More than 25 million Americans have diabetes, and another 80 million have prediabetes; taken together, approximately 1 in 3 Americans have diabetes or prediabetes (1). Diabetes is associated with obesity and physical inactivity; many built environment factors — attributes of the proximate environment — such as access to healthy foods (2), crime level (3), the rural–urban matrix (4–6), and walking (4) are correlated with diabetes prevalence. One of the great challenges in understanding the associations between built environment attributes and diabetes is that both factors vary across the United States. Although studies of diabetes have found spatial variations in incidence and prevalence, there is a paucity of information on how the spatial prevalence of diabetes may or may not be associated with the spatial prevalence of built environment attributes. The importance of understanding the covariance of diabetes with its correlates was made salient by Siordia and colleagues (7), who found that the relationship between poverty and diabetes prevalence varied across the United States, with poverty highly correlated with diabetes in some regions but not in others (7). This finding provided the impetus for our hypothesis that the relationship between diabetes prevalence and county-level built environment attributes is nonstationary (ie, the relationship varies across space).

The objective of our study was to determine how and where diabetes prevalence is associated with built environment attributes at the county level in the contiguous United States. This information could allow programs and interventions to better target populations and attributes of the built environment associated with high diabetes prevalence.

Our study used geographically weighted regression (GWR), a tool that is increasingly used by public health researchers to understand the nuances of such issues as access to health care, disease distribution, and spatial variation in magnitude of health outcome predictors (8–10).

Diabetes prevalence in the United States at the county level ranged from 3.8% to 17.8% and was significantly clustered (Moran’s I = 0.35; z = 540.2; P < .001). We found clusters of high diabetes prevalence in the Southeast and clusters of low diabetes prevalence in Colorado. Diabetes prevalence was significantly correlated with numerous independent variables. Because the percentage of neighborhood food deserts was not significantly correlated at the county level, it was not included in the OLS model. The following 9 variables were included in the OLS model: population density, percentage nonwhite, percentage Hispanic, percentage living below the federal poverty level, percentage with less than high school education, percentage unemployed, percentage obese, percentage physically inactive, and percentage that walked or cycled to work. The OLS model was significant (F _9,3099 = 495.87, P < .001). The model explained 58.8% of the variance in county-level diabetes prevalence. The VIF for all variables was less than 4.0, a commonly used cutoff point, suggesting no multicollinearity (Table 1).

The residuals of the OLS model were spatially autocorrelated (Moran’s I = 0.13; z = 26.4; P < .001). The OLS model overestimated diabetes prevalence for Colorado and New Mexico counties. Similarly, it underestimated the outcomes for Alabama and West Virginia counties.

The GWR model produced coefficients for each county (Table 2, Figure 1, Figure 2). The change in both magnitude and direction of the coefficients suggests spatial nonstationarity of the relationship between the predictors and diabetes prevalence. The direction of the relationship in most counties was as expected. Only a few counties had opposite relationships for the predictors in the GWR model. In most counties, walking or cycling to work was associated with lower diabetes prevalence. However, a few clusters of rural counties in Minnesota, North Dakota, and South Dakota show an association between walking or cycling to work and higher diabetes prevalence. Such nonstationarity demands a more nuanced analysis with a contextual focus. For example, high rates of walking or cycling to work are often associated with multimodal transportation that also includes public transit, which is less likely to be available in rural communities (19).

The adjusted R² for the local GWR model ranged from 0.06 to 0.94; the adjusted R² in the OLS model was 0.58. Explicitly, the global OLS R² of 0.58 masks a wide distribution of local associations between the predictors and diabetes prevalence. Without GWR, we would have been unable to estimate local models. In counties in North Dakota, South Dakota, and Montana, the GWR model explained up to 94% of the variance in diabetes prevalence. However, in Washington and Oregon, the model did not explain much of the variance (6%–37%), a spatial variation that would have been missed with the OLS model alone. Residuals for the GWR model, although significant, were less spatially autocorrelated than residuals for the OLS model (Moran’s I = 0.01; z = 3.74; P < .001). Compared with OLS, the GWR model greatly improved model fit. The GWR model explained more variance in diabetes prevalence and reduced the AICc (ΔR² = 0.22; ΔAICc = 2,008.4).

Poverty level, physical inactivity, and walking or cycling to work were each significantly associated with county-level diabetes prevalence, relationships that were spatially nonstationary across the United States. The variation in parameter estimates from GWR suggests the need to apply this spatial analysis tool to other diabetes studies that have been restricted to global models (2,4). In the global OLS model, 58.8% of county-level diabetes prevalence was explained by race, poverty, obesity, physical inactivity, and walking or cycling to work. However, at an individual county level, the explanatory percentage ranged from 6% to 94%, and the individual county-level models were significantly clustered. This clustering suggests that local contexts, policies, programs, and built environment attributes are associated with diabetes prevalence and that the amplitude of such contexts, policies, programs, and environments varies across the nation.

The dissimilarity in variable coefficients was not a factor of one county alone but was a factor of multiple proximal counties, perhaps because of policy and programmatic spillover from neighboring counties and diffusion of innovation (20). The percentage of nonwhite population in a county had the greatest effect in the Southeast and the Rockies, from Arizona and New Mexico to Idaho and Montana. States in these regions have a high proportion of African Americans, Hispanics, or Native Americans, races/ethnicities with disproportionately high rates of diabetes (21). In several regions (including the Midwest, the Ohio Valley, and New England), poverty had a greater association with diabetes prevalence than any other variable. Physical inactivity had the greatest effect in the Southeast and the Southwest, a pattern similar to that of obesity prevalence (22). Walking or cycling to work was most associated with diabetes prevalence in the Mississippi Valley, the panhandles of Texas and Oklahoma, and south Florida, areas not generally associated with walking or cycling because of their hot summers.

The relationships among nonwhite populations, poverty, physical inactivity, and diabetes are not new (3,4,7). Others found these relationships have a spatial component (23). With the exception of recent work by Siordia and colleagues (7), there has been no investigation into the nonstationarity of these relationships. Similarly, the strong association between walking or cycling to work and diabetes is consistent with findings of other studies (24), but it has not been investigated for spatial heterogeneity or nonstationarity. That there is a significant association between nonwhite populations, poverty, physical inactivity, and diabetes and that this relationship has a spatial but nonstationary association highlights the need for local, context-specific diabetes prevention programs.

There are limitations to GWR and our analyses. GWR equates the local regression coefficients based on those geographic areas (eg, counties) most proximate to the area of interest. That is, the regression equation and coefficients for a county in Missouri are most influenced by bordering counties and other nearby counties, but not influenced by counties in Colorado or North Carolina. This concept is essential for local planning and related to Tobler’s first law of geography, that “everything is related to everything else, but near things are more related than distant things” (25). However, the distance of influence (of predictors or potential interventions) is theoretically unknown and perhaps inconsistent across a geographic area (eg, the continental United States). We chose to use an adaptive kernel bandwidth, which accounted for differences in the size of counties and therefore the distance of influence. This choice should have helped adjust for the fact that, for example, North Carolina has 100 small counties and California has 58 larger counties spread over 3 times the landmass of North Carolina. Because of this discrepancy, the data point (county) was an estimate based on proximate counties as defined by the kernel type. GWR is also limited by the edge effect, whereby counties located on the edges of the United States (ie, coastal regions and the borders with Canada and Mexico) do not have the 360° influence of counties in the nation’s interior.

There are also limitations to our findings. The local R²s accounted for 6% to 94% of county-level diabetes prevalence. In large geographic areas in the Mid-Atlantic, upper Midwest, and Northwest, the 9 variables included in the model explained less than one-third of the variance in diabetes prevalence, which means that most factors associated with county-level diabetes prevalence in these geographic areas must have been missing from our model.

The primary strength of this study is the use of GWR in the analysis of the spatial distribution and correlates of diabetes prevalence. Siordia and colleagues (7) introduced the concept of spatial nonstationarity to the relationship between poverty and diabetes. Here, we extend their work by incorporating additional socioeconomic variables and built environment correlates with diabetes. GWR adds value to public health research and practice by emphasizing location-specific theories of health outcomes and tailored policies for intervention. It scrutinizes the assumption of global relationships between various predictors and health outcomes. Using GWR, public health researchers and practitioners can gain a nuanced understanding of health-related issues and respond to the notion that “all health is local” (25,26). In doing so, they can provide clarity for designing and funding context-specific public health programs and policies, especially for national programs that have local reach, including those of the CDC and the American Diabetes Association (ADA). Our analyses could also be used by local public health departments and ADA offices for resources such as MIYO (Make It Your own - http://www.miyoworks.org/) to tailor messages and materials for their target audiences. The use of GWR is a key advancement in public health research and practice because many health behaviors and outcomes vary spatially (eg, obesity) as do many common predictors (eg, race/ethnicity) (27).

Shedding light on spatial variations can provide new insights into well-established relationships. The methodology of GWR needs to be expanded to additional public health efforts to understand the impact of environment and place on health and how these relationships may vary across space. For diabetes prevalence, we presented an initial step in this direction, but much work remains before we understand why these variations exist and why race/ethnicity, poverty, physical inactivity, and active commuting have little explanatory effect in some regions but explain up to 94% of diabetes prevalence in other regions.