Adults with diabetes residing in “food swamps” have higher hospitalization rates

Objective To examine the relationship between food swamps and hospitalization rates among adults with diabetes. Data Sources Blue Cross Blue Shield Association Community Health Management Hub® 2014, AHRQ Health Care Cost and Utilization Project state inpatient databases 2014, and HHS Area Health Resources File 2010‐2014. Study Design Cross‐sectional analysis of 784 counties across 15 states. Food swamps were measured using a ratio of fast food outlets to grocers. Multivariate linear regression estimated the association of food swamp severity and hospitalization rates. Population‐weighted models were controlled for comorbidities; Medicaid; emergency room utilization; percentage of population that is female, Black, Hispanic, and over age 65; and state fixed effects. Analyses were stratified by rural‐urban category. Principal Findings Adults with diabetes residing in more severe food swamps had higher hospitalization rates. In adjusted analyses, a one unit higher food swamp score was significantly associated with 49.79 (95 percent confidence interval (CI) = 19.28, 80.29) additional all‐cause hospitalizations and 19.12 (95 percent CI = 11.09, 27.15) additional ambulatory care‐sensitive hospitalizations per 1000 adults with diabetes. The food swamp/all‐cause hospitalization rate relationship was stronger in rural counties than urban counties. Conclusions Food swamps are significantly associated with higher hospitalization rates among adults with diabetes. Improving the local food environment may help reduce this disparity.

service delivery models aimed at achieving better care at lower costs. 1 Management of diabetes and its complications is highly dependent on dietary intake. The quantity and quality of foods consumed affects one's blood glucose levels, and untreated hyperglycemia in diabetics can lead over time to cardiovascular disease, kidney disease, and infection from damaged blood vessels as well as other emergency complications such as diabetic ketoacidosis and hyperglycemic hyperosmolar syndrome. 2 As such, individuals with diabetes are encouraged to limit their intake of processed carbohydrates, saturated and trans-fatty acids, cholesterol, and sodium. 3 Dietary intake, however, is impacted by the nature of the local food environment. Distance to and density of neighborhood grocery stores, fast food outlets, and convenience stores have been found to be associated with fruit and vegetable consumption and other dietary quality measures as well as with obesity. [4][5][6] Research in this area emphasizes that the availability of both healthy and unhealthy foods can be influential in predicting dietary outcomes, leading some recent studies to focus on the number of unhealthy outlets relative to the number of healthy outlets as a measure of the food environment rather than absolute measures. The term "food swamps" was coined in 2009 by Rose et al 7 to describe those areas with high relative measures, defining them as places in which large numbers of unhealthy energy-dense food offerings inundate or "swamp out" the relatively few existing healthy food offerings. Studies using relative food environment measures as predictors have found that food swamps have modest but significant associations with obesity 6,[8][9][10] and that these associations may perhaps be stronger than those of "food deserts," which are areas in which residents must travel long distances to reach grocery stores. 11 If food swamps are associated with obesity through dietary intake, it is likely that, in addition, adults with diabetes residing in food swamps are more vulnerable and prone to diabetic exacerbations and complications caused by poor dietary intake. If so, this vulnerability may be placing diabetics living in these areas at a distinct disadvantage that exists entirely separate from the health care system and creating a disparity in health outcomes and service utilization.
This analysis assesses the degree to which counties in which large relative numbers of outlets selling energy-dense foods overwhelm healthy food options have significantly higher hospitalization rates among adults with diabetes, controlling for other area health system-related and sociodemographic characteristics. Further, we examine whether or not the food swamp-hospitalization rate relationship varies in urban and rural counties. Urban and rural areas differ markedly in their transportation resources and the types of retail outlets that choose to locate within them. [12][13][14] It is, therefore, possible that the association is stronger in rural areas, where supermarkets and robust public transportation systems that facilitate access are lacking.
Some previous work has studied the association between food environment and diabetes prevalence 8,15,16 and incidence, 17,18 but, to our knowledge, the relationship between food environment and hospital utilization among adults with diabetes has not been assessed. This study adds to the limited research using relative food environment measures in the U.S. context, particularly across multiple geographic areas.

| Outcome
This analysis examines two measures of hospitalization rates among adults with diabetes: rates of all-cause hospitalizations and rates of hospitalizations for ambulatory care-sensitive conditions (ACSC).
All-cause hospitalizations refer to inpatient hospitalizations among adults with diabetes over the age of 20 for any reason over the course of the year. Rates were calculated by summing all admissions with any-listed diagnosis of Clinical Classification Software code 49 ("diabetes mellitus without complication") or 50 ("diabetes mellitus with complications"). Common principal diagnoses for all-cause admissions include septicemia, pneumonia, kidney failure, subendocardial infarction, and osteoarthrosis of the leg. Sums were divided by each county's diabetes prevalence rate as estimated by the CDC BRFSS. All-cause hospitalizations were analyzed to assess potential spillover effects of the local food environments on health care utilization; poor glycemic control among adults with diabetes can result in complications that may initially seem unrelated to diabetes, such as those listed above, and these diagnoses may not be marked specifically as diabetes complications in inpatient records. 22 The ACSC hospitalization rate considers only admissions with a principal diagnosis that meets the AHRQ Prevention Quality Overall ACSC Composite specifications. These diagnoses include diabetes with shortterm complications, diabetes with long-term complications, uncontrolled diabetes without complications, diabetes with lower-extremity amputation, chronic obstructive pulmonary disease, asthma, hypertension, heart failure, dehydration, bacterial pneumonia, or urinary tract infection. This measure was included because these diagnoses are preventable through access to high-quality ambulatory care services 23,24 and exacerbated diabetes often manifests itself in such conditions. 25 Rates are presented as the number of hospitalizations per every 1000 adults with diabetes.

| Main independent variables
Food swamp severity is measured on a continuous scale. Food swamp scores represent the ratio of the number of fast food outlets to the number of grocers in a county, adjusted for population density and average disposable income and standardized so that values fall between zero and ten. Estimates were calculated at the zip code level by BCBSA. Zip codes were allocated to their respective counties by assigning each to the county in which the majority of its population resides. Food swamp scores were then estimated at the county level by weighted averaging based on the proportion of the county population that each zip code contributes, according to the most recent U.S. Census.

| Control variables
Analyses also controlled for several health system-related and sociodemographic variables that are often associated with hospitalizations among adults with diabetes. 26,27 Health systems variables include the average number of comorbidities per hospitalized diabetic patient in the county, the percentage of the diabetic population enrolled in Medicaid, and the percentage of the hospitalizations that were admitted to the hospital via the emergency department.
Sociodemographic variables include the percentage of the county population that is female, non-Hispanic black, Hispanic, and over the age of 65.
Rural-urban categorization was determined according to the methodology used by the U.S. Census Bureau, which places counties into one of three categories based on the percentage of the population that is considered rural as of the 2010 census. Counties are classified as "completely rural" if 100 percent of the population is rural, "mostly rural" if 50-99.9 percent of the population is rural, and "mostly urban" if less than 50 percent of the population is rural. 28

| Statistical analyses
Multivariate linear regression was used to estimate the association between food swamp severity and hospitalization rates (all-cause and ACSC) among adults with diabetes at the county level, controlling for health system-related and sociodemographic covariates. The models included state fixed effects to account for clustering within states and were weighted according to county population so that larger counties contributed more to estimates than smaller counties.
The models follow the following specifications: where the hospitalization rate among adults with diabetes for a county (i) within a state (j) is a function of the food swamp score in county (i) in state (j), a number of health system-related and sociodemographic covariates in county (i) in state (j), an intercept α j for each state (j), and a county-specific error term ϵ. The health system-related covariates include comorbidity burden, Medicaid enrollment, and emergency department utilization, and the sociodemographic covariates include the proportions of the county population that are female, non-Hispanic black, Hispanic, and over age 65. Each of these additional covariates had variance inflation factors under 2.0, indicating that collinearity was not an issue for the final models. Income-related control variables were excluded, as the food swamp variable was already adjusted for average disposable income.
Both models were then stratified by rural-urban category.
To account for potential selection effects, we estimated the final regression models using inverse propensity treatment weights. To do so, we first created a dichotomous variable classifying a county as a food swamp if its food swamp score was above the median and not a food swamp if it was below the median.
Propensity scores were estimated using a number of covariates, including the average comorbidity burden among the diabetic population and the percentage of the population that was non-Hispanic All-cause hospitalization rate ij = 1 (food swamp score) ij black, between the ages of 20 and 45, over age 65, had a college degree, on Medicaid, living in urban areas, and living in poverty, all of which either have been linked to poor food environments and hospitalization rates in the literature or varied substantially by food swamp status in this data.

| Sensitivity analyses
We conducted a series of sensitivity tests to assess the robustness of the main regression model results. First, as a falsification test, 29 sometimes called a negative control, 30 the final model was tested on a clinical outcome that should theoretically be unrelated to diet quality. The outcome chosen was the county rate of hospitalizations with a principal diagnosis of a mood disorder (including bipolar disorder, manic affective disorder, major depressive disorder, and other unspecified mood disorders) among all adults.
Second, to build confidence in the concurrent validity of our food swamp measure, the models were estimated using the Retail Food Environment Index (RFEI) as the predictor. The RFEI consists of the ratio of fast food outlets and convenience stores to grocery stores and supermarkets in an area and has been employed in previous food swamp analyses. 8,9,11 We constructed RFEI estimates using data from the USDA Food Environment Atlas, divided the ratios by county population estimates, and standardized the variable to achieve greater comparability to the food swamp score. These models were also additionally controlled for median income, as the food swamp score is adjusted for income.
Third, to control for socioeconomic status and area deprivation beyond the incorporated adjustment for median disposable income, we included a measure of the percentage of households in a county that are vacant. Many other measures of deprivation were strongly correlated with the food swamp score and would create collinearity among variables when included the model. The vacant homes variable, however, was only moderately correlated and inclusion resulted in variance inflation factors that all remained below 2.0, so it was added to the model to more strongly control for county socioeconomic status.
Fourth, the use of a ratio as a dependent variable may result in spurious correlation between the dependent variable and the predictor if the predictor is correlated with the ratio's denominator but not with the numerator, conditional on the denominator. 31 To rule out such correlation, we decomposed the hospitalization rate variable and ran the linear models regressing the number of all-cause and ACSC hospitalizations (log transformed) on the previously included predictors and the number of adults with diabetes in the county (log transformed).
Finally, to improve causal inference with the cross-sectional data, an instrumental variable approach was attempted. All instruments, including highway exits, which has been used in previous studies of fast food access, 11,32 were unsuccessful. While this instrument works well when obesity is the outcome, it did not satisfy the exclusion restriction in this analysis, as highway access is related to transportation which is related to health services access.

| Descriptive statistics
The mean all-cause hospitalization rate across all counties in the

| Multivariate regression
Results of the multivariate linear regression are displayed in  Table 3). The association is similar but slightly greater in magnitude in mostly rural counties compared to mostly urban counties; a one point higher food swamp score is associated with However, no significant association was found between food swamp scores and ACSC hospitalization rates in completely rural or mostly rural counties.

| Sensitivity analyses
Our falsification test indicated that food swamp severity is not significantly related to the rate of mood disorder hospitalizations in the full model or in stratified models.
The multivariate model with population-adjusted RFEI as the predictor and additionally controlling for median income found county RFEI to be significantly associated with all-cause hospitalization rates and ACSC hospitalization rates.
Food swamp scores remain significantly associated with both allcause and ACSC hospitalization rates when additionally controlling for the percentage of homes that are vacant; in fact, the magnitude of the association is stronger in both models. These results are presented in Table 4. The pattern seen in the stratified analysis for allcause hospitalizations also remains.
The models utilizing the number of hospitalizations conditional on the number of adults with diabetes rather than the hospitalization rate indicate that higher food swamp scores remain strongly significantly associated with greater all-cause hospitalizations and ACSC hospitalization rates.  Notes: Presented are coefficient estimates from a state-level fixedeffects regression weighted according to county population. Robust standard errors are in parentheses. ***P < 0.001; **P < 0.01; *P < 0.05. residents of food swamps being at a notable disadvantage with regard to complications and hospitalizations in comparison with their counterparts in better food environments. As such, they add support to population-wide policies that seek to regulate the food environment, such as zoning restrictions for fast food outlets, as they may help reduce this disparity. The findings also suggest that the food environment could perhaps stymy some health system attempts to dimensions. This analysis focuses on availability, which is only one of these dimensions. The others include accessibility, accommodation, affordability, and acceptability, and it is possible that differences in these food environment dimensions between rural and urban contexts, which have been well noted, may moderate the observed association. For instance, rural areas are primarily served by small grocery stores; any supermarkets that do exist are concentrated in regional hubs. 36,37 These smaller stores often have a limited, less appealing, and more expensive selection of healthy foods than do the supermarkets found in urban areas. [12][13][14] Thus, although the ratio of fast food outlets to grocers may be similar in urban and rural areas, this lack of acceptable and affordable healthy options may be spurring rural residents, including those with diabetes, to purchase more unhealthy items. In fact, several studies, including some longitudinal studies, have found a positive association between small grocery stores and increased BMI in contrast to either a negative association or no association between supermarket availability and BMI. 6,38 In addition, the concentration of supermarkets in regional hubs means that rural residents must travel farther distances than their urban counterparts to access healthy foods, which is also made more difficult by the fact that rural areas often lack any public transportation. 13,36 This may further entice them to purchase more easily accessible unhealthy foods.

| D ISCUSS I ON
The observed stronger association between food swamps and hospitalization rates in rural counties may simply be the result of rural areas' poorer access to primary care, which is likely to be cor-

| Limitations
The study findings should be considered in light of some limitations.
First, the analyses are cross-sectional, which inhibits any claims about the causality or temporal ordering of the food swamp-hospitalization rate relationship. The results from propensity score-weighted regression analyses suggest that food swamps may exacerbate diabetes and result in higher hospitalization rates, but these conclusions would be While a more granular geographic unit of analysis may have yielded more precise results, there are also benefits to using the county as the unit of analysis. Zip codes and block groups may capture the food environment near the home, but most residents travel outside of their immediate areas for work and other daily activities and are therefore likely to be impacted by the food environment of neighboring areas.
Counties, however, are more likely to capture the majority of their daily routes and, thus, their exposure. 38 Additionally, land use and zoning regulations often occur at the county level, so estimates for smaller units may be harder to apply to any potential policy change discussions. 11 Nonetheless, zip code-level analyses should be pursued when data are available.

| CON CLUS ION
Food swamps appear to be linked to a disparity in hospitalization rates among adults with diabetes. In this analysis, we find that the degree to which unhealthy food outlets outnumber healthier options is associated with all-cause and ambulatory care-sensitive hospitalizations among diabetic adults, controlling for relevant health system-related and sociodemographic covariates. Further, the association of food swamp severity with all-cause hospitalizations is stronger in completely rural counties. Policy makers and health systems may want to consider the nature of the food environment in future efforts to address disparities in diabetes management, reduce preventable hospital utilization among adults with diabetes, and improve population health. Hector P. Rodriguez https://orcid.org/0000-0002-6564-2229