team via ArcGIS Online account

^†Databases updated monthly. Data collection occurred in September 2023

Healthcare Accessibility

Many components of the built and social environment coalesce to influence population- and individual-level health (Marmot and Wilkinson 1999). Healthcare services and community engagement from health practitioners are drivers of overall health (Maani and Galea 2020; Dave, Wolfe, and Corbie-Smith 2021). All healthcare accessibility measures here are computed using the enhanced 2-step floating catchment area (E2SFCA) methodology (Luo and Qi 2009), a ratio of supply-to-demand where supply is the capacity of each clinic or healthcare service and demand is the estimated population within a service area. The E2SFCA method defines two catchments – the total number of services available to a population and the total population served by a given service point. In line with similar healthcare accessibility studies (e.g., Wang 2012; Delamater 2013; McGrail and Humphreys 2014) and based on input from the research team members who work with health care providers and patients in the study area, we utilized a 2-hour travel time threshold for measures of accessibility to hospitals, pediatrics, and behavioral health, and a 60-minute travel time threshold for primary care. A limitation of the E2SFCA methodology is that it assumes that the location of service providers is the only measure of how populations obtain services, which is an incomplete picture of healthcare in both urban and rural contexts. For example, health practitioners in rural settings commonly travel long distances to a dispersed patient population, meaning that the physical locations of clinics only partially account for the breadth of services offered (Philo, Parr, and Burns 2003). Given this, we recognize that the physical location of clinics is not necessarily a definitive or complete measure of health outcomes in rural places. However, we use distribution of service points as an indicator of rural landscapes of care and expect it to be more relevant and granular when compared with rural-urban classification codes as a proxy for health care access.

Hospitals.

Hospitals are critical service points for both acute and long-term health for diverse populations. In the 19^th century, hospitals often served as critical welfare institutions for patients who endured adverse structural determinants of health, demonstrating a broad capacity for holistic care (Sullivan 2019). Though the focus of hospitals has generally become more clinical and healthcare has become a marketplace for patients who pay for care (Rosenberg 1982; Sullivan 2019), the multidirectional relationship between community health and hospital performance remains a cornerstone of community-level resilience (Maani and Gaela 2020; Baker et al. 2021; Dave, Wolfe, and Corbie-Smith 2021). We model hospital capacity in the E2SFCA algorithm as the number of beds available to patients within a 2-hour driving time. Data enumerating total hospital beds as of 2020 are derived from the Definitive Healthcare USA Hospital Beds dataset available as an ArcGIS Online feature service (Definitive Healthcare 2020).

Primary Care.

Primary care is a fundamental element of population health and encompasses a diversity of basic healthcare services (WHO 2021). Primary care practitioners (PCPs) often serve as the first point of contact for patients seeking more specific forms of healthcare. Spatial accessibility to PCPs (as opposed to nonspatial barriers to access like health insurance) varies across populations and is often an important indicator of community health (Lardier et al. 2023). We derived the practice location and specialties of PCPs from the National Provider Inventory (NPI) as of September 2023. We then grouped co-located points into singular practice locations such that capacity is the number of providers in a given clinic. The same approach was employed for behavioral health and pediatric specialists. We considered supply from all PCPs within a 1-hour drive time from 10-km grid centroids. The NPI codes used to select PCP specialties are provided in the appendix.

Behavioral Health.

Rural populations in the US are more likely to experience adverse mental health outcomes than urban populations despite similar prevalence of mental illness, underscoring the need for more effective provision of rural behavioral healthcare (Jacob, Bourke, and Luloff 1997; Philo, Parr, and Burns 2003; Morales, Barksdale, and Beckel-Mitchener 2020). Efforts to close this gap include integrating behavioral health care with primary care and pediatric clinics (Hine et al. 2017) and increasing the focus on mobile behavioral health clinics and telemedicine (Wynn and Sherrod 2012; Malone et al. 2020; Peritogiannis, Papathanasiou, and Giotakos 2022). Similar to more urbanized settings, a main challenge of behavioral healthcare in rural places is tailoring services to meet the specific needs of diverse communities and navigating the plural nonsocial, demographic, socioeconomic, interpersonal, and ideological factors that shape the utilization and effectiveness of healthcare provision (see Bachrach 1983; Philo, Parr, and Burns 2003). We modeled access to behavioral health provider locations derived from the September 2023 NPI within a 2-hour drive from 10-km grid centroids.

Pediatrics.

Pediatric populations often require specialized healthcare and may be particularly vulnerable to accessibility disparities (Brown, França, and McManus 2021). Limited access to pediatric medicine is linked to numerous adverse health outcomes that may be mitigated by telemedicine (Curfman et al. 2021; Shah and Badawy 2021). However, the reliability of telemedicine is dependent on access to reliable telecommunications services – an infrastructural burden that may disproportionately affect poorer and other disadvantaged populations. While critical pediatric services are typically consolidated into larger urban clinics, poverty and minority status are typically stronger indicators of these disparities than rural-urban codes (Brown, França, and McManus 2021). The supply to pediatric clinics derived from the September 2023 NPI is modeled within a 2-hour drive from 10-km grid centroids. Rather than using total population estimates as the demand for pediatricians, the estimated population was adjusted to only include people 18 years of age and younger using dasymetric apportionment (Wright 1936; Eicher and Brewer 2001) to multiply the population estimate at each Landscan grid cell by the estimated proportion of the population younger than 18 years old in the census tract in which its centroid resides.

Access to Retail Food

Diet is often implicated in chronic diseases and the physical availability of nutritious food is considered an important determinant of health (Ahalya et al. 2017; Dangerfield et al. 2021). Highlighting the complexity of preventable health outcomes rural populations face, food insecurity is also frequently comorbid with poor healthcare access (Kushel et al. 2006; Ma, Gee, and Kushel 2008), racial and gender disparities (Adam et al. 2010; Tipper 2010), and a loss of multigenerational knowledge (Gilio-Whitaker 2019; Cardarelli et al. 2020).

Recent population health research has shifted its focus from an enumeration and analyses of food availability to the qualitative composition of food environments (Ahalya et al. 2017). The consumer nutrition environment refers to the quality of food available to consumers (Ahalya et al. 2017; Pulker, Thornton, and Trapp 2018). A similar but distinct concept, the food retail environment, refers to the diversity and availability of food retailers within an area or region. How people act within or utilize their food environment varies by region and across the urban and rural gradient according to qualitative factors like cost, convenience, variety, quality, and service. As such, rural populations may tend to commute across greater distances to obtain food from a wider variety of retailers (Lee, Ralston, and Truby 2011; Dangerfield et al. 2021). Given this, Dangerfield et al. (2021) argue that it’s inappropriate to frame food insecurity in rural areas simply as a function of spatial proximity to all retailers.

We adapt the modified retail food environment index (mRFEI), a ratio of retailers of unhealthy food, like convenience stores and fast-food restaurants, to all retail food providers. The mRFEI has notably been calculated by the U.S. Agency for Toxic Substances and Disease Registry (ATSDR) and U.S. Centers for Disease Control and Prevention (CDC) for all census tracts (CDC 2012). To account for variable, often longer, drive times to obtain retail food from rural places, we calculate a drive time-dependent mRFEI (${\mathrm{m}\mathrm{R}\mathrm{F}\mathrm{E}\mathrm{I}}_{\mathrm{d}\mathrm{d}}$) as:

$${mRFEI}_{dd}=\frac{{\sum }_{DB\in {DT}_j}\left(P_{DB}\times \frac{F\prime }{F}\right)}{{\sum }_{DA\in {DT}_j}{P}_{DA}}$$ (1)

where $P_{DB}$ is the population of a given dissemination block, $DB$ (in this case, 10-km grid cell), $P_{DA}$ is the population of a given $DA$, which is the service area of every retailer within a 15-, 30-, 45-, or 60-minute drive time, ${DT}_j$, F is all food retailers in the $DA$, and $\mathrm{F}\mathrm{’}$ is convenience stores and fast food restaurants in the $DA$.

Telecommunications

Telecommunications technology is fundamental for receiving many services, including multiple types of healthcare. Many of our partner communities from other ongoing environmental health research are situated in areas with no cellular coverage and limited broadband internet service. We have observed community liaisons commuting up to two hours to attend remote meetings and field staff who cannot remotely upload collected data to central systems. Furthermore, these data-poor regions are a critical gap in geospatial big data representation.

Here, we model the average drive time in any direction from each 10-km grid centroid to the polygon boundary of reliable 4G LTE mobile data coverage as of May 15, 2021 from four major providers (Federal Communications Commission 2021). We defined grid centroids within the coverage area as 0 indicating an average drive time of 0 minutes to access reliable mobile data.

Land Use Mix

Land use mix (LUM) captures the variety of zoning and land use types within a given area. As LUM increases, the travel time to engage in multiple different activities may decrease (Northridge, Sclar, and Biswas 2003). An assumption in some public health research is that heterogeneous land uses encourage alternative modes of transportation to automobiles, increasing physical activity and interaction with green and blue spaces, which are widely associated with beneficial health outcomes (Northridge, Sclar, and Biswas 2003; Schulz and Northridge 2004; Wolch, Byrne, and Newell 2014; Coutts and Hahn 2015; Jennings and Gaither 2015; Laatikainen, Hasanzadeh, and Kyttä 2018).

We modeled LUM by calculating a fishnet-constrained LUM index (Xu et al. 2017) to measure the entropy of developed land uses as:

$$LUM=\frac{-\left[\sum p_{i,j}\ln p_{i,j}\right]}{\ln n_j}$$ (2)

where $p_{i,j}$ is the proportion of land use as developed land (all intensities) from the 30-meter 2020 National Land Cover Database, $i$, within 10-km grid cells, $j$, and $n_j$ is the number of distinct land uses in area $j$.

Population Sparsity

Population sparsity is a shorthand indicator of rurality. Despite some early attempts to measure sparsity (e.g., Openshaw and Coombes 1991), population density tends to be the most common metric of sparse rural populations (Le Tourneau 2020). However, the distribution of population density across large regions is highly right-skewed (Cohen and Greaney 2023), meaning that density alone makes it difficult to capture the degree of spatial isolation that rural populations experience. We represent population sparsity as the minimum bounding geometry of a subpopulation of a predefined size. Specifically, we constructed circles radiating from each 10-km grid centroid to select an estimated LandScan population subset of size n. We calculated this minimum bounding circle for three values of n defined as ¼, ½, and ¾ standard deviations of the total estimated population of the study area. The output is the diameter of the minimum bounding circle for each grid cell – more sparsely populated areas have larger diameters than densely populated ones.

Forest-Based Classification

To determine how well rural-urban classification schemes capture heterogeneous contexts of place within areal units and across classification factor levels, and if they adequately distinguish between different contexts related to population-level health across similar classes, we used all place-based variables discussed above to predict each classification level. We used a forest-based classification (FBC) to fit all variables to common rurality classification schemes, specifically, the RUCA, RUCC, UIC, URC, and FAR codes. Rurality classification codes were geographically co-registered with the point locations of 10-km grid centroids. All codes and factor levels used are described in Table 2. FBC is a machine-learning model based on the random forest algorithm (Breiman 2001) that makes multiple decision trees to derive many predictions based on input data. The model then votes on each individual tree’s prediction to make a final forest-based prediction. FBC models are an effective way to account for model overfitting (Breiman 2001).

Table 2.

Rural-urban classification codes and factor levels.

Code	Aggregation Unit	Factor Levels	Measurement	Citation(s)
RUCA	Census tract	1–10 (urban to rural)	Commuter flows to and from metropolitan cores. Derived from American Community Survey.	(U.S. Department of Agriculture 2020)
RUCA – 4 classes	Census tract	1 (urban): 1.0, 1.1, 2.0, 2.1, 3.0, 4.1, 5.1, 7.1, 8.1, 10.1 2 (large rural city/town): 4.0, 4.2, 5.0, 5.2, 6.0, 6.1 3 (small rural town): 7.0, 7.2, 7.3, 7.4, 8.0, 8.2, 8.3, 8.4, 9.0, 9.1, 9.2 4 (isolated small rural town): 10.0, 10.2, 10.3, 10.4, 10.5, 10.6	-	(Lin and Wimberly 2017)
FAR	Zip code tabulation area (ZCTA)	0–4 (not frontier to frontier)	Population density. Derived from U.S. Census.	(U.S. Department of Agriculture 2012)
UIC	County	1–12 (urban to rural)	Metropolitan and Micropolitan Statistical Area (MSA) designations.
UIC – 3 classes	County	1 (metropolitan): 1, 2 2 (micropolitan): 3, 5, 8 3 (non-core): 4, 6, 7, 9–12	-	(Zahnd, Askelson, et al. 2019)
URC	County	1–6 (urban to rural)	Metropolitan and Micropolitan Statistical Area (MSA) designations.	(National Center for Health Statistics 2013)
RUCC	County	1–9 (urban to rural)	Metropolitan and Micropolitan Statistical Area (MSA) designations.	(U.S. Department of Agriculture 2013a)
RUCC – 2 classes	County	1 (urban): 1, 2, 3 2 (rural): 4–9	-	(Palmer et al. 2013; Weaver et al. 2013; Zahnd, Askelson, et al. 2019)

Each model uses the following explanatory variables: hospital accessibility, behavioral healthcare accessibility, pediatrics accessibility (all within a 2-hour drive time), primary healthcare accessibility within a 60-minute drive time, access to retail food at 15, 30, 45, and 60-minute drive times, average drive time to reliable mobile data coverage, land use mix, 10-km grid cell population, and population sparsity grouping ¼, ½, and ¾ standard deviation of the total study area population. We included an interaction term between population and land use mix to account for grid cells with low entropy and estimated populations of 0. We did not attempt to define a more parsimonious model for any of the schemes. The fitted models predict the rurality classifications according to each scheme and then further return a binary (1, 0) value at each grid centroid indicating whether the model correctly predicted the given rurality classification. The output analysis is further stratified both by the total study area population and the population living on Tribal reservations enumerated by misclassified grid centroids to compare the percentages of “misclassified” populations in two distinct types of rural areas.

We evaluated each model in terms of prediction accuracy, or agreement between predictor variables and rural-urban classification schemes. We assess the effects of prediction accuracy in three ways. First, we stratified prediction accuracy in all grid cells by those whose centroid is either on or off Tribal reservations to illustrate how heterogeneous accessibility to services and adverse health outcomes result in different degrees of prediction accuracy. Second, we stratified the estimated population living within 10-km grid cells where the FBC failed to predict rural-urban classifications by urban (0) and rural (1) to test if the commonplace strategy of dichotomizing rural and urban classification codes in research results in more prediction error in rural places. Finally, we compare the proportion of the estimated population either living on or off Tribal reservations that lives in 10-km grid cells where the model either predicted or failed to predict the outcome by rural-urban classification code. It is not the intention of this paper to rank or recommend which classification schemes outperform others. Additionally, it is expected that the classification schemes with more categories will inherently produce a generally lower prediction accuracy overall when compared with those with fewer categories. Therefore, taken together, our approach is to construct a comprehensive critique of using rural-urban classification schemes in population health research, due to their deviation from SDH foundations.

We computed each FBC in ArcGIS Pro (version 3.2.2) with 1000 trees, 5 randomly sampled variables, 100% data available per tree, and 30% training data using a parallel processing factor of 100% on a PC with 16 cores (12^th Gen Intel 19–12900K CPU, NVIDIA GeForce RTX 3060 GPU). Each model took approximately 4 hours to complete.

Results

Table 3 presents model diagnostics as well as counts and proportions of the estimated population within misclassified grid cells for each scheme. The mean standard errors (MSE) range from 36.17 to 76.96. RUCC (9 classes) and the UIC have the highest MSE and the 2-factor RUCC classification has the lowest MSE. We further explored contingency tables of model predictions for each rural-urban classification scheme and computed both Cohen’s $k$ and Gwet’s ${AC}_1$ statistics to measure reliability of model predictions following recommendations from Zec et al. (2017) and Vach and Gerke (2023). We also considered the weighted average F1 scores for each classification, which provide a balanced measure of precision and recall across categories, with each category’s contribution proportional to its size. This approach calculates the harmonic mean of precision and recall for each category and then weights it by the number of samples, offering a realistic assessment of model performance. All metrics demonstrate low agreement between predicted and observed values for all models although the 2-factor RUCC code is the best predicted scheme of the eight despite overall poor prediction consistency ($k=0.24,{AC}_1=0.38$, F1 = 0.64).

Table 3.

FBC model diagnostics

Scheme	MSE	k	AC 1	F1	Pop. Misclassified (%)	Tribal Reservation Pop. Misclassified (%)
RUCA	53.81	0.14	0.42	0.35	1,163,930 (7%)	134,929 (15%)
RUCA – 4 classes	53.13	0.16	0.33	0.42	1,042,174 (7)	311,904 (34)
FAR	47.81	0.29	0.44	0.50	353,740 (2)	34,101 (4)
UIC	74.59	0.08	0.20	0.22	2,740,416 (20)	125,932 (14)
UIC – 3 classes	53.76	0.18	0.23	0.46	973,842 (6)	110,209 (12)
URC	60.58	0.10	0.30	0.34	2,906,610 (21)	65,027 (7)
RUCC	76.96	0.09	0.13	0.21	2,488,798 (18)	132,124 (15)
RUCC – 2 classes	36.17	0.24	0.38	0.64	692,364 (4)	67,130 (7)

The range of the proportion of misclassified population living within 10-km grid cells is 2–21% (353,740 – 2,906,610 people) (Table 3), with the poorest prediction accuracy for URC codes. We see lower levels of prediction accuracy among grid cells whose centroids are within Tribal reservations; 4–34% respectively (67,130 – 311,904 people). Notably, the RUCA codes reclassified into four factor levels, an approach that has been commonly used in population health research, demonstrates more error among this subpopulation despite generally similar overall MSE and agreement with the other schemes, suggesting that a significant proportion of overall error in this model occurs on Tribal reservations.

Summary statistics of the distribution of percent variance explained by each variable in all eight models are listed in Table 4. The average drive time to reliable mobile data coverage is the most influential variable in most models, with the highest percent variance explained (12%) in models of UIC, URC, and RUCC, all of which are aggregated to counties. Population sparsity measured by grouping ¼ standard deviation of the total study area population tended to be least important predictor, which at the lowest explained 3% of variance in models for RUCC (both full and with two classes), URC, UIC, and RUCA collapsed to four classes. The interaction term between LUM and grid population explains between 7 and 9% of model variance, indicating that LUM is moderated by the spatial distribution of people.

Table 4.

Summary statistics of variance explained by variables in all models.

Variable	Mean (SD)	Min	Max
Telecommunications	11.375 (0.518)	11	12
Food Access (15m)	10.250 (0.707)	9	11
Population	9.250 (0.707)	8	10
Food Access (30m)	9.000 (0.535)	8	10
Population * LUM	8.000 (0.535)	7	9
Food Access (45m)	7.250 (0.463)	7	8
Primary Care Accessibility (60m)	7.125 (0.641)	6	8
Food Access (60m)	6.125 (0.835)	5	7
Pediatrics Accessibility	5.875 (0.641)	5	7
Behavioral Healthcare Accessibility	4.875 (0.354)	4	5
Sparsity (0.75 SD)	4.875 (1.356)	3	7
Sparsity (0.5 SD)	4.250 (0.886)	3	6
Land Use Mix	4.125 (0.354)	4	5
Hospital Accessibility	3.875 (0.641)	3	5
Sparsity (0.25 SD)	3.500 (0.535)	3	4

We evaluated the prediction accuracy by factor level to understand the ability of each classification scheme to discriminate between different degrees of rurality in terms of their physical context related to population-level health. We further stratified this evaluation by grid cells whose centroids are either within or outside of Tribal reservation boundaries (Figure 2). In general, we find that except for UIC codes, all classification schemes reliably identify the most rural classification levels. All classification codes poorly distinguish between our contextual models at factor levels designed to identify different degrees of rurality.

Figure 2.

Prediction accuracy by rural-urban classification scheme and code level, stratified by observations on or off Tribal reservations. (Note: No grid cells off Tribal reservations are classified as RUCA level 6 and no cells on Tribal reservations are classified as UIC levels 11 or 12 (white cells).)

Next, to reflect the common practice of refactoring urban-rural classification codes to dichotomous variables (Long, Delamater, and Holmes 2021), and to standardize the classification schemes for a better comparison to avoid the effect of imbalanced classification categories, We estimated the population at each grid cell whose centroid falls within an areal unit reclassified as either urban/metropolitan (0) or rural/non-metro (1) following the approach detailed by Zahnd et al. (2022). We plot the proportion of this population living in grid cells where prediction accuracy is 0 in Figure 3 and compared distributions using the Wilcoxon rank-sum test to account for the nonparametric distribution of each population. Results demonstrate that group means are significantly dissimilar (W = 103, p < 0.01) and that the mean proportion of misclassified populations in rural-coded areas is higher than in urban-coded ones (0.51 and 0.22, respectively).

Figure 3.

Proportion of population misclassified by binary urban/rural designation (Wilcoxon rank-sum W = 103, p < 0.01). Red diamonds indicate median value of distributions.

Finally, differences in the proportion of the misclassified population in each model stratified by grid cells either on or off Tribal reservation land are listed in Table 5. We find that in general relatively low proportions of the overall population are estimated to live in grid cells where there is a mismatch between models and classification schemes. There are only slight differences between the proportion of the population living in misclassified cells on and off Tribal reservations. A notable exception is the RUCA scheme reclassified to four codes, where more than one-third of the estimated population living on Tribal reservations live in grid cells that are misclassified by the model.

Table 5.

Estimated population and proportion of total population living in misclassified grid cells by classification scheme

	Est. Pop. Misclassified (Proportion)
	Off-Reservation	On-Reservation
FAR	319,639 (0.021)	34,101 (0.043)
RUCA	1,029,001 (0.065)	134,929 (0.149)
RUCA (4 classes)	730,270 (0.046)	311,904 (0.345)
RUCC	2,356,674 (0.151)	132,124 (0.146)
RUCC (2 classes)	625,234 (0.039)	67,130 (0.0742)
UIC	2,614,484 (0.164)	125,932 (0.139)
UIC (3 classes)	863,633 (0.0543)	110,209 (0.122)
URC	2,841,583 (0.179)	65,027 (0.072)

Throughout the results and discussion, we refer to people living on land held in federal trust for Tribal reservations, and caution readers not to conflate this term with any racial or ethnic group because Tribal membership is not always a necessary condition for living on reservation land. Therefore, we emphasize that this does not reflect the proportion the population identifying as Indigenous.

Discussion

The models presented here demonstrate agreement between urban-rural classification schemes and variables representing SDH within the built environment collectively. Again, it is not the intention of the paper to rank or recommend which extant classification scheme for use in environmental health research. Evaluating MSE alone, it might be tempting to conclude that variance in the RUCC codes refactored to two classes (urban/rural) is best explained by the model (MSE = 36.17). Not only do other model diagnostics contradict this conclusion ($k=0.24,{AC}_1=0.38$), but overall MSE provides no insight into model robustness especially considering the fact that variables with fewer classification categories are expected to return lower MSE in general. Moreover, when input-output geospatial models suggest potential causal relationships, assuming that the relationships are correctly specified, they also suggest that the areal units are correctly specified. However, it is hard, if not impossible, to determine the extent to which the configuration or size of areal units drives statistically significant results. This is a fundamental tenet of the uncertain geographic context problem (UGCoP) (Kwan 2012b, 2012a), which is similar to but distinct from the modifiable areal unit problem (MAUP) (Wong 2011). The configuration of areal units may be correctly specified in terms of the MAUP, but we cannot conclude that the model has not returned false positives. Conversely, if a zone is incorrectly specified, it may either be due to its spatial configuration or because of spatial or temporal nonstationarity. Therefore, model fit alone cannot justify the zoning or scale of contextual units (Kwan 2012b, 964). In the context of health research, where populations are stratified by urban or rural, rurality may appear to be a significant driver of health outcomes in rural areas, but other contextual factors independent of rurality may be more influential but nonetheless hidden by aggregation.

Moreover, research has shown that as the size of areal units increases, it is almost impossible to get poor model fit (Fotheringham and Wong 1991). This is especially important when considering how the areal units used in the present classification schemes are drawn. Both census tracts and ZCTAs are designed to contain a relatively fixed population – per the U.S. Census, for example, tracts should have an average population of 4,000 and a maximum of 8,000 people (U.S. Census Bureau 2018). This means that census tracts in sparsely populated areas will, in turn, be larger. This example of the scale problem of the MAUP predisposes rural areas to potentially overinflated statistical associations.

Beyond statistical completeness, we are critically interested in who is encompassed by margins of error. As presented in Table 5, an estimated 692,364 people (4.3%) live in places where the RUCC’s two-factor rural-urban classification does not capture the heterogeneity of infrastructure in rural places. This metric reaches an astounding 2.9 million people (20.9%) with the URC scheme, which is specifically designed to capture ‘the associations between urbanization level of residence and health and to monitor the health of urban and rural residents (National Center for Health Statistics 2013).’

Incidentally, the ancillary finding that model error reduces with fewer factor levels is not surprising because these are more parsimonious models. However, as model complexity is systematically reduced, so too is important contextual nuance (Hirschman 1985; Neal 1996). Here, we demonstrate the effect that model oversimplification exerts on grid centroids within Tribal reservations (Table 5). Specifically, the error in the four-factor RUCA code is cause for alarm because RUCA codes are often preferred for their flexibility when reducing degrees of freedom in multivariate models (Hall, Kaufman, and Ricketts 2006).

Nonetheless, it remains common practice to aggregate factor levels in regression models to return more reliable coefficient estimates. However, parsimony is often favored at the expense of context:

Deliberately limiting the complexity of the model is not fruitful when the problem is evidently complex. Instead, if a simple model is found that outperforms some particular complex model, the appropriate response is to define a different complex model that captures whatever aspect of the problem led to the simple model performing well (Neal 1996, 103–104).

To this end, we argue that stratifying health outcomes by broad rural-urban classifications provides limited insight into how the specific context of rural places influences those outcomes. Rather than abandoning such classifications or favoring one scheme over another, researchers should integrate detailed place-based health measures—such as those used in this study—to develop more precise models that explain why rural-urban stratification yields certain results. While these classifications remain central to policymaking, the increasing availability of high-resolution geospatial datasets for rural areas enables researchers to move beyond generalized categories, leading to more accurate and meaningful analyses of place-based health disparities. Future work should prioritize making such datasets publicly available and easily accessible to researchers whose work informs policy.

All models we present here demonstrate general agreement in the relative importance of individual variables overall (Table 4). Access to reliable mobile data coverage tends to explain the most variance and underscores the persistence of the digital divide along rural and urban lines (Li, Chen, and Wu 2020). Despite research suggesting that the digital divide is more socioeconomic than spatial (Hindman 2000), the post-pandemic emphasis on and rapid proliferation of telemedicine services (OASH 2020) may further entrench the health and economic disparities faced by rural populations.

Population is a strong predictor in all models, demonstrating the overall influence cities have on the classification schemes. This also explains why the interaction between LUM and population is influential. As the estimated grid population approaches or equals zero the LUM index follows, right-skewing the numerical distribution of LUM and emphasizing urban places. We also see the effect of population density on the population sparsity variables, which measure the minimum bounding geometry from each 10-km grid cell required to group a predefined population. The minimum bounding geometry and population size are positively correlated, and the distribution of population sparsity values becomes less skewed as population increases, meaning that there is more variance in sparsity values for the model to discriminate between rural and urban grid cells.

The second strongest overall explanatory variable is the mRFEI-dd within a 15-minute drive time from grid centroids. This is likely for two reasons related to the influence of population on models. First, within a 15-minute drive time cutoff more food retailers will be captured in urbanized areas than rural ones. Second, the mRFEI-dd model is qualitative in that it orders retailers by the types of food sold and creates a ratio of unhealthy food options to all food options. Mathematically, this means that if no food retailers are within a 15-minute drive time area the ratio of F’ to F is undefined, making a clear distinction between areas where food retailers are concentrated from more rural and frontier places with few retailers.

Conclusion

By evaluating the agreement between SDH variables and classifications of rurality, we aimed to problematize the assumption that stratifying populations by rural or urban in health research is truly informative or sufficient. Rurality and rural-urban classification schemes are extensively deliberated in GeoHealth literature. The social dimensions of geospatial data intersect with rurality/urbanity through geospatial datasets produced by the state (and indeed corporate actors with business motivations). Despite a rich body of literature problematizing a rural-urban binary, this dichotomy remains relevant and continues to inform policy, economic planning, business development, and political identities. Given the pervasive nature of both rural-urban concepts and data utilization, we advocate for critical approaches to evaluating geospatial data.

Here, we demonstrate the volatility of urban-rural classification schemes and focus on their individual efficacy as either confounding or exposure variables in population health research. They fail to capture the diversity of rural spaces, whether or not rurality is an indicator of disparate access to services or increased exposure to adverse health outcomes. The inability of urban-rural classification schemes to represent these deficiencies is more pronounced on Tribal reservations, places that require careful attention when considering the burdens faced by populations living there. Despite providing nominal degrees of rurality, these classifications hinge on the assumption that rurality is homogenous within areal units and that the definitive characteristics of rurality manifest similarly across regions regardless of local specificity and the plural lived experiences of local populations.

Inadequate representation of rural populations in health studies can potentially have severe consequences. Watson et al. (2022) argue that clinical trials in rural places occur far less often than in urban ones, potentially driving adverse health outcomes among rural populations at disproportionate rates. They conclude that future clinical research needs to ‘systematically assess and report rurality (Watson et al. 2022, 838)’. However, complications arise when completeness or robustness of datasets is conflated with accuracy, and when precise location data do not fully represent true causally relevant areas. Approximations in spatial data are often overlooked and are instead treated as objective truths (Parrinello, Benson, and von Hardenberg 2020). If the rural-urban classifications discussed here are the most readily available datasets, any systematic assessment of rurality is likely to overlook important contextual nuance about heterogeneous rural places and populations. As stated above, our argument is not that rural-urban classifications are necessarily bad or even harmful, but rather that they may be interpreted as conveying information beyond their scope. Our recommendation is to avoid parsimony where possible by hypothesizing and testing what contextual characteristics of place are relevant to health outcomes.

Geospatial data scientists – including the authors of this paper – play a role in shepherding the GeoHealth datasets used by experts in other academic domains. Critical engagement with GeoHealth data (and geospatial data writ large) needs to be bottom-up and needs to pave the way for health and environmental research, which always affects the world and its diverse inhabitants.

Biographies

Daniel Beene, Ph.D. is a postdoctoral fellow at the Johns Hopkins Bloomberg School of Public Health in the Department of Epidemiology. He received his PhD from the University of New Mexico. His contribution includes conceptualization, methodology, software, validation, analysis, data curation, visualization, writing, and editing.

Yan Lin, Ph.D. is an associate professor in the Department of Geography and the Social Science Research Institute at The Pennsylvania State University. Dr. Lin received her Ph.D. from Texas State University. Her contribution includes conceptualization, supervision, writing, and editing.

Joseph H. Hoover, Ph.D. is an assistant professor in the Department of Environmental Science at the University of Arizona. He received his Ph.D. from the University of Denver. His contribution includes supervision, writing, and editing.

Xun Shi Ph.D. is a professor of geography and chair of the Department of Geography at Dartmouth College. He received his Ph.D. from the University of Wisconsin-Madison. His contribution includes writing and editing.

Appendix A: NPI and NACIS Codes Used

The following NPI codes were used to identify primary care physicians: 207Q00000X – Family Medicine Physician, 208D00000X – General Practice Physician, 207R00000X – Internal Medicine Physician; behavioral health specialties: 101200000X – Drama Therapist, 103TA0400X – Addiction (Substance Use Disorder) Psychologist, 101Y00000X – Behavioral Health and Social Service Providers: Counselor, 101YM0800X – Behavioral Health and Social Service Providers: Counselor, Mental Health, 101YP1600X – Behavioral Health and Social Service Providers: Counselor, Pastoral, 101YP2500X – Behavioral Health and Social Service Providers: Counselor, Professional, 101YS0200X – Behavioral Health and Social Service Providers: Counselor, School, 102L00000X – Behavioral Health and Social Service Providers: Psychoanalyst, 102X00000X – Behavioral Health and Social Service Providers: Poetry Therapist, 103G00000X – Behavioral Health and Social Service Providers: Clinical Neuropsychologist, 103GC0700X – Behavioral Health and Social Service Providers: Clinical Neuropsychologist, Clinical, 103T00000X – Behavioral Health and Social Service Providers: Psychologist, 103TA0700X – Behavioral Health and Social Service Providers: Psychologist, Adult Development and, 103TB0200X – Behavioral Health and Social Service Providers: Psychologist, Cognitive and Behavioral, 103TC0700X – Behavioral Health and Social Service Providers: Psychologist, Clinical, 103TC1900X – Behavioral Health and Social Service Providers: Psychologist, Counseling, 103TC2200X – Behavioral Health and Social Service Providers: Psychologist, Clinical Child, 103TE1000X – Behavioral Health and Social Service Providers: Psychologist, Educational, 103TF0000X – Behavioral Health and Social Service Providers: Psychologist, Family, 103TF0200X – Behavioral Health and Social Service Providers: Psychologist, Forensic, 103TH0004X – Behavioral Health and Social Service Providers: Psychologist, Health, 103TH0100X – Behavioral Health and Social Service Providers: Psychologist, Health Service, 103TM1800X – Behavioral Health and Social Service Providers: Psychologist, Mental Retardation and, 103TP0016X – Behavioral Health and Social Service Providers: Psychologist, Prescribing (Medical), 103TP0814X – Behavioral Health and Social Service Providers: Psychologist, Psychoanalysis, 103TP2700X – Behavioral Health and Social Service Providers: Psychologist, Psychotherapy, 103TP2701X – Behavioral Health and Social Service Providers: Psychologist, Group Psychotherapy, 103TR0400X – Behavioral Health and Social Service Providers: Psychologist, Rehabilitation, 103TS0200X – Behavioral Health and Social Service Providers: Psychologist, School, 104100000X – Behavioral Health and Social Service Providers: Social Worker, 1041C0700X – Behavioral Health and Social Service Providers: Social Worker, Clinical, 1041S0200X – Behavioral Health and Social Service Providers: Social Worker, School, 106E00000X – Behavioral Health and Social Service Providers: Assistant Behavior Analyst, 106H00000X – Behavioral Health and Social Service Providers: Marriage and Family Therapist, 163WP0807X – Nursing Service Providers: Registered Nurse, Psychiatric/Mental Health, Child, 163WP0808X – Nursing Service Providers: Registered Nurse, Psychiatric/Mental Health, 163WP0809X – Nursing Service Providers: Registered Nurse, Psychiatric/Mental Health, Adult, 363LP0808X – Physician Assistants and Advanced Practice Nursing Providers: Nurse Practitioner, 364SP0807X – Physician Assistants and Advanced Practice Nursing Providers: Clinical Nurse Specialist, 364SP0808X – Physician Assistants and Advanced Practice Nursing Providers: Clinical Nurse Specialist, 364SP0809X – Physician Assistants and Advanced Practice Nursing Providers: Clinical Nurse Specialist, 364SP0810X – Physician Assistants and Advanced Practice Nursing Providers: Clinical Nurse Specialist, 364SP0811X – Physician Assistants and Advanced Practice Nursing Providers: Clinical Nurse Specialist, 364SP0812X – Physician Assistants and Advanced Practice Nursing Providers: Clinical Nurse Specialist, 364SP0813X – Physician Assistants and Advanced Practice Nursing Providers: Clinical Nurse Specialist, 2084P0800X – Psychiatry; pediatric specialties: 208000000X – Pediatrics Physician, 2080N0001X – Neonatal-Perinatal Medicine Physician.

The following North American Industry Classification System (NAICS) codes were used to identify food retailers: 44512001 – Convenience Stores, 44523001 – Farm Markets, 44511001 – Food Markets, 44511002 – Food Products-Retail, 44523003 – Fruits & Vegetables & Produce-Retail, 44511003 – Grocers-Retail, 72251304 – Deli-Bakery, 72251302 – Delicatessens, 72251301 – Foods-Carry Out, and 72251303 – Sandwiches.

Data and Codes Availability Statement

All codes and data required to reproduce layers in the present study are provided via Figshare repository here: https://doi.org/10.6084/m9.figshare.25727235. Code for the enhanced two-step floating catchment area (E2SFCA) requires the following inputs: origins (population grid centroids), destinations (healthcare clinics with capacity), and origin-destination (OD) cost matrices of population served by healthcare providers. Because these matrices are computed using proprietary ArcGIS StreetMap Premium subscription-based cloud data service, which is not available for download, they are provided in the repository. Detailed instructions on calculating origin-destination cost matrix with parameters used in this study are included in the Figshare repository. We also provide the service area polygon shapefiles at 15-, 30-, 45-, and 60-minute drive times and point locations of food retailers used to compute respective mRFEI-dd calculations. Finally, we provide the Python code used to compute population sparsity.

The present study utilized the following publicly available datasets:

1. USDA Rural-Urban Commuting Area (RUCA) Codes (2010), available for download from https://www.ers.usda.gov/data-products/rural-urban-commuting-area-codes/.

2. USDA Frontier and Remote Area (FAR) Codes (2010), available for download from https://www.ers.usda.gov/data-products/frontier-and-remote-area-codes/.

3. USDA Urban Influence (UIC) Codes (2013), available for download from https://www.ers.usda.gov/data-products/urban-influence-codes/.

4. USDA Urban Influence Codes (UIC) (2013), available for download from https://www.ers.usda.gov/data-products/urban-influence-codes/.

5. CDC-NCHS Urban-Rural Classification (URC) Scheme for Counties (2013), available for download from https://www.cdc.gov/nchs/data_access/urban_rural.htm.

6. USDA Rural-Urban Continuum Codes (RUCC) (2013), available for download from https://www.ers.usda.gov/data-products/rural-urban-continuum-codes/.

7. ORNL LandScan (2020), available for download from https://landscan.ornl.gov/.

8. National Provider Inventory (NPI) codes (as of September 2023), available for download from https://download.cms.gov/nppes/NPI_Files.html.

9. Federal Communications Commission (FCC) 4G LTE Coverage (AT&T Mobility, T-Mobile, UScellular, and Verizon), available for download from https://www.fcc.gov/BroadbandData/MobileMaps/mobile-map.

10. USGS National Land Cover Database (NLCD) (2019), available for download from https://www.usgs.gov/centers/eros/science/national-land-cover-database.