Economic Conditions, Supermarket Distance, and Modeled Diabetes Prevalence Across U.S. Counties
Rohan Tiwari
21/11/26
Diabetes is one of the most common and expensive chronic diseases in the United States, and the burden is not evenly distributed geographically. The issue of food accessibility in policy discourse often revolves around the concept of distance, namely how far people live from supermarkets. This paper poses a narrower, descriptive question: in the context of U.S. counties, does the modeled prevalence of adult diabetes correlate more with economic factors or with a distance-based measure of poor supermarket accessibility? In an ecological design using publicly available county data from the USDA Food Environment Atlas, we related the modeled 2019 diabetes rate from the CDC to the poverty rate, median household income, food insecurity, and the percentage of the population living far away from a supermarket using correlations, multiple regressions with state-clustered inference, comparison within states, and out-of-sample predictions. There are two limitations inherent in this data that influence our findings. First, methods documentation for these CDC estimates indicates that county diabetes prevalence was estimated with a model that uses county poverty as an input variable, thus, some of the relationship between the latter and the outcome is built into the model; second, the economic variables (from 2021 to 2023) were measured after the outcome (in 2019). Within these constraints, poverty showed the strongest correlation with the modeled diabetes rate (r = +0.44), while the distance measure demonstrated a weak, slightly negative correlation (r = −0.08). A comparison of the variables within states yielded similar results (r = +0.33 and r = −0.01, respectively), and the association between the poverty rate and diabetes was still significant after adjusting for clustering at the state level. Together, the four predictors provided marginally better fit compared to the poverty rate alone (R² = 0.196 versus 0.190) and made little difference in out-of-sample predictions; a simple state-average baseline predicted more accurately than the regression in 199 out of 200 random samples. State-level analysis and obesity rates (which are not modeled from county poverty rates) also placed the economic measures ahead of store distance. These are correlations with modeled estimates, not causal effects, and they involve one distance measure rather than the entire food desert concept, which includes low income.
