Neighborhood Food Environment and Walkability Predict Obesity in New York City

Background Differences in the neighborhood food environment may contribute to disparities in obesity. Objectives The purpose of this study was to examine the association of neighborhood food environments with body mass index (BMI) and obesity after control for neighborhood walkability. Methods This study employed a cross-sectional, multilevel analysis of BMI and obesity among 13,102 adult residents of New York City. We constructed measures of the food environment and walkability for the neighborhood, defined as a half-mile buffer around the study subject’s home address. Results Density of BMI-healthy food outlets (supermarkets, fruit and vegetable markets, and natural food stores) was inversely associated with BMI. Mean adjusted BMI was similar in the first two quintiles of healthy food density (0 and 1.13 stores/km2, respectively), but declined across the three higher quintiles and was 0.80 units lower [95% confidence interval (CI), 0.27–1.32] in the fifth quintile (10.98 stores/km2) than in the first. The prevalence ratio for obesity comparing the fifth quintile of healthy food density with the lowest two quintiles combined was 0.87 (95% CI, 0.78–0.97). These associations remained after control for two neighborhood walkability measures, population density and land-use mix. The prevalence ratio for obesity for the fourth versus first quartile of population density was 0.84 (95% CI, 0.73–0.96) and for land-use mix was 0.91 (95% CI, 0.86–0.97). Increasing density of food outlets categorized as BMI-unhealthy was not significantly associated with BMI or obesity. Conclusions Access to BMI-healthy food stores is associated with lower BMI and lower prevalence of obesity.


Research
The United States faces an epidemic of over weight and obesity (Ogden et al. 2006). Analyses of National Health and Nutrition Examination Survey data for [1999][2000][2001][2002][2003][2004] show that 32% of Americans > 20 years of age are obese (Ogden et al. 2006). New York City Department of Health and Mental Hygiene statistics show that New York City, our study site, likewise faces high rates of overweight/ obesity (Roberts et al. 2005). There is a grow ing understanding that the availability of resi dential neighborhood resources that support physical activity and healthy food choices may influence obesity rates (Larkin 2003;Rao et al. 2007).
Previous studies linking food environ ment measures with dietary intake or obesity have found mixed results. Proximity to super markets has been positively associated with consumption of a healthy diet (Laraia et al. 2004;Morland et al. 2002;Zenk et al. 2005) and negatively associated with overweight or obesity (Morland et al. 2006). Individuals with access to lowerpriced fruits and veg etables have a lower body mass index (BMI) (Sturm and Datar 2005), whereas those living near convenience stores have higher rates of overweight and obesity (Morland et al. 2006). To date, however, there is no evidence that proximity to fastfood restaurants influences diet or obesity risk (Burdette and Whitaker 2004;Jeffery et al. 2006;Liu et al. 2007). Most analyses relating the density of other types of restaurants or grocery stores to BMI or obesity risk found no significant associ ation (Jeffery et al. 2006;Liu et al. 2007;Powell et al. 2007;Sturm and Datar 2005).
Two concerns can be raised about this existing literature. First, most analyses do not control for built environment characteristics, such as landuse mix and population density, associated with pedestrian travel and lower BMI, which also tend to covary with density of retail food outlets (Rundle et al. 2007). Second, with a few exceptions (e.g., Morland et al. 2006), most studies relating the food environment to diet or body weight focus on just a few types of food outlets rather than considering the food environment as a whole. Because density measures for different types of food outlets are likely to be correlated with each other and with commercial space avail ability in general, their individual associations with BMI may be difficult to disentangle.
In this study we related the food environ ment to BMI and obesity in New York City. Addressing the concerns noted above, the analysis controls for neighborhood and built environment features already shown to influ ence BMI and includes measures of all res taurants, grocery stores, and specialty food vendors in the city (Rundle et al. 2007). To address problems of multicollinearity raised by simultaneous inclusion of a large num ber of food outlet measures, we constructed density measures for three food environment categories: BMIhealthy food outlets such as supermarkets and fruit and vegetable mar kets, BMIunhealthy food outlets such as fast food restaurants and convenience stores, and a BMIintermediate category.

Materials and Methods
The analyses presented here employed data collected during the baseline enrollment of subjects for the New York Cancer Project, a study of residents of New York City and the surrounding suburbs that has been described extensively elsewhere (Mitchell et al. 2004;Rundle et al. 2007). Of the total sample, 14,147 individuals had geocoded addresses falling within New York City boundaries, and 13,102 had a BMI < 70 and complete data for objectively measured height and weight and questionnaire measures of age, race and ethnicity, sex, income, and educa tional attainment. Table 1 shows descriptive statistics for individual characteristics. The demographic profile and spatial distribution of the sample are similar to those derived from the 2000 U.S. Census and from the 2002 New York City Community Health Survey (Rundle et al. 2007). Analyses of BMI, indi vidual demographic variables, and appended neighborhood characteristics were approved by the Columbia University Medical Center Institutional Review Board.
Neighborhood measures. We defined a study subject's neighborhood as a half mile (805 m) "network buffer" around his or her residential address, comprising loca tions reachable within a halfmile walk along the street network. Most urban planners assume that a halfmile is a walkable distance (Agrawal et al. 2008;Calthorpe 1993;Cervero 2006). We constructed sociodemographic and built environment measures, including food environment variables, for each indi vidual's neighborhood. To control for the effects of neighborhood composition on BMI, our models adjusted for the proportion of residents below the federal poverty line, pro portion black, and proportion Hispanic using data from the 2000 U.S. Census summary file 3 (U.S. Census Bureau 2000).
We assessed the possible confounding effects of the following measures of neighbor hood walkability: population density, density of bus and subway stops, percentage of com muters using public transit, landuse mix, and proportion of land zoned to permit commercial development (Rundle et al. 2007). We calcu lated population density, expressed as persons per square kilometer of land area, and the per centage of commuters using public transit from 2000 U.S. Census data (U.S. Census Bureau 2000). We based the numbers of bus and sub way stops per square kilometer on data from the Department of City Planning (DCP). We con structed the proportion of the buffer zoned to permit commercial development and a meas ure of residential/commercial landuse mix using the Primary Land Use Tax Lot Output data, a parcellevel data set also available from DCP. Landuse mix is an index of the extent to which a neighborhood supports both commercial and residential lands uses, with the index tending toward 1 as the mix of residential and commer cial floor area approaches a 1:1 ratio.
Food environment measures. We derived food environment measures from 2001 data purchased from Dun & Bradstreet (D&B; unpublished data). The data include busi ness name, geocoded location, and detailed Standard Industrial Classification (SIC) industry codes (http://www.osha.gov/pls/imis/ sic_manual.html) for food establishments. A priori, we grouped food outlets into catego ries hypothesized to provide BMIhealthy or BMIunhealthy food, with one intermediate category for food outlets whose classification was uncertain. We classified supermarkets and fruit and vegetable markets as BMIhealthy based on evidence associating proximity to supermarkets with better dietary patterns and lower BMI (Laraia et al. 2004;Morland et al. 2002Morland et al. , 2006Zenk et al. 2005), lower fruit and vegetable prices with slower growth in BMI (Powell et al. 2007;Sturm and Datar 2005), and daily vegetable consump tion with lower rates of obesity (LahtiKoski et al. 2002). Although supermarkets sell a range of food, including both healthy and unhealthy options, we consider them healthy food outlets because they offer local residents the opportunity to purchase healthy food.
No evidence is available linking natural food stores to diet or BMI, but food products typi cally available at natural food stores tend to be healthier; thus, we also categorized natural food stores as BMIhealthy food outlets.
The category of BMIunhealthy food out lets included fastfood restaurants, a choice based on extensive evidence linking fastfood consumption with high energy intake, fat intake, BMI, and weight gain (Befort et al. 2006;Bowman and Vinyard 2004;Duerksen et al. 2007;Duffey et al. 2007;French et al. 2001;Jeffery et al. 2006;Jeffery and French 1998;Thompson et al. 2004). The BMI unhealthy food index also included conve nience stores (Morland et al. 2006) and meat markets (Gillis and BarOr 2003;LahtiKoski et al. 2002). We classified pizzerias, bakeries, and candy and nut stores as BMIunhealthy based on the energy density of the types of foods sold there. Because "bodegas" or very small grocery stores tend to sell energydense foods and few fruits and vegetables, they were classed as BMIunhealthy (Kaufman and Karpati 2007).
The BMIintermediate category comprised food outlets for which evidence was insuf ficient for placement in the other two cate gories. This category included nonfastfood restaurants-that is, restaurants excluding fast food and pizzerias. Although eating food pre pared away from home has sometimes been associated with poor diet and higher weight (Gillis and BarOr 2003;Guthrie et al. 2002;Yao et al. 2003), research on consumption of food from nonfastfood restaurants has found no effect on weight or weight gain (Duffey et al. 2007;Jeffery et al. 2006;Thompson et al. 2004), and one study found higher vege table consumption among adolescents who ate more frequently at nonfastfood restau rants (Befort et al. 2006). The intermediate category also includes mediumsized grocery stores and specialty stores, as well as fish mar kets. Although some evidence associates fish intake with weight loss (Thorsdottir et al. 2007), fish markets in New York City often sell fried fish and fried seafood for immediate consumption; thus, the implication of this food outlet type for weight is unclear.
We identified most food outlet types by SIC code number alone: fruit and vegetable markets (#5431), natural or health food stores (#549901), fish markets (#542101), specialty food stores (#5451 and #5499, excluding #549901), convenience stores (#541102), bakeries (#5461), candy and nut stores (#5441), and meat markets (#542102). We distinguished three categories of grocery stores, excluding convenience stores. We identified "supermarkets" as grocery stores (#5411) with at least $2 million in annual sales or, for estab lishments with missing data on annual sales, at least 18 employees. (Among establishments with annual sales data, 18 employees was the threshold at which at least half had annual sales of ≥ $2 million.) "Mediumsized gro cery stores" were nonsupermarket groceries with at least five employees. "Bodegas" were grocery stores with fewer than five employees. We identified nationalchain fastfood restau rants through text searches in the D&B "com pany name" and "tradestyle" fields for names appearing in Technomic Inc.'s list of the top 100 limitedservice chain brands (Technomic Inc. 2006). We identified as local fast food those restaurants that were not already identi fied as a nationalchain fastfood restaurant and that had an SIC code indicating fast food (#58120300, #58120307, or #58120308), as well as the restaurants with names match ing those on this list of local fastfood restau rants. We identified nonfastfood restaurants with "pizza" or "pizzeria" in their name, or with SIC codes of #58120600, #58120601, or #58120602, as pizzerias. We categorized all other establishments with an SIC code of 5812 as nonfastfood restaurants.
The density per square kilometer of estab lishments falling within each of these three categories was calculated for each subject's unique network buffer. Subjects were then categorized into increasing quintiles for each of the three food outlet categories.
Statistical analysis. We calculated adjusted mean BMI for each quintile of retail density for the three food categories using cross sectional, multilevel modeling (Diez Roux 2000) with the Proc Mixed procedure (Singer 1998) in SAS (version 9; SAS Institute Inc., Cary, NC). Because each of the neighborhoodlevel among those living in proximity to each other, to exist across a geographic scale larger than the halfmile buffers. To account for this, we esti mated our multilevel models with community district as a level 2 clustering factor. New York City's 59 community districts correspond to named areas such as the Upper West Side and Chinatown. Although we measured no predic tive variables at level 2, the use of this nested data structure allowed for valid estimation of standard errors. We adjusted analyses for indi vidual and neighborhood sociodemographic characteristics and then for the five neighbor hood walkability measures. We evaluated the five walkability measures as possible confound ers individually and in combination. We mutu ally adjusted all analyses for quintiles of each of the three food categories. We calculated separate prevalence ratios for overweight and obesity compared with normal weight for increasing quintiles of retail food density categories using Poisson regression with robust variance estimates (Spiegelman and Hertzmark 2005). We used community district as a clustering variable to correct the standard errors for intercorrela tions among individuals across larger areas of the city and to generate robust SE estimates.

Results
The data set initially received from D&B included 32,949 retail food businesses for New York City. After correction of geocoded addresses and removal of duplicate records, businesses likely to be defunct, and records likely to represent back offices and corporate offices, the data set included 29,976 busi nesses, of which 29,858 fell within the bounds of study subjects' neighborhoods. Table 2 dis plays descriptive statistics for the BMIhealthy, BMIunhealthy, and BMIintermediate categories as well as for specific food outlet types. Density of intermediate and unhealthy food outlets was much higher than density of healthy food outlets. Almost all study sub jects lived within a halfmile of an unhealthy food outlet, with an average density of 31 such outlets per square kilometer. By contrast, only 82% lived within a halfmile of a healthy food outlet, with an average density of four outlets per square kilometer. Density measures for food outlet types were significantly correlated across neighborhoods, with correlation coef ficients ranging from 0.38 (convenience stores and supermarkets) to 0.85 (nonfastfood restaurants and pizza restaurants). Figure 1 maps the density of BMIhealthy food outlets, expressed in outlets per square kilometer, across the city. Outlet density was highest in highwalkability areas of the city, such as Manhattan, and lowest in lowwalk ability areas, such as Staten Island. Outlet den sity also varied by neighborhood income and race/ethnic composition, with higher densities in affluent and predominantly white neigh borhoods in the southern half of Manhattan and lower densities in the poor and predomi nantly black or Latino neighborhoods in the Multilevel analyses of the association between BMI and the food environment measures showed significant associations only with access to BMIhealthy food. We also assessed possible confounding effects of built environment variables. Population density, which has previously been inversely associ ated with BMI in analyses of the same data set, had an appreciable confounding effect, but further control for landuse mix, percent commercial area, and access to and neighbor hood use of public transit did not alter the results. Table 3 shows adjusted mean BMI for each quintile of the three food catego ries and the median density of food outlets for each category; Figure 2 displays the asso ciation between healthy food outlet density and BMI based on this analysis. The adjusted mean BMI in the fifth quintile of healthy food was 0.80 units [95% confidence interval (CI), 0.27-1.32, p < 0.01] lower than in the first quintile of healthy food. Population den sity and landuse mix remained significantly inversely associated with BMI after control ling for measures of the food environment. Increasing density of the BMIunhealthy and BMIintermediate food categories was not associated with BMI, and analyses of selected subcategories of BMIunhealthy food (fast food, pizzerias, and convenience stores) found no significant associations.
Because there was little difference in the adjusted mean BMI of individuals living in the first and second quintile of BMIhealthy food density, we collapsed these two catego ries into a single reference category to increase statistical power for analyses of the prevalence of overweight and obesity. The reference cat egory had a median density of 0.76 healthy food outlets per square kilometer. Table 4 shows the prevalence ratios for overweight and obesity by increasing density of healthy food outlets, increasing population density, and landuse mix. Controlling for popula tion density and landuse mix, the prevalence of overweight and obesity were both lower among individuals with the highest density of healthy food outlets. Controlling for other features of the built environment did not alter the prevalence ratio for healthy food density.
Our previous work showed that increas ing landuse mix and population density were inversely associated with BMI; this associa tion remained after control for the density of BMIhealthy, BMIunhealthy, and BMI intermediate food outlets (Rundle et al. 2007). The prevalence ratio for obesity comparing the fourth and first quartiles of landuse mix was 0.91 (95% CI, 0.86-0.97) and comparing the fourth and first quartiles of population density was 0.84 (95% CI, 0.73-0.96).

Discussion
The results presented here indicate that the food environment is significantly associated with body size net of individual and neigh borhood characteristics and neighborhood walkability features. A higher local density of BMIhealthy food outlets was associated with a lower mean BMI, a lower prevalence of overweight, and a lower prevalence of obesity. BMIunhealthy food stores and restaurants were far more abundant than healthy ones, but the density of these unhealthy food outlets was not significantly associated with BMI or with body size categories. Of studies relating the food environment to body size, this work is among the first to measure the food envi ronment comprehensively and to account for the effects of other built environment factors associated with obesity. The apparent effect of the food environment, while modest, is net of the significant associations between indices of neighborhood walkability and BMI. Our prior work showed that built environment features related to walkability were associ ated with approximately a 10% difference in the prevalence of obesity (Rundle et al. 2007). Even after control for measures of the food environment, the estimated effects of these built environment variables remained and were of a similar magnitude. Considered together, food environment and neighbor hood walkability may have a substantial effect on body size.
Although the observed associations between BMI and the density of BMI healthy food establishments were consistent with expectations, we had also hypothesized that increasing density of BMIunhealthy food options would be positively associated with BMI. Because the density of unhealthy food outlets is correlated with commercial activity in general as well as other features of the urban landscape that promote pedes trian activity, we expected that associations  between unhealthy food density and BMI had been masked in prior research and would be observed after control for such built environ ment features. Consistent with other studies in this area, however, we found no associa tion between density of unhealthy food and BMI or obesity. This lack of association may reflect the ubiquity of unhealthy food in an urban environment; as Table 2 shows, virtu ally all New York City neighborhoods pro vide many opportunities to eat poorly. In addition, unhealthy convenience foods may be consumed near the workplace or during travel about the city, making the density of unhealthy foods in the residential neighbor hood less relevant. Alternatively, the null find ings may reflect undercounting of unhealthy food outlets in the most disadvantaged urban neighborhoods. As the case of New York City shows, the penetration of nationalchain fast food is low in some of the poorest neighbor hoods; this niche in the food environment is filled by inexpensive ethnic restaurants sell ing highcalorie takeout food (Graham et al. 2006). Better measures of the food environ ment may show an association of unhealthy food outlets with body size. One limitation of this study and, indeed, of most studies on this topic is that our data are observational and crosssectional. Observed associations may be attributable to selfselec tion of individuals into neighborhoods that support their preferred lifestyle; for instance, individuals who prefer to consume healthy foods may move to neighborhoods with more healthy food outlets. Conversely, retailers selling healthy foods may choose to locate in neighborhoods where they believe the popula tion will be most receptive to their products. In addition, questions might be raised about two potential sources of error in the food environment measures. The first is incom plete coverage of the D&B data. Because the D&B data are used primarily for marketing purposes, coverage may be less complete in areas less attractive to marketers, such as low income neighborhoods. For error in the D&B database to bias our results, it would have to be correlated with the spatial distribution of BMI. Our analyses control for neighborhood sociodemographic composition, which may be an important correlate of measurement error in the D&B data. Second, measurement error may be caused by misclassification of food outlets into the BMIhealthy, BMIunhealthy, and BMIintermediate categories. Some food outlets, such as fruit and vegetable markets, are internally relatively homogeneous, whereas others, such as grocery stores or nonfastfood restaurants, may have significant internal het erogeneity. Withincategory heterogeneity in food selection may bias food environment coefficients toward zero or create interactions between neighborhood composition and food environment characteristics. Although the analyses controlled for neighborhood socio demographic composition and for landuse mix and commercial space, variables that might be expected to influence the extent of withincategory measurement error, mea surement error remains a concern. A further limitation is the mismatch between the time period of the survey (2000)(2001)(2002) and the time period of food environment measures (2001), population census measures (2000), and landuse and zoning data (2003); because neighborhood demographic and built environ ment charac teristics typically change slowly, these discrepancies should not affect the results significantly. Limitations also include the lack of an audit to verify types of food sold in different types of stores.
A distinctive feature of this study is its use of broad categories to characterize the food environment based on the existing literature. Although this analytic strategy sacrifices the opportunity to identify associations between specific food outlet types and BMI, it has several advantages. First, although some in the public health and medical communities, as well as the popular media, have focused on the contribution of fast food to the obe sity epidemic, other types of food outlets also sell highenergydensity food; comprehensive measures of the food environment provide a more accurate account of the food choices available to urban residents (Stender et al. 2007;Wallis 2004). Second, density measures for the 14 individual food outlet categories are significantly correlated; reflecting this multi collinearity, models including all 14 meas ures are quite unstable. Third, reducing the number of food outlet measures made it less likely that one would be significant simply by chance. Specific choices about how to group food outlet types can certainly be debated and can be tested in replication.
The research reported here adds to our knowledge about the relationship between the food environment and obesity with evi dence that access to BMIhealthy food out lets such as supermarkets, fruit and vegetable markets, and natural food stores is inversely associated with obesity. This protective asso ciation is net of urban design features that promote pedestrian activity and lower BMI, as well as the density of other types of food outlets. Although not identifying a specific culprit within the retail food environment for the obesity epidemic, these analyses indicate that retail outlets providing opportunities for healthier food purchases are associated with lower BMI. If the results of our observational research are confirmed by future studies that permit causal inference, this evidence would suggest that increasing access to healthy food outlets is likely to do more to address the obe sity epidemic than limiting unhealthy food outlets. Given the recent proliferation of ini tiatives to promote access to supermarkets, farmers markets, and fruit and vegetable stands and to limit fastfood outlets (Abdollah 2007;Lee 2007;Marter 2007), study of the causal relationship between the food environment and diet or body size should be a priority for future research.