We used data from the 2007 and 2009 waves of the California Health Interview Survey (CHIS), a random-digit–dial telephone survey of California’s noninstitutionalized population (13
). In 2007 and 2009, the CHIS included interviews of 98,662 adults aged 18 years or older. The sampling weights provided by CHIS account for unequal sampling probabilities and nonresponse to allow generalization to the study population. We limited our analysis to 97,678 adults aged 18 years or older, excluding 533 (0.5%) pregnant women and 451 (0.5%) respondents whose information was provided through a proxy interview.
Self-reported measures on dietary intake included the number of times per week the following were consumed during the month before the interview: fruits (excluding juices), vegetables (excluding fried potatoes), sugar-sweetened soft drinks (excluding diet soft drinks), and fried potatoes (including French fries, home fries, and hash browns). The measures also included the number of times fast food had been consumed in the week before the interview. The information was collected through questions in the following format: “During the past month [‘or in the past 7 days’ for fast-food consumption], how often did you eat [food item name]?” The question was often followed by clarification of the food items. The response was standardized to reflect mean per-week intake frequency.
Anthropometric measures included BMI, calculated as self-reported weight in kilograms divided by self-reported height in meters squared. We defined “overweight or obese” as a BMI of 25.0 or more and “obese” as a BMI of 30.0 or more, according to World Health Organization classifications for adults (14
Whereas most large survey studies use a predefined administrative unit such as census tract or zip code to define neighborhood, we defined neighborhood on the basis of geographic distance from a respondent’s residence, and we defined neighborhood food environment by counting the number of different types of food outlets within those distances. We drew circular buffers of varying radii (0.25, 0.5, 1.0, 1.5, and 3.0 miles) centered on each respondent’s residential address. Because 1 mile is often used as a threshold for walkable distance (6
), we considered distances of 0.25, 0.5, and 1.0 miles to be within walking distance. We measured Euclidean distance (straight line distance between 2 points) using ArcMap 9.1 (ESRI, Redlands, California).
We used food outlet data from the 2008 release of InfoUSA (15
). We overlaid food outlet locations on the buffers around respondents’ residences and counted the number of different types of food outlets in each buffer. We classified fast-food restaurants, full-service restaurants, convenience stores, small food stores, mid-size grocery stores, and large supermarkets by using the North American Industry Classification System (NAICS) (16
). NAICS does not have a code for fast-food restaurants; we identified 63 major fast-food franchises that have main menus that include items such as hot dogs, hamburgers, pizza, fried chicken, submarine sandwiches, or tacos by NAICS codes 72221105–6. Full-service restaurants were identified by NAICS codes 72211001–20; convenience stores, code 44512001; and small food stores (annual sales <$1 million), mid-size grocery stores (annual sales of $1–$5 million), and large supermarkets (annual sales >$5 million), codes 44511001–3, respectively.
To examine the association between neighborhood food environment and dietary intake, we performed negative binomial regression analysis. The dietary intake measures were the dependent variables, and the numbers of different types of food outlets in the buffers were the explanatory variables. Negative binomial regression is a generalization of the Poisson model in which the Poisson parameter has a random component (17
). We performed separate regressions for each dietary intake measure and buffer size, calculated average marginal effects (AMEs), which measure an estimated change in the outcome in the observed unit associated with 1 unit change in the regressor of interest, and applied the Bonferroni adjustment for multiple comparisons. We controlled for potentially confounding individual and neighborhood factors. Individual-level control variables included sex, age (in years and age squared), race/ethnicity (white, African American, Hispanic, Asian or Pacific Islander, Native American, other race/multirace), household size, annual household income (in natural logarithm), education (not a high school graduate, high school graduate, high school graduate but not college graduate, college graduate, and more than college degree), marital status (married, divorced/separated/widowed, single), parental status (has a child or has no child), physical activity (regular activity, some activity, sedentary), and survey year. Although we were interested in vehicle ownership, this information was not collected in CHIS 2009 and therefore not included in our analysis. Proxies for neighborhood-level control variables, which were obtained from 2000 Census data (18
), included population density, median household income, and proportion of non-Hispanic white residents of a respondent’s residential census tract. These factors did not match our definition of neighborhood for the explanatory variables, but they could be derived only from predefined administrative units.
To examine the association between the neighborhood food environment and BMI measures, we regressed BMI (through ordinary least squares [OLS]) and its dichotomous cutoffs of BMI of 25.0 or more and BMI of 30.0 or more (through logistic regression) on the same set of explanatory and control variables. Again, we performed separate regressions for each buffer size, calculated AMEs for logistic regression models, and applied the Bonferroni adjustment for multiple comparisons. We then compared the significant associations of food environments with dietary intakes and BMI measures to examine the hypothesis that the neighborhood food environment influences BMI through its influence on dietary intakes.
We performed several sensitivity analyses: stratified analysis by urbanicity (urban vs nonurban residents) and income level (federal poverty level [FPL] ≤130% vs FPL >130%) and analysis by density (instead of raw counts) of food outlets by census tract (the number of food outlets in census tract per 1,000 population and the number of food outlets in census tract per square mile). We defined “low income” as FPL of 130% or less. All analyses were performed in Stata 12.1 (StataCorp LP, College Station, Texas). We weighted the regression using sampling weights and estimated P values based on heteroscedasticity-robust standard errors obtained by using the Eicker–Huber–White sandwich estimator.