An estimated 10.5% of the US population has diabetes, and these 34 million individuals are at an increased risk for coronary artery disease, cerebrovascular disease, and other complications [1,2]. Approximately 90% to 95% of people with diabetes have type 2 diabetes (T2D) [1]. Another 88 million individuals have prediabetes, defined as having elevated glucose levels above normal but below the threshold for diabetes, and are at elevated risk of developing T2D [3,4]. Diabetes prevalence and incidence vary substantially by geographic region [5-7]. In 2013, there was a six-fold difference between counties with the lowest and highest diabetes prevalence [6]. A large body of literature links community and environmental factors (hereafter referred to as community factors) to T2D and obesity, one of the risk factors for T2D [8-15]; however, the mechanisms for these links remain poorly understood. Moreover, there are inconsistencies in this body of literature. Identifying the community factors driving T2D disparities and the pathways through which these factors influence T2D is essential to informing geographically targeted policy interventions that reduce the risk of T2D and related outcomes.

Researchers have identified consistent associations of community-level socioeconomic factors (eg, community poverty rate) with T2D prevalence and incidence. However, findings related to aspects of the built (eg, food and physical activity establishment environment, land use environment) and natural environment (eg, greenness) and T2D risk have been less uniform [9-19]. This may be due, in part, to methodologic variations in measuring community factors, including differences in spatial scales, data sources, and measurement approaches [20].

Among the challenges to creating a cohesive body of research is a lack of consistent approaches to conceptualizing and operationalizing the geographic area in which community factors are thought to be relevant to health [19-22]. Furthermore, the size and boundaries of spatial scales most relevant to health may vary according to community type (eg, across the gradient from urban to rural). Community type is also an important consideration in measurement development, as measurement of the same community factors may require different approaches [23]. For example, car ownership may be a basic necessity for individuals living in rural areas but more of a luxury for individuals living in urban areas with good public transportation options. Thus, car ownership may work differently as an indicator of the neighborhood socioeconomic environment (NSEE) in urban versus rural areas [24]. This differential item functioning may contribute to inconsistencies observed in the literature as the same measure (eg, proportion who own a car) could hold different meanings in different community types.

Community and health research is also vulnerable to structural confounding, which occurs when individual and contextual factors strongly predict residence in a certain community [25,26]. In the presence of structural confounding, certain measures, owing to social sorting, are largely nonoverlapping, resulting in an inability to examine their independent influences on the outcome of interest. For example, in some settings, the distribution of persons across categories of the NSEE and racial residential segregation may reveal a lack of comparable groups across key strata, resulting in analytical challenges of nonpositivity that prohibit causal contrasts across levels of exposure [26]. Finally, capturing the complex interactions among multiple community factors on health can be challenging [27]. As a result, previous research has largely evaluated co-occurring community factors in isolation.

The increasing availability of longitudinal, individual-level data from electronic health record (EHR) networks [28] and cohort studies, coupled with advances in geographic information systems (GISs), provides new opportunities to examine the effects of community factors on health. In 2017, the Diabetes LEAD (Location, Environmental Attributes, and Disparities) Network was established to identify the contributions of modifiable community factors on T2D risk. The Network includes researchers from 4 academic institutions who collaborate to address the methodological challenges previously described to investigate a range of community factors across the United States. The Network aims to guide policy decision making to reduce the burden of T2D across the United States. This paper aims to describe the Diabetes LEAD Network, its study populations, and the methodologies used to investigate the community factors that are associated with T2D onset and related outcomes.

The Diabetes LEAD Network is a research collaboration of 4 academic centers: Drexel University, Geisinger and Johns Hopkins University (G/JHU), New York University School of Medicine (NYU), and the University of Alabama at Birmingham (UAB). The Centers for Disease Control and Prevention (CDC) funded the Network to bring together institutions with diverse but complementary expertise and a rich array of data assets. Three study sites—G/JHU, NYU, and UAB—use longitudinal data, such as EHRs, administrative claims, and survey data on distinct populations and geographies in the United States (Tables 1-3; Figures 1-3). Drexel, the data coordinating center (DCC), is leading the development of a set of harmonized community factors, health outcomes, and analysis plans (Tables 4 and 5) that will be applied to each study site’s cohort and geography. Each site has its own set of study aims that examine community factors and T2D outcomes, including T2D onset, obesity, and other cardiometabolic conditions. Working collaboratively, the study sites and the CDC also developed a shared set of aims that complement site-specific aims (Textbox 1). We first describe the shared Network aims and then describe the site-specific aims.

The Network aims to evaluate the association of community factors and T2D outcomes (aims 1 and 2) and to evaluate and address the previously described methodological challenges of community and health research (aims 3 and 4):

The Diabetes LEAD Network draws from individuals living in all 50 US states (Figures 1-3). The G/JHU participants were selected from among 1.6 million individuals in the Geisinger EHR, spanning 37 counties in central and northeastern Pennsylvania (Table 1). The G/JHU participants range in age from 10 to 97 years, are predominately White (542,128/578,458, 93.71%) and non-Hispanic (553,211/578,458, 95.66%), reflecting the region’s population [32]. More than 40% of G/JHU participants reside in areas that the US Census Bureau categorizes as rural. The NYU population of veterans spans all US counties. NYU assembled a baseline cohort of veterans who do not have diabetes (n=6,082,246) and are receiving primary care through Veterans Affairs. Veterans are mostly male (5,578,056/6,082,246, 91.71%) and predominantly White (4,411,233/6,082,246, 72.53%) followed by Black (920,596/6,082,246, 15.14%).

These demographic groups are represented by large sample sizes. The UAB population includes participants from the Reasons for Geographic and Racial Differences in Stroke (REGARDS) study cohort [33]. At baseline (2003-2007), the REGARDS study enrolled 30,239 non-Hispanic Black and non-Hispanic White adults aged 45 years and older with oversampling of participants from the Stroke Belt region (Alabama, Arkansas, Georgia, Louisiana, Mississippi, North Carolina, South Carolina, and Tennessee) [34]. To assess T2D onset, participants without T2D at baseline and who completed the follow-up in-home examination (n=11,199) will be evaluated. Patients were not invited to comment on the cohort development or study design.

The DCC is leading the development of harmonized, Network-wide approaches to measuring community factors of interest and T2D outcomes. To develop measures of community factors (Table 4), the DCC is using archival data available at the national level, including publicly available data (eg, US Census) and data elements previously created for the Retail Environment and Cardiovascular Disease (RECVD) study. The RECVD study has longitudinal measures of food, fitness, and social establishments based on the National Establishment Time Series (NETS), a data source that includes information on more than 58 million US business establishments from 1990 to 2014. For each community factor, the DCC has partnered with a study site with relevant expertise to make decisions regarding data sources, spatial scale, exposure assignment, and approach to measurement.

The DCC is applying a range of measurement techniques to define community factors, including data reduction and measurement models. Measurement development is stratified by community type at the census tract level using a modification of the Rural-Urban Commuting Area (RUCA) from the US Department of Agriculture developed by the Network [35]. After collapsing the original 10 RUCA categories into 3, the DCC further divided census tracts within urbanized areas into 2 categories based on land area, resulting in 4 community-type categories that reflect distinct typologies along the rural-urban continuum.

To the extent possible, the Network is harmonizing approaches to measure T2D onset (Table 5) and diabetes-related outcomes. G/JHU and NYU have worked together to develop EHR-based algorithms based on their previous work [36,37] and diagnosis criteria from the American Diabetes Association [38], using a combination of diagnosis codes, medications, and laboratory measures. With coordination from the DCC, the sites are also standardizing approaches to measure potential confounders, mediators, and effect modifiers.

For each Network-wide aim, the study sites will conduct analyses among their study populations based on a common analytic plan. The DCC is coordinating the development of the analytic plan, harmonizing analytical approaches, including the selection of confounding, mediating, and modifying variables of interest, model building, and model diagnostics. Sites will conduct site-specific sensitivity analyses that include relevant data elements that may not be available Network-wide. This approach allows us to examine consistency in results while leveraging the unique data available at individual sites.

For aims 1 and 2, the Network will employ methods to account for group-level and individual-level data, including multilevel models, Bayesian approaches, and generalized estimating equation models. The Network will conduct causal mediation analysis for aim 1 [39]. For aim 2, we will evaluate effect modification through the inference of interaction terms, creating cross-products between our contextual domains of interest and a predetermined set of individual- and community-level variables, such as age, sex, and race. To guide model development for these aims, sites are developing causal diagrams to formulate and test theoretically based pathways, identify potential confounding influences, and account for potential interaction between measures (Figure 4) [40,41]. To assess spatial residual autocorrelation, the Network will calculate I statistics by Moran (local and global) [42] and use modeling approaches that account for spatial residual autocorrelation, if needed. The Network will conduct sensitivity analyses to evaluate how approaches to measurement of outcomes and community factors impact observed associations.

For aim 3, the Network is exploring strategies to evaluate and address nonpositivity, including propensity scores [43], stratification by community type for all analyses, and latent profile analysis to evaluate community typology [44]. For aim 4, we are evaluating community factors at multiple spatial scales (eg, census tract, network buffer around population-centroid). We will compare the associations of community factors measured at different scales with T2D to assess the influence of the modifiable areal unit problem [21]. The study sites also have site-specific aims, as described below.

The Diabetes LEAD Network leverages a breadth of expertise and data to advance knowledge regarding modifiable community risk factors for T2D onset and related outcomes. The Network brings strengths to its collective mission to provide scientific evidence for targeted interventions and policies. First, the sites provide the Network with community data sources that collectively ensure widespread geographic coverage and variation of community types across a rural-to-urban spectrum. It was of particular importance to ensure representation by rural communities, since CDC reports that diabetes is 17% more prevalent in rural than urban areas [55]. Furthermore, understanding the association between community and health requires the assessment of a heterogeneous set of communities [12,56]. Each of the study sites contributes unique data sources to achieve this goal. The NYU cohort spans the nation, the UAB REGARDS study cohort offers a national study with in-depth data from a high-risk region (Stroke Belt), and the G/JHU population is a regionally representative sample with high rural representation and primary data collection.

A second contribution of the Network is the development of measures of 6 community factors (Table 4) to be examined across diverse geographies and community types. These measures are being developed with consideration of community types, examining community factors within strata of different community types defined along a rural-urban spectrum to avoid potential differential item functioning and nonpositivity [23]. In addition, Network investigators are examining individual and joint associations to better understand how these community factors work in concert to contribute to excess T2D risk. With access to individuals’ residential and commercial addresses, the Network is evaluating spatial scales of various types and sizes to better understand the impact of scale on the findings.

Third, the range of expertise across institutions allows the Network to address methodological challenges common to community and health research [20]. In addition, the Network is advancing methods for conducting research on the role of community and health using data obtained from EHR systems, including extracting historical residential addresses to allow for time-varying exposure estimation. Finally, the Network is harmonizing community factor definitions and analytical approaches to facilitate comparable analyses and replicating analyses across 3 different study populations. This effort to harmonize approaches across multiple settings and populations will advance both the field of community and health research and generate data needed to guide evidence-based policies for T2D prevention in the United States.

There are some limitations to the Diabetes LEAD Network research portfolio. First, although access to longitudinal data will mitigate issues of temporality, the available data do not allow for investigation of early life exposures (eg, childhood) that may influence T2D risk [57]. Second, while the diversity of populations and geographies across the study sites is advantageous to expand the generalizability of findings and include previously underrepresented settings (ie, rural), it complicates comparison of results across sites. However, by harmonizing measurement and analytical approaches the Network will be well positioned to pinpoint reasons for any potential conflicting results that arise. Finally, despite employing advanced analytic approaches, the studies are all observational in design; thus, they are potentially constrained with respect to causal inference due to the risk of residual confounding and neighborhood self-selection [58,59]. The Network will consider methodological approaches such as propensity scores to address this limitation [43].

T2D is a leading cause of morbidity in the United States, with select populations, often defined by geography, affected by a disproportionate burden of disease. The Diabetes LEAD Network identifies modifiable community factors that influence geographic disparities in T2D risk across diverse communities and identifies policy levers to ameliorate these disparities.