Urban food deserts

In urban areas, people often have unequal access to food retailers that carry healthy food because of their socioeconomic status and residential location. Impoverished neighborhoods, predominantly occupied by minority groups, were further away from the nearest supermarket than were wealthier neighborhoods in large metropolitan areas, such as Detroit [20], Los Angeles [19], and London, Ontario [25]. In addition, the food environment was generally less diverse and lower quality in poor and minority neighborhoods than in wealthier and predominantly white neighborhoods of urban counties in Maryland, North Carolina, and New York [22] and in New Haven, Connecticut [26]. In Tunisia, people who frequently shopped at supermarkets had a slightly improved diet quality, compared to those who shopped at other smaller retail outlets [27].

In contrast, fast-food restaurants and convenience stores were concentrated within walking distances from schools in deprived neighborhoods of Chicago [28] and East Los Angeles [29] and were close to residents with lower socioeconomic status [21, 30, 31] which suggests that children and low-income residents have greater access to poor quality food in urban areas. Energy-dense foods were not only less expensive than healthier foods, but also more resistant to inflation, resulting in the highest rates of obesity among residents of limited economic means [32]. Therefore, pricing interventions of taxes on energy-dense food, subsidies for the purchase of healthy food, and improved economic and physical access to supermarket may be essential for health promotion [33]. These studies suggest that people in disadvantaged urban communities are forced to travel longer distance to reach food retailers that carry affordable healthy foods, compared to people living in wealthier neighborhoods, where supermarkets are more abundant.

Rural food deserts

Rural food deserts, in contrast, have aspects that are fundamentally different from urban food deserts. Rural neighborhoods have fewer chain supermarkets than urban areas and have poor geographic access to supermarkets and healthy foods [31]. Rural areas offer fewer public transportation services than in urban areas, and this lack of transportation infrastructure is the most defining characteristics in rural communities with limited food access [6]. Large food retailers are often limited to larger towns, resulting in greater travel distances to reach services among isolated rural residents, such as in the Lower Mississippi Delta [34] as well as in rural Georgia [9]. In parts of rural Texas, neighborhoods were further away from supermarkets or full-line grocery stores and closer to convenience stores, although some of the most deprived neighborhoods had better access to the nontraditional food stores that sold a variety of fruit and vegetables [35, 36]. Yet, the types of food stores available and the range of healthy foods offered vary greatly across diverse low-income communities [37]. In addition, food consumption patterns were associated with poverty levels in rural communities, indicating that the diet quality was lower among adults with insufficient food supply [38]. This may be in part due to the fact that convenience stores and smaller grocery stores that are common in many rural communities generally have lower availability of healthy foods at higher costs than supermarkets [6, 39].

From an economic perspective, over the past 30 years the restructuring of food retail industries has occurred such that local grocery stores that once served small rural communities have been closed and replaced by or consolidated into a few regional and national chain grocers and supercenters [40]. As a result of this restructuring, remote rural residents must travel longer distances to reach a few large food retailers, which generally carry a wide variety of healthy food items at a relatively low cost. This uneven distribution of large food retailers creates problems of food deserts in rural communities. Rural residents with lower income, particularly younger age groups, are more likely to be food insecure than other rural populations [41]. With regard to geographic regions, the Great Plains are especially lacking in easy access to grocery stores [42].

To date, most studies have focused on the accessibility of large supermarkets that carry a large assortment of fresh fruit and vegetables and have been conducted in specific urban areas [15, 19, 20, 43, 44]. Other studies in urban environments have documented the association of obesity with neighborhood deprivation and access to fast food retailers [30, 45] or examined the spatial patterns of both supermarkets and fast food and/or convenience store locations in the context of nutrition and obesity [15, 21, 22, 29]. Studies of food environments in rural areas have become more common recently, and they have also focused on specific local communities and regions, [9, 24, 34–38, 41]. As a result, it is difficult to generalize their results beyond the local study areas.

There is some evidence that people with no supermarkets near their homes are less likely to have a healthy diet than those with the most stores [46]. However, there is a need for more information about how supermarket accessibility is associated with obesity and healthy food consumption nationwide, and how distance to supermarket is related to obesity prevalence and fruit and vegetable consumption in both urban and rural areas. This research examines geographic patterns of supermarket accessibility at the national level by mapping population-weighted mean distance to supermarket for every county in the conterminous United States. The associations of distance to supermarket with obesity and fruit and vegetable consumption are then investigated by linking the supermarket accessibility data with health surveillance data from the Behavioral Risk Factor Surveillance System (BRFSS). Although the term 'food desert' is multifaceted in its usage, this research focuses on distance as a physical constraint that may limit access to healthy food. Based on the existing literature, we test two hypotheses: (1) the population-weighted mean distances to supermarket are positively related to obesity prevalence and negatively related to fruit and vegetable consumption at the national level and (2) the associations of obesity and fruit and vegetable consumption with distance to supermarket vary between metropolitan and nonmetropolitan counties, reflecting the unique characteristics of urban versus rural food environments.

Data

Data on obesity and fruit and vegetable (F/V) consumption from 2000 to 2006 were obtained from the BRFSS. Established in 1984 by the Centers for Disease Control and Prevention (CDC), BRFSS is an annual cross-sectional telephone survey that provides state-specific data on health risk behaviors and other information related to chronic diseases. The BRFSS uses random digit dialing and collects data from non-institutionalized civilian adults aged over 18 years. We requested non-public use BRFSS data via a written request to the CDC, in which a county FIPS code was assigned to all individual records. Sample weights were provided to adjust for the differences in probability of sample selection and sample design weights were used for deriving representative population-based estimates. In performing statistical analyses, sample weights (_FINALWT), provided by the BRFSS, were used to calculate representative population-based estimates and to adjust for non-telephone coverage areas and non-response associated with telephone surveys.

We examined obesity and F/V consumption as response variables. For obesity, BRFSS respondents provided information about their height and weight from which their body mass index (BMI) was computed. BMI was calculated as weight in kilograms divided by height in meters squared (kg/m²), and respondents whose BMI was over 30 kg/m² were considered obese. For F/V consumption, the BRFSS variable _FRTINDX was used. The index was derived from the six questions of servings per day F/V consumption: (1) "How often do you drink fruit juice such as orange, grapefruit, or tomato?", (2) "Not counting juice, how often do you eat fruit?", (3) "How often do you eat green salad?", (4) "How often do you eat potatoes not including French fries, fried potatoes, or potato chips?", (5) "How often do you eat carrots?", and (6) "Not counting carrots, potatoes, or salad, how many servings of vegetables do you usually eat? (Example: A serving of vegetables at both lunch and dinner would be two servings.)" Respondents were classified into those who had five or more servings per day versus less than five. The F/V consumption variable was not uniformly available in all states for 2001, 2004, and 2006, thus this variable was analyzed using only the four available years (2000, 2002, 2003, and 2005).

The sample sizes for obesity and F/V consumption variables were 1,477,828 and 836,281, respectively, after excluding missing values and non-responses. Imputation to replace missing data was not considered because of the large amount of missing data on F/V in 2001, 2004, and 2006. About 4.5% of all respondents did not provide information on their height and weight and nearly 80% of these respondents were female in both metro and nonmetro counties. There was, however, no difference between height/weight respondents and non-respondents in F/V consumption in both metro and nonmetro counties, and other demographic characteristics were similar. The percent of respondents who did not provide information on their F/V consumption in the selected four years was 1.6% in metro and 1.0% in nonmetro counties. The difference between F/V respondents and non-respondents in obesity prevalence was small in both metro and nonmetro counties, and other demographic characteristics were similar.

Data on supermarkets were obtained from the 2006 Census Zip Code Business Pattern (ZBP) classified by the North American Industry Classification System (NAICS). Two NAICS codes (445110 and 452910) were combined to form a single supermarket and supercenter category (hereafter referred to as supermarkets) and georeferenced at the ZIP Code level. These two NAICS codes classified supermarkets and other grocery stores (except convenience stores), and warehouse clubs and supercenters, respectively. They were selected because both types of establishments primarily engaged in retailing a general line of food and groceries including fresh fruit and vegetables.

Population-weighted mean distance to supermarket

Because BRFSS data were not available at finer spatial resolutions than counties, supermarket distances were summarized at the county level. We computed the population-weighted mean distance to supermarket for each county. We obtained population-weighted ZIP Code Tabulation Area (ZCTA) centroids for the entire United States from the Missouri Census Data Center website at http://mcdc.missouri.edu/pub/data/georef/ and joined the NAICS ZIP code data at the national level with three supermarket size categories to the ZCTA centroids. Here, we demonstrate how population-weighted mean distance to supermarket is calculated using South Dakota as an example. The ZCTA centroids that contained different sizes of supermarkets in South Dakota are shown in Figure 1(a). Large cities, such as Sioux Falls and Rapid City, had all three sizes of supermarkets, while smaller cities often contained only small- to medium-sized supermarkets, particularly in sparsely populated areas of the western central South Dakota. Some ZCTA centroids in remote communities, however, contained no supermarkets of any size.

A 1,000 meter resolution raster dataset of Euclidian distance to the nearest supermarket was generated. We used Euclidian distance as a computationally efficient alternative to street network distance because of our focus at the national level. The distance surface to the nearest ZCTA centroid containing either large or medium supermarkets in South Dakota is shown in Figure 1(b). Eastern South Dakota and areas near large cities had shorter mean distance to large or medium supermarkets, compared to the western central part of the state. Zonal statistics were used to compute the mean distance to ZCTA centroids that contained supermarkets for each Census block group (CBG).

where D _j is the population-weighted mean distance to supermarket for county j, D _ij is the mean distance to supermarket in CBG i, P _ij is the population in CBG i. Summations were across all CBGs within county j. The population-weighted mean distances to large or medium supermarkets for all counties in South Dakota are shown in Figure 1(d). Counties in eastern South Dakota and counties with large cities had shorter distances to large or medium supermarkets, while counties in the west-central part of the state had longer distances.

Population-weighted means based on CBG population were used to take into account spatially heterogeneous populations within counties. For example, most of the population of Pennington County lives in Rapid City, whereas the eastern portion of the county has a much lower population density (Figure 1c). The population-weighted mean distance to supermarkets accounts for the fact that populations tend to be concentrated in cities and towns where supermarkets are also located. The population-weighted mean distance is thus typically shorter than the unweighted mean distance (5.6 km versus 6.8 km in the case of Pennington County.)

We repeated these steps to map the population-weighted mean distance to supermarket at the national level for each of three supermarket size categories: large supermarkets; large and medium supermarkets combined; and large, medium, and small supermarkets combined (Figure 2). Large supermarkets were the least numerous, thus, distance to large supermarkets was longer than the other categories, particularly in nonmetropolitan counties. Small and medium supermarkets were more numerous than large supermarkets. Thus, distances to supermarkets decreased when the medium and small store size categories were added in new locations where large supermarkets were absent (see example in Figure 1(a) for locations of supermarket by size).

Logistic regression models

We developed a series of logistic regression models for obesity and F/V consumption as response variables. We ran the models for metropolitan and nonmetropolitan areas of the United States separately. A metropolitan area is based on urbanized areas of 50,000 or more population within a county or a group of counties. Metropolitan areas consist of core urban area counties along with adjacent counties that have a high degree of social and economic integration measured by commuting ties [51]. In 2000, about 83% of Americans lived in metropolitan areas as defined by the Office of Management and Budget [52]. Nonmetropolitan areas included micropolitan counties with urban clusters of at least 10,000 but less than 50,000, and more sparsely populated noncore counties.

Population-weighted mean distance to supermarket was included as a county-level metric of variability in supermarket accessibility. Because of the strong positive skew of the supermarket distance variables (Figure 3), we transformed these distance variables using a common (base 10) logarithm. We also included geographic variables using the four Census regions (Northeast, Midwest, South, and West) to control for other unmeasured sources of regional variability unrelated to supermarket distance. The Northeast was selected as a reference group. Odds ratio (OR) estimates and 95% confidence intervals (CI) of the individual- and county-level variables were summarized for each model.

In performing all statistical analyses, sample weights (_FINALWT) provided by the BRFSS were used to obtain representative population-based estimates. The SURVEYMEANS and SURVEYFREQ procedures were used to compute descriptive summaries of obesity, F/V consumption, and population characteristics for metropolitan and nonmetropolitan areas [53]. Because independent variables were measured at two different levels (demographic characteristics at the individual level and supermarket distance and region at the county level) we used a generalized linear model for logistic regression, wherein counties were specified as clusters to account for correlations among individuals sampled within each county. Models were fitted using the GENMOD procedure, which used generalized estimating equations to account for the correlated data structure. All data processing and analyses were carried out using SAS version 9.2. All maps were created using ArcGIS Version 9.3.

Distance to supermarket

Metropolitan areas had overall shorter distances to supermarkets for all three store size categories (Figure 3), reflecting higher population densities in metropolitan areas. Nonmetropolitan areas, in contrast, had longer distances to supermarkets, but distances diminished in some areas as more stores of smaller sizes were included (Figure 2c). Distance to supermarket in suburban and exurban counties of the East Coast and Great Lakes regions was particularly sensitive to the addition of smaller supermarkets (Figure 2c compared to Figure 2a). Nonmetropolitan areas of the Great Plains and West, however, still had relatively long distances to supermarkets even if smaller supermarkets were included.

Obesity - metropolitan area model

Obesity - nonmetropolitan area model

The model for obesity in nonmetropolitan counties was similar to the metropolitan area model with higher ORs occurring among the middle-age cohorts (aged 45-54 and 55-64) compared to the reference group (Table 3). The ORs for sex, race/ethnicity, education, and income in the nonmetropolitan area model were generally similar to the metropolitan area model. Obesity risk in the Midwest and South in the nonmetropolitan area model was greater than in the Northeast. Obesity risk in the West, on the contrary, was less than in the Northeast. Unlike metropolitan areas, distance to supermarket was not associated with obesity prevalence for any of the supermarket size categories (Additional File 1. Appendix 1).

F/V consumption - metropolitan area model

F/V consumption - nonmetropolitan area model

The associations of individual-level covariates with F/V consumption in the nonmetropolitan area models were similar to those in the metropolitan area models (Table 4). In the nonmetropolitan area model, the middle-age cohort (aged 45-54) was not significantly different from the reference group. In addition, the odds of consuming F/V among non-Hispanic black and non-Hispanic Asian was not significantly different from that of non-Hispanic white. The higher education and income groups had greater odds of consuming F/V, compared to the reference group. The Midwest and South had lower F/V consumption, compared to the Northeast. Distance to supermarket was not associated with the prevalence of F/V consumption for any size categories in the nonmetropolitan area models (Additional File 2. Appendix 2).

Caveats and limitation

Analyses using data drawn from telephone surveys may be subject to bias because BMI and F/V consumption measures in the BRFSS are based on individuals' recall. Studies showed that population-level bias in self-reported height and weight was larger in telephone surveys, compared to in-person interviews that involved physical examinations. BRFSS underestimates the overall prevalence of overweight and obesity, and this bias could be attributed to such survey mode effects [67, 68]. Our study, however, focused on studying the correlates of spatial variability in obesity rather than precisely quantifying the prevalence of obesity from multiple databases, and these self-reporting biases should not influence our ability to detect these relationships. In addition we only examined one broad metric of healthy food consumption (the number of daily servings of fruit and vegetables). Different results could be obtained based on more detailed measurements on the quality and quantity of healthy food consumption [69].

The advantage of using the BRFSS data for this analysis was the availability of a representative sample covering virtually all counties within the conterminous United States. However, because counties are the finest geographic resolution at which the BRFSS data can be obtained, we had to summarize distance to supermarket as a population-weighted mean at the county level. In our analyses, this statistic was assumed to define the food environment for all individuals within the county. In reality, most counties encompass a mixture of neighborhoods with varying scales of supermarket accessibility, and this variability is not captured by our county-level indices. Different results could, therefore, be obtained by more localized analyses accounting for neighborhood and household effects.

Because we obtained supermarket data from the Zip Code Business Patterns dataset classified by the NAICS, we geocoded supermarkets to the population-weighted centroid of each ZCTA. Population-weighted centroids generally fall within the largest town or a segment of a city in the ZCTA, which is where supermarkets and other large retail stores are most likely to be located. Although we did not have the exact spatial locations of supermarkets, this lack of precision is probably not critical because distances were aggregated at the county level. More detailed, localized studies based on geocoded street addresses of individual residences and supermarkets in rural areas are needed in order to corroborate the results of this study, and provide additional insights into the proximal effects of supermarket accessibility in rural areas.

In addition, industry classification systems may be increasingly irrelevant as local economies become more diversified. Business establishments may involve sector changes, and newly emerging firms may not be recognized by existing classifications [70]. Despite this limitation, a study showed that business listings provided by commercial databases based on the NAICS and direct observations of business establishments were highly compatible at the neighborhood level [71]. Although rural areas may be associated with greater uncertainty for business existence and operations, the NAICS is a standardized business classification used by Federal statistical agencies for analyzing the U.S. economy.