Exploratory spatial analysis of food insecurity and diabetes: an application of multiscale geographically weighted regression

ABSTRACT Both food insecurity and diabetes are important public health concerns. For example, diabetes remained a leading cause of mortality both in the United States (U.S.) and globally during the past decade. Meanwhile, an estimated 11% of households endured food insecurity in the U.S. in 2021 with higher rates in the Southeast region. While significant advancements have occurred in better understanding food insecurity, its relationship with diabetes has yielded mixed results. Due to these inconsistent findings, a better understanding of how food insecurity may be associated with diabetes, particularly at the county-level, can improve population health and well-being. This study advanced such an area by undertaking a cross-sectional observational study for the Southeastern region of the U.S. (e.g. Alabama, Arkansas, Mississippi, and Tennessee with 319 counties as the unit of analysis), an area disproportionately impacted by both food insecurity and diabetes. The overall design applied the socio-ecological perspective within a multiscale geographically weighted regression framework or MGWR. MGWR is a recent development and a more advanced approach which allowed adjustments for geographic space and scale. Results showed food insecurity (estimate of 0.28) was positively associated with diabetes but the relation varied in magnitude and significance across space (range of −0.10 to 1.19). That is, food insecurity exhibited a strong, positive association for northwestern Arkansas, a mild association for central Mississippi and Tennessee, and a weaker association for southern Alabama and eastern Tennessee. Households without internet access also exhibited a positive association (estimate of 0.15), as did convenience stores per thousand (estimate of 0.12). These findings add value to our understanding of how geographic space and scale matter when examining health. Public health practitioners can recognize such variations when devising targeted interventions for food insecurity and diabetes care.


Food insecurity and diabetes
Food insecurity impacted approximately 11% of households in the United States (U.S.) during 2021 with much higher rates for the Southeast region (e.g.15-17% for Arkansas, Alabama, Mississippi) (Coleman-Jensen et al. 2022).This meant at-least one household member experienced a reduced or skipped meal or was worried food would run out or not last.In formal terms, food insecurity is a household-level economic and social condition of limited or uncertain access to adequate food for a healthy life.To better understand food insecurity and its relation to well-being, this study explored how the spatial distribution of food insecurity was associated with health and environmental attributes within a local context (Gundersen and Ziliak 2015;Garcia, Haddix, and Barnett 2018;Gundersen and Ziliak 2018;Jablonski, McFadden, and Colpaart 2016).And one health attribute which could benefit from greater research pertains to diabetes.Diabetes is another important public health concern for it is a leading cause of mortality both in the U.S. and globally (Williams and Loeffler 2019; Centers for Disease Control and Prevention 2020; Ampofo and Boateng 2020).While significant advancements have occurred in managing and treating diabetes by better understanding the role of social and economic factors (i.e.education, income, race, poverty, unemployment), health access, health behaviours, and insurance status, its relationship with food insecurity has yielded mixed or inconsistent results (Beltrán et al. 2022;Ziso, Chun, and Puglisi 2022;Gucciardi et al. 2014;Hill-Briggs et al. 2021).Also noteworthy, the recent National Clinical Care Commission (NCCC) put forth the importance of managing diabetes as a societal problem, in addition to a biomedical condition, and emphasized the utility of examining environmental and social factors in its report to Congress (Schillinger et al. 2023).Similarly, the current Centers for Disease Control and Prevention (CDC) communication noted the significance of food insecurity on diabetes (Centers for Disease Control and Prevention 2022).Therefore, further research can add value to this body of literature.This study also serves a useful purpose for public health planning by offering another perspective into how food insecurity may be associated with diabetes within a local context by formulating a socioecological regression analysis which modelled variation across both space and scale (i.e.multiscale geographically weighted regression or MGWR).This is a more advanced approach and a recent development in geospatial techniques (Oshan et al. 2019).

Scholarly contribution and conceptual framework
The importance of understanding the covariance of food insecurity and diabetes with its correlates was made salient by Gucciardi and colleagues with a comprehensive screening of 539 articles and a final review of 39 articles (Gucciardi et al. 2014).Their synthesis found food insecurity was a risk factor for diabetes occurrence and self-management (i.e.awareness about learning how to manage the condition with dietary and lifestyle changes).Given the complex relationship between the two, they suggested more research is needed and scholars consider advancing the field by undertaking population-based studies with larger sample sizes.Gucciardi and colleagues concluded by proposing that researchers show an appreciation for multiple perspectives and employ approaches which recognize community-level factors (Gucciardi et al. 2014).This study followed their recommendations by undertaking a geographic examination utilizing a socioecological framework.
Another notable review examined an extensive body of literature from multiple fields and found a relationship between food insecurity and diabetes (Hill-Briggs et al. 2021).These scholars similarly conveyed the need for more research as to better inform policymakers about practical approaches for diabetes management.This work emphasized the need for programs which add food insecurity intervention as part of the clinical management of diabetes.This comprehensive review was also salient because it noted the unequal distribution of food insecurity in the U.S.More specifically, reports show food insecurity was common in the Southeast but less noticeable in the Midwest and Northeast (Feeding America 2022).This review, when combined with another recent study, provided additional support for researching food insecurity and diabetes (Walker et al. 2021).Utilizing a sample of nearly 207,000,000 adults from the National Health and Nutrition Examination Survey (NHANES) from 2003-2016, Walker et al. (2021) ) found food insecurity was associated with higher diagnostic measures for the diabetes diagnosed population.More striking, this association was much larger for the undiagnosed population.These scholars, correspondingly, espoused the need to screen for food insecurity as an additional approach for diabetes care.This finding and the afore-mentioned reviews provided the impetus for the present study's objective: the relationship between food insecurity and diabetes with countylevel health-environmental attributes should be explored with current data to inform public health practitioners about targeted interventions for diabetes care in the Southeastern region of the U.S., an area disproportionately impacted by both.Equally important, there may be a possibility that such a relationship may be nonstationary.That is, the relationship varies across space and scale.Although studies of food insecurity have found spatial variations in its incidence, there is a paucity of information on how the spatial distribution of food insecurity may or may not be associated with the spatial prevalence of diabetes.The present study contributed to this area by employing multiscale geographically weighted regression (MGWR), which was a more advanced spatial modelling approach to account for local variability with food insecurity (main variable of interest) and diabetes (outcome variable).MGWR is an improvement over geographically weighted regression (GWR) because it adjusts for both space and scale.While GWR also creates regression equations for all features (i.e.local model to account for space), the same scale is applied to all the predictor variables.In contrast, MGWR relaxes this assumption of spatial stationarity and allows the variables to operate at different scales.
As for a conceptual framework, this study applied the socio-ecological perspective to examine the association between food insecurity and diabetes (Gundersen and Ziliak 2018;Bronfenbrenner 1991;Carter, Dubois, and Tremblay 2014;Crowe, Lacy, and Columbus 2018).In this widely used framework, the county served as the unit of analysis with the focus on health and environmental factors as associated community-level structures (Bronfenbrenner 1991;Deller, Canto, and Brown 2017).A formal mathematical notation is provided in the Material and Method Section under subheading 2.3.

Study setting
The study area consisted of the following states in the Southeast region of the U.S.: Alabama, Arkansas, Mississippi, and Tennessee with the county as the unit of analysis (n = 319; see Figure 1: Study Area).Estimates from the 2020 Map the Meal Gap show food insecurity approximated 14-17% for these states with several counties approaching 25% (Feeding America 2022).Unlike the Midwest or Northeast, food insecurity has been more common in the Southeast region of the U.S. In addition to a considerable food insecure population, these states also maintained a significant percentage of the population with diabetes while a few of these states had among the highest rates in the nation, thereby supporting this region as a suitable candidate to study (Centers for Disease Control and Prevention 2020; Sharma 2014).Other studies have correspondingly found a greater burden of diabetes in the Southeast region of the U.S. Additional states (e.g.Florida, Georgia, and Louisiana) were considered but were excluded due to missing data for several of the counties.

Data sources and measures
Data were aggregated from multiple sources since no one dataset contained all the variables with the county as the unit of analysis.Diabetes, the dependent variable, was defined as the prevalence of diagnosed Type 1 and Type 2 diabetes among adults aged 18 and older.It was obtained from the 2019 Behavioral Risk Factor Surveillance System (BRFSS) and was measured at the county-level.The primary independent variable corresponded to food insecurity.Food insecurity was defined as the percentage of households having difficulty providing adequate food for family members at the countylevel for 2019.These data were collected from the Current Population Survey from the Bureau of Labor Statistics and modelled utilizing the Core Food Insecurity Model by Map the Meal program.One advantage of employing this measure was Map the Meal applied five-year averages from the American Community Survey (ACS) when modelling food insecurity, thereby limiting year-to-year fluctuations when examining smaller geographies.An important technical point to note about Map the Meal methodology pertains to the food insecurity model, which was developed in two stages with the following variables: unemployment, poverty, median income, percent Hispanic, percent Black, homeownership percentage, and disability at the statelevel (first step) and then another regression with these same variables at the county-level (second step).As such, the above-listed variables were not included in the present analysis since they were already formulated in the food insecurity estimate and adding them would introduce multicollinearity into the model (Gundersen et al. 2022).
The 2019 BRFSS was also utilized to obtain the healthoriented independent variables: percentage adult smoker and percentage obese.Percentage adult smoker measured the population aged 18 and older that reported smoking cigarettes every day or some days for each county.Percentage obese was one of the four categories (e.g.underweight, normal, overweight, obese) calculated from the body mass index or BMI and this variable measured the percentage of the population with a BMI>30 kilograms/metres 2 for each county.Smoking was added because previous and current reviews noted greater tobacco use with an increased risk of diabetes and a similar relationship has been found for obesity (Maddatu, Anderson-Baucum, and Evans-Molina 2017;Ioannidis 2008;Lazar 2005).That is, previous research has found these variables to be associated with diabetes.The primary care physician ratio was obtained from the Area Health Resource Files for 2019.It measured the number of individuals served by a physician for each county.The mental health provider ratio was obtained from the Centers for Medicare and Medicaid Services (CMS) National Provider Identifier Registry for 2021.It measured the number of individuals served by a mental health provider for each county.Uninsured was collected by the U.S. Census Bureau (Small Area Health Insurance program) and it measured the percentage of the population under age 65 without any type of health insurance at the county-level for 2019.These three variables were added to account for availability of care, another important determinant for health (Fields, Bigbee, and Bell 2016).
The environmental-oriented independent variables corresponded to PM 2.5 , percentage of population with access to exercise opportunities, percentage of households without internet access, convenience stores per thousand persons, fast food restaurants per thousand persons, and rural value.To expound, PM 2.5 measured ambient air quality for particle pollution smaller than 2.5 micrometres for each county for 2018 and this data was collected by the NASA Applied Sciences Program.This variable was added because a recent Lancet Planetary Health study found greater attributable burden of incident diabetes in areas with higher air pollutant for both the U.S. and globally (Bowe et al. 2018).Access to exercise measured the percentage of the population residing within a half mile of a park or one mile of a physical activity centre or recreational facility.It was measured at the county-level for 2019.This variable was added because a recent systematic review and analysis found areas with greater recreational facilities and additional green spaces (i.e.parks) were associated with lower diabetes risk (Den Braver et al. 2018).Percentage of the households without internet access was measured at the county-level for 2020 and it was collected as a survey response in the five-year American Community Survey (ACS).This variable was included because recent search suggests broadband access and internet availability should be considered as social determinants of health (Bauerly et al. 2019;Holtz 2020).Equally important, households with limited broadband internet access have been associated with lower levels of patient engagement when managing electronic healthcare information and resources (Rodriguez et al. 2020).In sum, areas which lack internet access can result in diabetic patients receiving limited engagement with health providers by means of video conferencing, telemonitoring, and e-consultations, approaches which have delivered positive outcomes when compared with standard care (Agarwal, Simmonds, and Myers 2022).Convenience stores per thousand and grocery stores per thousand measured available food stores at the county-level and were obtained from the 2017 Food Environment Atlas/USDA.These variables were added because several cross-sectional and longitudinal studies have found food access and availability as associated factors for diabetes (Hill-Briggs et al. 2021).The afore-mentioned health-environment oriented variables were also selected based on the comprehensive reviews and recent research reports (Gundersen and Ziliak 2018;Jablonski, McFadden, and Colpaart 2016;Gucciardi et al. 2014;Leonard et al. 2020).

Multiscale geographically weighted regression
The spatial data used in this study violated major assumptions of the global regression framework: (1) residuals displayed spatial autocorrelation and (2) the independent variables did not present spatial stationarity.Based on this preliminary analysis, global ordinary least squares or OLS regression estimates would not be representative of the study area (Fotheringham, Brunsdon, and Charlton 2003).In contrast, multiscale geographically weighted regression (MGWR) relaxed these assumptions by creating local models (i.e. one for each county) and such an approach allowed adjustments for space and scale.As stated in the Introduction, another key contribution of this study was applying this more advanced spatial technique.To expound, OLS regression does not account for variation in space because only one equation is created for the entire study area (i.e.global).Geographically weighted regression (GWR) overcomes this limitation by employing a weighted least squares procedure to estimate equations for all features in the study area (i.e.local).However, GWR does not adjust for the bandwidth between the different variables.That is, the same spatial scale is applied to all variables.The more innovative MGWR approach overcomes both these limitations by creating local regression equations which can also vary in scale.Stated differently, the predictor variables can operate at different scales with some having an influence within a larger space and others within a smaller space.To provide support for this framework, formal checks for model performance and comparison were undertaken with an explanation provided in the Results section under subheading 3.3.
In terms of a mathematical formulation, MGWR can be written as: where y i was the measure of diabetes while bwj represented the bandwidth (i.e.adjustment for scale) and (u i ,v i ) denoted the centroid for each county.βx ij was the estimated parameter for variable j with respect to county i, β 0 (u i ,v i ) was the intercept, and ε i was the random error term.

Preliminary steps and model formulation
After undertaking the preliminary measure of exploratory data analysis, the next step was to employ MGWR by selecting the weighting function by (1) considering the type of kernel (i.e.Gaussian or bisquare), (2) determining whether the kernel would be adaptive or fixed, and (3) applying a selection method to determine the bandwidth (or distance) for the kernel.The bi-square weighting function with an adaptive kernel was employed because the distribution of the observations varied across space.Specifically, counties were smaller and closer together in the northeast (i.e.Tennessee) relative to the southeast and southwest region (i.e.Arkansas and Mississippi).Due to this uneven distribution, the use of an adaptive kernel allowed adjustments to be made so counties closer to i would have greater influence on the estimation of β bwj (u i ,v i ).For this study area, a dynamic process of adjusting weights can be more accurate since it reduces the variability in estimates from areas of low/high density (Van Kerm 2003).Finally, a process that minimized the corrected Akaike Information Criteria (AICc) was used to determine the bandwidth (i.e.optimal kernel size).This was accomplished by using an iterative process which successively narrowed the range of values inside the optimal bandwidth by searching and then comparing the lowest obtained AICc scores.These analyses were conducted using Stata V15, GeoDa, and MGWR2.2 while the mapping exploration was performed using QGIS (StataCorp 2015; QGIS Development Team 2019).

Descriptive analysis
The mean for food insecurity ranged from 13% to 20% and it was the highest for Mississippi and the lowest for Tennessee (see Table 1).The choropleth map revealed greater detail by showing individual values at the countylevel while the cluster map uncovered adjacent areas for high and low values for food insecurity (see Figures 2 and  3).To expound, the choropleth map represented a thematic map for the variable of interest where the intensity of the colour scheme signified areas with greater values.The Jenks natural breaks classification was utilized since this optimization method allowed for visual appeal (i.e.easier map reading) by minimizing within-group variance while maximizing between-group variance.For example, some counties exhibited much higher values for food insecurity (i.e.22-36%) near the southern portion of Alabama.Higher values were also found for counties bordering eastern Arkansas with western Mississippi, as depicted by a more intense or darker colour scheme.These same areas also reflected a cluster of high values.To expound, the local indicators of spatial autocorrelation (LISA) cluster map provided a visual representation of high (or low) value counties next to other high (or low) value counties, as well as high-low, low-high, and not statistically significant counties.For example, adjacent counties with food insecurity percentages exceeding 25% were depicted with the red colour.In contrast, a cluster of low values was noted for central Tennessee and this was depicted with the blue colour.
The mean for diabetes ranged from 14% to 16%.The average for diabetes estimated the highest for Alabama and Mississippi at nearly 16%.Unlike food insecurity,  a clear pattern for high or low values was not evident.Instead, the choropleth map showed several counties throughout the study area with high values (i.e.above 25%).That is, a high percentage with diabetes in several counties in Alabama, Arkansas, and Mississippi.In contrast, the LISA cluster map did not uncover many areas with a collection of high values next to other high values for diabetes (only three counties in southern Alabama) but reflected a collection of low values next to other low values for the northern portion of central Tennessee.In terms of health-oriented variables, the overall mean for the region approximated 38% for obesity and 23% for smoking.In terms of environment-oriented variables, PM 2.5 was consistent throughout the region with an estimate of 10.0.Consistent estimates for the region were also found for households without internet access at 29%.Population with access to exercise exhibited greater variation with Tennessee being the highest at 53% and Mississippi being the lowest at 40%.Not surprising given the focus on the Southeast region of the U.S., the rural value approximated at the higher range at 0.67.

Regression analysis
The OLS model provided baseline estimates (see Table 2).Results showed food insecurity (0.16, p ≤ 0.01), obesity (0.26, p ≤ 0.01), smoking (0.15, p ≤ 0.01), households without internet (0.15, p ≤ 0.01), and convenience  stores per thousand (0.16, p ≤ 0.01) were important considerations.Although useful, this formulation did not adjust for space and scale of the health-environment attributes resulting from local variation whereas the MGWR model accounted for this (see Table 3).MGWR results suggested food insecurity was positively associated with diabetes with a mean estimate of 0.28.In formal terms, a one percentage point increase in food insecurity was associated with a 0.28% point increase in diabetes.In terms of variation across geographic space, food insecurity values ranged from −0.10 to 1.19.Capturing this local context was evident with the MGWR estimates map, which revealed the association of food insecurity with diabetes was notable in particular regions of the study area (Figure 4).For example, food insecurity approximated higher for the northern areas of Arkansas (darker shaded areas).A moderate association was found for central Mississippi and a weaker association for counties in northeastern Tennessee (lighter shaded areas).
Given the importance of conveying information by employing a bivariate map, Figure 4 should be explained in greater detail.This map depicted both the MGWR estimates and significance for food insecurity in one map.With this simplified visual depiction, the reader need not compare two, side-by-side maps to determine which areas were impacted by food insecurity.Results show distinct regions where food insecurity was strongly associated with diabetes.This was reflected with the 3 × 3 matrix, which depicted all nine combinations with nine distinct colours for the estimates and significance (i.e. three categories for the estimates and three categories for significance for a total of nine possibilities).For example, areas with the blue colour shading (top right corner of the matrix) reflected estimates in the range of 0.59 to 1.19 with significance values less than 0.01.A case in point for this would be the large area extending from northwest to southeast Arkansas.Areas with the grey colour shading (bottom left corner of the matrix) reflected estimates in the range of −0.10 to 0.15 with significance values greater than 0.05 but less than 0.10.A case in point for this would be the large area in southern Alabama.While the bivariate map was useful, one should briefly make note of the adjusted R-squared map (Figure 5) to better understand how much of the variation was explained by the model for various areas.The largest values were found for southern Mississippi and southern Alabama.More specifically, nearly half of the counties in each state approximated between 0.43 and 0.47.Overall, the values ranged from 0.31 to 0.47 with lower values for the northern counties and higher values for the southern counties.
MGWR analyses also suggested health-oriented factors played a role when explaining diabetes.A positive association was noted for obesity and smoking.Specifically, a one percentage point increase in obesity was associated with a 0.18% point increase in diabetes.The association for smoking estimated higher at 0.24 with a minimum value of 0.17 and a maximum value of 0.31.Finally, environment-oriented factors were associated with diabetes.For example, an increase in households without internet access reflected a positive association while an increase in the percentage of the population with access to exercise opportunities revealed a negative association.Lastly, an increase in convenience stores per thousand persons reflected greater diabetes with a mean estimate of 0.12, a minimum of −0.08, and a maximum of 0.28.

MGWR performance and comparison
To ensure MGWR was a suitable framework, several diagnostic measures were evaluated.These measures allow researchers to select between regression frameworks (i.e.OLS, GWR, MGWR).The adjusted R-squared, which is widely used, quantifies how much of the variation is explained by the model after adjusting for the number of predictors with a value closer to one signifying a better fitting model.For OLS, the adjusted R-squared approximated 0.25.For MGWR, the value was 0.37.While the MGWR framework was an improvement, the value was slightly low and should be interpreted with care.Such moderate values are not uncommon in social science research given the complex nature of the variables.Two other measures were examined: the Residual Sum of Squares (RSS) and the corrected Akaike Information Criterion (AICc).The RSS is calculated as the sum of the squares of the residuals (i.e.difference between the value of the dependent variable and the predicted value for each observation) and it provides a measure of the unexplained variation.When comparing models, a lower RSS value signifies model improvement.The RSS estimated 230.31 for OLS and was much lower at 174.94 for MGWR.The AICc provides another measure of the estimation error of the model.A lower score signifies model improvement.The AICc estimated 832.95 for OLS and 812.65 for MGWR.In addition to these diagnostic measures, MGWR presented other advantages over both OLS and GWR by better modelling spatial heterogeneity, diminishing collinearity, and reducing bias in the parameter estimates (Schillinger et al. 2023).For example, the local test of variation offered evidence to further support MGWR.A Monte Carlo test for spatial variability with 1,000 simulations found local variation for all the variables except access to exercise (p = 0.08), convenient stores (p = 0.06), and obesity (p = 0.05).Stated differently, MGWR provided results which accounted for differences in scale by creating equations with local variation for the predictor variables and this would not be possible with OLS or GWR.
In addition to the bi-square adaptive kernel, a fixed Gaussian kernel was implemented for the weighting scheme.With this approach, a value equal to one was assigned to the regression feature (feature i) and this value gradually decreased for surrounding features (features j).The MGWR Gaussian kernel results were similar to the adaptive kernel model except for slightly lower values for access to exercise opportunities and air particulate matter and slightly higher values for obesity and households without internet access.The final specification test was an examination of multicollinearity.In general terms, multicollinearity exists when a linear relationship exists between two or more independent variables.A test of variance inflation factors (VIF) did not find evidence of multicollinearity with values for a few variables approaching two, several variables approximating between two and three, and few in the range of three to four (e.g.3.15 for food insecurity, 4.16 for households without internet access, and 3.01 for obesity).Values which are near or above five suggest multicollinearity.

Discussion and concluding remarks
The percentage of the population with food insecurity was strongly associated with diabetes for several counties in Arkansas and Mississippi and less so for Tennessee (see Figure 4).This finding suggested local variation in food insecurity had a distinct impact on diabetes and affirmed the importance of capturing the local context.The local context should also be considered because it may explain why some studies have found a strong association between food insecurity and diabetes while others have found a weak relationship.That is, the association between food insecurity and diabetes was variable throughout the study area and reflected an estimate near zero in some regions but over one in others.As to why food insecurity was positively associated with diabetes, one explanation for this is communities which cannot consume nutritionally rich foods on a regular basis have difficulty maintaining and improving health, which could have an adverse impact on diabetes care (Gucciardi et al. 2014;Hill-Briggs et al. 2021).In counties with greater food insecurity, lack of consistent access to healthy food options exacerbates diabetes and other comorbidities.Additionally, increased food insecurity detracts from diabetes selfmanagement and the stressors associated with not having adequate food may explain the greater burden for diabetes (Hill-Briggs et al. 2021).This can be better understood by referring to a novel study which examined the role of stressors with food intake (Keenan et al. 2021).Specifically, household food insecurity modelled by considering distress and the role of distress on body mass index (BMI) and diet quality.Results suggested viable pathways from household food insecurity to diet quality and distress with distress leading to negative health outcomes.These scholars reiterated the significant role of dietary intake in populations with high food insecurity.Based on the results of the present study and findings from the work by Keenan et al. (2021), local public health officials should advance measures which alleviate food insecurity and doing so could also reduce stressors associated with diabetes management and care.As a case in point, public health professionals can utilize the MGWR estimates and significance map to conduct a community food insecurity assessment in the Capital River area (Mississippi), Middle Tennessee, Ozark Plateaus (Arkansas), and the Piedmont Upland (Alabama; see Figure 4).Such an assessment serves as an important first step because it could identify probable causes and potential interventions for food insecurity and diabetes management.A recent review article in the Proceedings of the Nutrition Society noted food insecurity interventions (e.g.food banks, community shops, nutrition assistance) may be unique to an area due to the availability of local resources.Another recent article in the Annual Review of Public Health emphasized the importance of aligning food insecurity policies and programs with community needs and resources (Loopstra 2018;Seligman and Berkowitz 2019).This could be a future direction of research for the afore-mentioned areas.That is, a detailed analysis of food insecurity, its consequences, and its relation to diabetes by studying smaller geographic units (i.e.communities) for the aforementioned regions.
The present study also found counties with a greater number of households without internet access were associated with higher levels of diabetes.This association existed throughout the study area with several counties approaching the second highest estimated coefficient from the model with a value of 0.43.This estimate remained robust even after accounting for health-oriented and environment-oriented attributes.This was an important finding because lack of internet access may serve as a barrier for quality health care.For example, a recent study found a positive association between census tracts/neighbourhoods with internet access and use of patient portals for electronic transmission of information relating to health monitoring and health information (Perzynski et al. 2017).Similar results were noted from a national study which examined approximately 43 million patients across 15,191 census tracts (Rodriguez et al. 2020).Another recent systematic review examined 97 studies (from a possible list of 5,225) and found internet access remained as a consistent factor for (1) intention to use, (2) subjective use, and (3) objective use for electronic personal health records (ePHR) for health management and care (Abd-Alrazaq et al. 2019).In counties with limited internet access, patients and providers may be restricted in sending and receiving electronic health information and this has the potential to adversely impact diabetes care and management.In rural and remote areas of the U.S. South, limited internet access can hamper the ability to effectively deliver telehealth or coordinate telemonitoring of diabetes care.
This study offered key insights which add to the scholarship in food insecurity and diabetes.For one, employing MGWR to better understand spatial relationships and how such relationships can vary in magnitude and significance with associated factors has not been examined.Such an approach is more accurate because different social and environmental factors will exert varying levels of influence on the outcome variable across the study area.For example, access to exercise and obesity did not show much local variation while internet access and smoking exhibited a much wider range.By relaxing the assumption that all predictor variables operate at the same scale, MGWR utilizes a more realistic approach.Recognizing the extent to which different social and environmental factors impact diabetes can be useful when deciding upon the type and level of public health intervention needed for specific areas.Second, this study advanced the field by applying recommendations from leading scholars by examining micro-geographies (i.e.regional scope) of food insecurity (Gundersen and Ziliak 2018).Further research is needed in this area given the complex nature of food insecurity with respect to place and health.In terms of limitations, this study cannot be generalized to other regions of the U.S.While food insecurity remains high for other states in the Southeast region, the patterns and associations with respect to diabetes may be different and other health-environment attributes can play a larger role.This may also partly explain why the adjusted R-squared was a low value.Another limitation was this study did not account for price fluctuations in food.Although the food insecurity measure was derived by utilizing five-year averages from the American Community Survey, thereby limiting year-to-year fluctuations, dramatic price changes in food items could exacerbate food insecurity, particularly for lowincome areas.Two additional limitations refer to the modifiable areal unit problem and ecological fallacy (Buzzelli 2020).While MGWR has the capability to reduce the areal problem by adjusting for both space and scale, one should exercise caution when drawing inferences from aggregated data based to the scale of the unit of analysis.One should also be attentive when drawing inferences with aggregated data as to not commit the ecological fallacy of drawing conclusions about individuals when the unit of analysis may be the community or county.Lastly, the MGWR formulation did not attempt to model causality and results should not be interpreted in such a manner.In closing, this study found a positive association between food insecurity and diabetes but the association varied in magnitude and significance across the study area.Public health practitioners and policymakers can utilize Figure 4 (i.e.bivariate map) to devise targeted interventions at the county-level.Several counties with high rates of food insecurity and diabetes in the southern region can benefit from local food programs (Fleischhacker, Parks, and Yaroch 2019;Holben and Marshall 2017;Rollins et al. 2021).

Figure 1 .
Figure 1.Study area for Southeast region of the U.S.

Figure 2 .
Figure 2. Choropleth maps for food insecurity (top) and diabetes (bottom) depicting percentages for each county.

Figure 3 .
Figure 3. Cluster maps for food insecurity (top) and diabetes (bottom) with high areas in red and low areas in blue.HH signifies high values next to high values; HL signifies high values next to low values; LH signifies low values next to high values; LL signifies low values next to low values.

Figure 4 .
Figure 4. MGWR estimates and significance for food insecurity (main variable of interest).Darker shaded areas represent counties with larger estimates and greater significance.