4.1 Cohort demographics

The final cohort included \(n = 924\) people, \(63\) of whom had indications of a type-2 diabetes diagnosis in the EHR (i.e., were treated as cases). Our estimated diabetes prevalence of \(7\%\) (95% CI: \(5\%, 8\%\)) was lower than the NC state prevalence of \(12\%\) (American Diabetes Association 2024). The cohort was \(61\%\) female, \(71\%\) White race, \(19\%\) Black race, and \(6\%\) Hispanic with a median age of \(49\) years old.

These patients were chosen to be representative of people routinely engaging in care at the Atrium Health Wake Forest Baptist healthcare system, but not beyond that. For example, we could not assume that they represented the state of NC or even the city or county where the hospital is located. This selection bias is a common challenge in analyzing EHR data since patients had to have access and ability to receive care in order to be considered (Bower et al. 2017). Using the 2020 U.S. Census population estimates (U.S. Census Bureau 2020a, b) and 2023 American Community Survey (U.S. Census Bureau 2023), we compared the patients in our case study to the entire state of NC. In addition to its lower prevalence of diabetes, we found the sample to have more females (\(61\%\) versus \(51\%\)), more people of White race (\(71\%\) versus \(62\%\)) and fewer people with Hispanic ethnicity (\(6\%\) versus \(11\%\)). The cohort was also older on average, with a median age of \(49\) versus \(39\) years old. The proportion of Black patients in our sample (\(19\%\)) was comparable to the state population (\(21\%\)).

4.2 Descriptive statistics of the food environment

The median straight-line proximity for patients in this study was \(1.14\) miles, compared to a median map-based proximity of \(1.70\) miles. The distributions for both proximity measures were right-skewed (Fig. 2a). Still, in addition to being farther, on average, map-based proximity had greater variability. Furthermore, the relationship between the straight-line and map-based proximities was very strong and linear (Fig. 2b) with a correlation of \(R = 0.97\). Building a simple linear regression model between them, \(\widehat = 0.25 + 1.30 X^*\), revealed that (i) straight-line proximity of \(0\) miles was expected to have map-based proximity of \(0.25\) miles and (ii) for every additional mile of straight-line distance, map-based proximity increased by \(1.30\) miles on average.

Fig. 2

Data visualizations comparing error-prone (straight-line) and error-free (map-based) proximity to healthy foods. In a), the distributions of both unbinned proximity measures are shown to both be right-skewed, but map-based proximity was farther, on average. In b), the relationship between unbinned straight-line and map-based proximity is seen to be strong and linear, but with straight-line systematically underestimating the latter (especially at higher values). For comparison, the dashed line denotes the line of equality (i.e., where straight-line and map-based proximity are equal.) The heatmap comparing binned proximity in c) shows that only \(450\) patients (\(49\%\)) had straight-line and map-based values that fell into the same category (bold squares), leaving \(474\) (\(51\%\)) patients’ measures in disagreement (i.e., misclassified)

Additionally, just under half of the patients in our study (\(n = 450\), \(49\%\)) had binned straight-line and map-based proximity in agreement (Fig. 2c). Among patients whose binned proximity was misclassified (i.e., they were categorized differently based on straight-line versus map-based proximity), most fell one map-based category higher (e.g., between \(2\)–\(3\) miles versus \(1\)–\(2\) miles). Still, some patients’ food access was multiple categories worse than straight-line distances indicated, like the seven with moderate straight-line access (\(2\)–\(3\) miles) but the worst map-based access (more than \(5\) miles).

We also compared patients’ neighborhood food insecurity based on the survey estimate versus the lower and upper bounds of their \(95\%\) CIs. The distributions of the three different food insecurity prevalences were all symmetric (Fig. 3a), but the upper bounds had the most variability. Based on the survey estimates, \(16.1\%\) of adults in patients’ neighborhoods, on average, reported being food insecure in the last year. This average was much higher than the recent statewide estimate of \(10.9\%\) (North Carolina Department of Health and Human Services 2023), as were the average rates based on each patient’s lower (\(14.5\%\)) and upper bounds (\(18\%\)). At a patient level, \(733\) people in our study (\(79\%\)) had a survey estimate of neighborhood food insecurity that was higher than the state’s estimate of \(10.9\%\) (Fig. 3b). While some of these patients’ CIs contained the state value (i.e., their lower bound was smaller than \(10.9\%\) and their upper bound was larger), the data suggested that \(669\) people in the study (\(72\%\)) resided in neighborhoods with more severe food insecurity than was typical for NC, while fewer (\(n = 172\), \(19\%\)) lived in neighborhoods with less severe food insecurity.

Fig. 3

Data visualizations comparing survey estimates to the upper and lower bounds of their \(95\%\) confidence intervals for neighborhood prevalence of food insecurity. In a), the distributions from all data sources are shown to be relatively symmetric – as expected, based on how they were estimated – and to have substantial overlap between them. In b), the survey estimate and \(95\%\) confidence interval for each patient’s neighborhood prevalence of food insecurity are displayed, from smallest to largest survey estimate, against the state’s prevalence of \(10.9\%\). The heatmap in c) indicates that \(555\) (\(60\%\)) of patients fall into the same categories based on the survey estimates as their lower bounds (bold squares). In d), \(456\) patients (\(49\%\)) are seen to be classified the same based on their survey estimates and upper bounds (bold squares)

After binning the different food insecurity values, we compared patients’ categorizations based on their survey estimate against those based on their CI bounds. The majority (\(n = 555\), \(60\%\)) fell into the same category based on their survey estimate and lower bound (Fig. 3c), while fewer (\(n = 456\), \(49\%\)) fell into the same category based on their survey estimate and upper bound (Fig. 3d). Patients who were differently classified fell only one category below or one category above their survey estimate, according to their lower and upper bounds, respectively. Note that we do not use the term “misclassified” here as the true categorization was unknown.

Only \(228\) patients (\(25\%\)) fell into the same category for neighborhood food insecurity regardless of which value we used. The remaining \(696\) patients (\(75\%\)) contributed different information to the health disparities models depending on which of the three available values based on the survey estimate, lower bound, or upper bound. Of these \(696\) patients, \(555\) (\(80\%\)) contributed two different binned food insecurity values, while \(141\) (\(20\%\)) contributed three different ones. That is, the data supported between two and three categorizations per patient, depending on their margin of error.

4.3 Disparities based on proximity to healthy foods

Higher densities and relative densities of healthy food options (full-service restaurants, grocery stores, direct farm sales, or supermarkets) have been associated with a lower prevalence and less risk of incident type-2 diabetes (Ahern et al. 2011; Kanchi et al. 2021). On the other hand, the proximity to these same healthy food options was not found to be associated with diabetes (Lotspeich et al. 2025a). Thus, further studies are required to help understand the connections between access to healthy foods and disease outcomes, particularly through an equity lens as we compare the health statuses of people in low- and high-access communities.

In the following subsections, we discuss our findings using various metrics to evaluate potential disparities in type-2 diabetes prevalence based on proximity to healthy foods and compare results based on straight-line (error-prone) versus map-based (error-free) measures. Ultimately, despite the differences between straight-line and map-based values discussed in Sect. 4.1, all methods led to the same inconclusive evidence about the presence of a disparity in type-2 diabetes between patients with better versus worse access to healthy foods. Still, many point estimates differed between error-prone and error-free measures, particularly with the binned rate ratios, and map-based values often led to better statistical efficiency (i.e., narrower CIs).

4.3.1 Rate Ratios

Based on binned straight-line distances, diabetes prevalence was \(7\%\) (95% CI: \(4\%\), \(9\%\)) among patients living up to 1 mile from their nearest healthy food retailer. Compared to this group, patients with proximity of \(3\)–\(4\) miles (\(\widehat} = 1.49\); 95% CI: \(0.67\text3.29\)) or more than \(5\) miles (\(\widehat} = 1.44\); 95% CI: \(0.58\text3.59\)) had much higher prevalence (Fig. 4). For proximity categories of \(2\)–\(3\) miles and closer, the estimated RRs were close to one. In the 4–5 mile category, there were zero diabetes cases. Despite larger RR estimates for some of the farther categories, all 95% CIs contained the null value of one, capturing the possibility of equal diabetes prevalence.

Fig. 4

Type-2 diabetes prevalence rate ratios (with 95% confidence intervals) based on unbinned and binned proximity to healthy foods (top) and neighborhood prevalence of food insecurity (bottom). The unbinned estimates come from log-binomial regression models, while the binned estimates are calculated using sample proportions in each category

Diabetes prevalence was slightly lower among patients with up to \(1\) mile proximity based on map-based distances, at \(6\%\) (95% CI: \(3\%\), \(9\%\)). Different bins stood out as having higher diabetes prevalence than this group based on map-based proximity than did with straight-line. Patients living between \(1\)–\(2\) and \(4\)–\(5\) map-based miles from their nearest healthy food retailer had \(1.38\) (95% CI: \(0.73\), \(2.60\)) and \(1.88\) (95% CI: \(0.83\), \(4.25\)) times, respectively, the diabetes prevalence of those living up to \(1\) mile away. One category, patients with map-based proximity between \(2\)–\(3\) miles, had a lower prevalence than the referent group (\(\widehat} = 0.66\); 95% CI: \(0.27\text1.78\)), which was not seen with the straight-line distances. Despite all of the map-based RR estimates falling above or below the null of one, all of the 95% CIs still passed through it.

Thus, none of the binned rate ratios, based on either distance calculation, were found to be significant. We have insufficient evidence to conclude that there were disparities in diabetes prevalence between patients with better versus worse proximity to healthy foods (binned), regardless of which distance measure we used. For additional details, including the diabetes prevalence in each bin, see Supplemental Table S1.

Based on the log-binomial regression models using the original unbinned data, we again did not find evidence of an association between proximity to healthy foods and changes in diabetes prevalence rates according to either straight-line (\(\widehat} = 1.00\); 95% CI: \(0.86\text1.14\)) or map-based (\(\widehat} = 1.01\); 95% CI: \(0.91\text1.12\)) distances. For example, based on the map-based estimates, we expect that diabetes prevalence among patients whose proximity to healthy foods was one mile worse (i.e., one mile farther) would be \(1.01\) times, or just \(1\%\) larger, than the prevalence among patients whose proximity was one mile better (i.e., one mile closer). While the log-binomial model estimates based on either distance metric were similar and inconclusive, the 95% CI based on map-based proximity was \(25\%\) narrower. As such, the more accurate distance measurements still led to better statistical precision here. However, based on either rate ratio approach (binned and model-free or unbinned and with a model), we had insufficient evidence to conclude that there was an access-based disparity in diabetes prevalence.

4.3.2 Relative index of inequality

The relative index of inequality compared the type-2 diabetes prevalence between patients with the farthest (more than \(5\) miles) and nearest (up to \(1\) mile) binned proximity to healthy foods. Based on straight line distances, the prevalence of diabetes for patients who lived more than \(5\) miles from their nearest healthy food retailer was expected to be \(\widehat} = 1.12\) times that of patients who live within \(1\) mile of their nearest healthy food retailer. However, with a 95% CI that passes through the null (i.e., the possibility that the true RII \(= 1\) and the population diabetes prevalence among patients with the worst and best access is the same) we have insufficient evidence to conclude that there was an access-based disparity based on straight-line proximity (95% CI: \(0.49\text2.75\)). Our estimate based on map-based distances was similar (\(\widehat} = 1.10\)), and while its 95% CI was again narrower (95% CI: \(0.45\text2.64\)), it still crossed through zero. Thus, this second method left us with the same inconclusive evidence about an access-based disparity in diabetes prevalence as the RRs did.

4.3.3 Concentration curve and concentration index

When ranking patients by binned proximity, the concentration curve based on straight-line proximity to healthy foods fell closely along the line of equality, aside from a small initial bump within the first \(10\%\) of patients (Fig. 5a). Based on binned map-based proximity, the concentration curve was slightly above the line of equality around \(25\%\) of patients but dropped below by \(50\%\) before deviating above again. Concentration indices of very nearly zero for both straight-line (\(\widehat_G = -0.02\); 95% CI: \(-0.15\), \(0.11\)) and map-based (\(\widehat_G = -0.01\); 95% CI: \(-0.15\), \(0.12\)) distance calculations suggested that there was not a disparity in diabetes cases based on binned proximity to healthy foods. While the map-based concentration curve deviated slightly more from the line of equality, these deviations seemed to cancel each other out, leading to a similar concentration index as the straight-line measures.

Fig. 5

Concentration curves based on binned and unbinned household proximity to healthy foods (top) and neighborhood prevalence of food insecurity (bottom). Patients were ranked from most to least disadvantaged, meaning from the farthest to the closest proximity to healthy foods and from the highest to lowest percent food insecure households. In the legend of each plot, C denotes the concentration index. The dashed lines denote the line of equality (i.e., where no disparity is present such that the cumulative proportion of diabetes cases is equally distributed across patients regardless of their food environment)

There were again some slight deviations from the line of equality when ranking patients based on unbinned proximity and small differences between the straight-line and map-based versions (Fig. 5b). In particular, based on straight-line proximity there were disproportionately fewer cases between \(25\)–\(50\%\) of patients and then disproportionately more between \(50\)–\(75\%\). Small differences between the concentration curves based on straight-line versus map-based distance calculations were also evident. Based on map-based proximity, the concentration curve dipped below the line of equality slightly later, around \(38\%\) of patients, before rising above it at \(70\%\). Still, deviations from the line of equality were not severe enough to suggest a health disparity using either distance measure to calculate proximity. The concentration indices using straight-line and map-based proximities were \(\widehat_W = 0.00\) (95% CI: \(-0.15\), \(0.14\)) and \(\widehat_W = -0.02\) (95% CI: \(-0.17\), \(0.13\)), respectively. These values indicated that the areas between the line of equality and the curves were very small and that the slight deviations noted previously largely canceled each other out. As such, the concentration indices agree that there was not an access-related disparity in the prevalence of type-2 diabetes in this study.

4.4 Disparities based on neighborhood-level prevalence of food insecurity

Most, if not all, of the established relationships between food insecurity and health found in the literature evaluated food insecurity at the individual rather than neighborhood level. People experiencing food insecurity tend to face lower quality diets, particularly since less healthy foods can be less expensive (Mello et al. 2010; Orr et al. 2019), and thus higher risks of diet-related noncommunicable diseases, like type-2 diabetes and obesity (Hernandez et al. 2017; Te Vazquez et al. 2021; Levi et al. 2023). Herein, we investigated whether similar connections held based on the age-adjusted prevalence of food insecurity beyond a patient’s home as well.

In the subsections that follow, we discuss our findings using various metrics to evaluate potential disparities in type-2 diabetes prevalence based on neighborhood food insecurity and compare results based on survey estimates (the model prediction) versus the most extreme estimates that were supported by the data (the 95% CI bounds). Based on the concentration curves and concentration indices, patients living in neighborhoods with a higher prevalence of neighborhood food insecurity were found to bear disproportionately more type-2 diabetes cases than those living in neighborhoods with a lower prevalence. Interestingly, these metrics differed in the severity of this disparity when using the binned survey estimates versus the CI bounds but were almost identical when using the unbinned versions. The rate ratios and relative index of inequality, which quantified different aspects of potential disparities, were all inconclusive.

4.4.1 Rate Ratios

Based on the binned survey estimates, type-2 diabetes prevalence was \(7\%\) (95% CI: \(3\%\), \(10\%\)) among patients living in neighborhoods with up to \(10\%\) food insecurity. This rate was higher for patients with more than \(20\%\) (\(\widehat} = 1.20\); 95% CI: \(0.61\text2.37\)) or \(17.5\)–\(20\%\) (\(\widehat} = 1.22\); 95% CI: \(0.53\text2.79\)) neighborhood food insecurity. It was lower for patients with \(15\)–\(17.5\%\) neighborhood food insecurity (\(\widehat} = 0.88\); 95% CI: \(0.35\text2.16\)) and much lower for those with \(10\)–\(12.5\%\) (\(\widehat} = 0.36\); 95% CI: \(0.08\text1.59\)). For patients with \(12.5\)–\(15\%\) neighborhood food insecurity, the prevalence of diabetes was similar to the least disadvantaged group. With all 95% CIs containing the null value of one, we have insufficient evidence to conclude that any of these between-category differences would hold in the population of all patients (Fig. 4).

By construction, each patient’s survey estimate for neighborhood food insecurity was contained between the lower and upper bounds of their 95% CI. One might expect that the rate ratio estimates based on these ordered values would follow a consistent ordering, as well. While the RRs based on the binned survey estimates fell between those based on the CI bounds for many categories (e.g., more than \(20\%\) and \(12.5\)–\(15\%\)), this trend was not always the case. For example, patients with \(17.5\)–\(20\%\) neighborhood food insecurity had the highest rate ratio based on the survey estimate (\(\widehat} = 1.22\); 95% CI: \(0.53\text2.79\)) rather than the lower (\(\widehat} = 1.08\); 95% CI: \(0.52\text2.72\)) or upper bounds (\(\widehat} = 1.08\); 95% CI: \(0.42\text2.77\)). For additional details, including the diabetes prevalence in each bin, see Supplemental Table S2.

The rate ratio estimates from the log-binomial regression models, fit to the unbinned data, were all greater than one, suggesting that as neighborhood food insecurity increased the expected prevalence of type-2 diabetes increased. The estimates were similar regardless of which food insecurity value was used: the lower bound (\(\widehat} = 1.04\); 95% CI: \(0.99\text1.09\)), survey estimate (\(\widehat} = 1.03\); 95% CI: \(0.99\text1.08\)), or upper bound (\(\widehat} = 1.03\); 95% CI: \(0.99\text1.08\)). For example, based on patients’ survey estimates, we expect diabetes prevalence to change by a factor of \(1.03\) (i.e., to increase by \(3\%\)) for every \(1\%\) higher the prevalence of neighborhood food insecurity was.

While subtle, the rate ratios from the log-binomial models did exhibit the ordering we expected (Fig. 4b), with the largest estimate based on the patients’ lower bound (their most conservative estimate of neighborhood food insecurity) and the smallest based on the patients’ upper bound (their most extreme estimate food insecurity). Also, the 95% CI was narrowest based on the upper bounds, which had the most variability (Fig. 3a). However, with all CIs containing the null (\(\text = 1\)), we did not find evidence of an association between neighborhood food insecurity and diabetes prevalence. Despite some large differences between the RRs based on survey estimates versus their lower or upper bounds, all three values led to the same conclusion. That is, we had insufficient evidence to conclude that there was a disparity in type-2 diabetes prevalence attributable to the prevalence of food insecurity in patients’ neighborhoods.

4.4.2 Relative index of inequality

The relative index of inequality compared the diabetes prevalence between patients living in neighborhoods with the highest (more than \(20\%\)) and lowest (up to \(10\%\)) binned prevalence of food insecurity. Based on survey estimates, the prevalence of type-2 diabetes among patients living in neighborhoods where \(> 20\%\) of people experienced food insecurity in the last year was expected to be \(\widehat} = 1.65\) times that among patients where \(\le 10\%\) of people did (95% CI: \(0.69\text4.05\)). The difference between patients in these two groups was even larger, on average, based on the CI lower (\(\widehat} = 2.26\); 95% CI: \(0.94\text5.57\)) and upper bounds (\(\widehat} = 1.81\); 95% CI: \(0.73\text4.61\)).

With corresponding 95% CIs for all three values capturing \(\text = 1\) (the null), however, we have insufficient evidence to conclude that there was a disparity in the prevalence of type-2 diabetes between patients residing in neighborhoods with the highest and lowest prevalence of food insecurity. In this case, despite large differences between the RII estimates based on the different food insecurity values, using any of the three led to this same conclusion.

4.4.3 Concentration curve and concentration index

When ranking patients by binned neighborhood food insecurity (from highest to lowest prevalence), the concentration curve based on the survey estimate initially deviated above the line of equality, before returning at approximately \(75\%\) of patients (Fig. 5). Based on the upper bound, the concentration curves followed a very similar trend, although it peaked slightly earlier (around \(68\%\) of patients) before dropping down to the line of equality. The concentration curve based on the lower bound increased the steepest from the line of equality (within \(13\%\) of patients) and remained higher than the other two curves, which largely overlapped, until around \(40\%\). It then peaked the earliest (at \(63\%\) of patients), before stabilizing and converging with the line of equality alongside the curves based on the survey estimate and upper bound.

Using the binned versions of food insecurity, the largest concentration index (in magnitude) was for the lower bound (\(\widehat_G = -0.13\); 95% CI: \(-0.27\), \(0.01\)), followed by the upper bound (\(\widehat_G = -0.09\); 95% CI: \(-0.22\), \(0.05\)) and survey estimate (\(\widehat_G = -0.08\); 95% CI: \(-0.22\), \(0.06\)). Interestingly, the lower bound, which was based on each patient’s smallest purported prevalence of food insecurity (i.e., the “best case scenario”), led to the largest concentration index and most severe estimated disparity. Still, these concentration indices were all fairly close to zero, and since the CIs all contained the null of zero, there was insufficient evidence of a disparity in diabetes prevalence based on binned food insecurity.

When ranking patients based on their original unbinned food insecurity instead, the concentration curves followed the same overall visual trend as with the binned version (Fig. 5). However, the curves briefly dipped below the line of equality for patients ranked \(0\)–\(5\%\), which was not seen with the binned curves, before quickly climbing above around \(13\%\) and remaining there until around \(80\%\). Additionally, there was no longer a visual separation between the curves based on different food insecurity values.

These similarities between the three curves were confirmed by the identical concentration indices and CIs (all \(\widehat_W = -0.16\); 95% CI: \(-0.33\), \(-0.01\)) based on the lower bound, survey estimate, and upper bound of neighborhood food insecurity. Notably, the concentration indices were all statistically significant since \(0\) was not contained in any of the 95% CIs. In this case, using the original unbinned values led to concentration indices farther from \(0\) than using the binned ones, and better statistical precision (i.e., narrower CIs). Thus, the unbinned concentration curves suggested a statistically significant disparity in diabetes cases for food insecurity, with more cases concentrated among patients with a higher neighborhood food insecurity. This finding was in line with much of the literature around individual-level food insecurity and health. Yet, binning food insecurity may have smoothed out more granular differences in diabetes cases among the patients.