Study Population and Data Sources
The Coronary Artery Risk Development in Young Adults (CARDIA) Study is a population-based prospective study of the determinants and evolution of cardiovascular risk factors among young adults. At baseline (1985–6), 5,115 eligible subjects, aged 18–30 years, were enrolled with balance according to race, gender, education (≤ and >high school) and age (18–24 and 25–30 years) from the populations of Birmingham, AL; Chicago, IL; Minneapolis, MN; and Oakland, CA. Specific recruitment procedures were described elsewhere.17
Follow-up examinations conducted in 1987–1988 (Year 2), 1990–1991 (Year 5), 1992–1993 (year 7), 1995–1996 (year 10), and 2000–2001 (year 15) had retention rates of 90%, 86%, 81%, 79%, and 74% of the surviving cohort, respectively.
Using a Geographic Information System (GIS), we linked time-varying, neighborhood-level food resource and U.S. Census data to CARDIA respondent residential locations in exam years 0, 7, 10, and 15 from geocoded home addresses. Among the 5115 participants at baseline, 48.2, 68.8, and 33.0% moved residential locations between years 0 and 7, 7 and 10, and 10 and 15, respectively.
Availability of neighborhood food resources
Fast-food chain restaurants, supermarkets (large grocery stores such as Kroger or Safeway), and smaller grocery stores were obtained from Dun and Bradstreet, a commercial dataset of U.S. business records. Food resources corresponding to each CARDIA exam period were extracted and classified according to 8-digit Standard Industrial Classification codes (Appendix Table e1
). Eight-digit codes were not available for 1985–86, so year 0 food stores were classified using 4-digit codes and textual queries designed for consistency with other exam years. Counts of each type of food resource were calculated within 1, 3, 5, and 8.05 kilometers (km) of each respondent’s residential location (Euclidean buffers), with the intent of capturing resources accessible by walking or by car. Specifically, 25% of all trips are less than 1.61 km (75% of these are by car), 62% of “social/recreational” trips are within 8.05 km,18
and 72% of walking trips are under 1 km19
(approximately a 15 minute walk). To test differences in how individual diet is related to food resources within varying distances, we examined food resources contained in concentric areas within 1 km, 1–2.9, 3–4.9, and 5–8.05 km of each respondent’s residence ().
Concentric areas in which food resource availability was measured.
Within each concentric area, we calculated fast food restaurant and grocery store counts per 10,000 population and, due to a smaller number of supermarkets, supermarket counts per 100,000 population. Population-scaled measures help to separate food resource availability from density of development, which is independently related to behavior20–22
and other neighborhood characteristics.23
Population within each area was derived from U.S. Census block-group population count, weighted according to the proportion of block-group area within each neighborhood buffer. While correlations of food resource availability among concentric areas were strong for food stores (up to 0.42 and 0.64 for supermarket and grocery store availability, respectively; ranged from −0.02 to 0.22 for fast food restaurants), examination of concentric areas allowed us to formally test associations across areas within the same model. Study conclusions were similar using 1, 3, 5, and 8 km buffers in separate models (reported in Appendix Tables e6–e8
Frequency of chain fast food consumption was ascertained at each exam year. Participants were asked “How many times in a week or month do you eat breakfast, lunch or dinner in a place such as McDonald’s, Burger King, Wendy’s, Arby’s, Pizza Hut, or Kentucky Fried Chicken?” Questions were open ended, but calculated to reflect a per-week consumption frequency.
Fruit and vegetable intake and overall diet quality was ascertained from an interviewer-administered, quantitative diet history of foods consumed over the past month and a questionnaire on usual dietary practices. Calculation of nutrient and energy intakes and validation of the CARDIA Diet History are described elsewhere.24–26
Diet quality was measured using the Diet Quality Index (DQI),27
which quantifies adherence to the 2005 Dietary Guidelines for Americans;28 Appendix Table e1
describes scoring criteria. Briefly, the DQI incorporates adherence to recommendations for nutrients, food groups, and broader health messages (diversity, moderation, and minimization of added sugars), each assigned scores ranging from 0 to 10 which were summed for a maximum score of 100. Higher values reflect healthier diets. Adherence to fruit and vegetable recommendations, a common marker of healthy dietary patterns,12
was derived from DQI components. This dichotomous measure also addressed highly skewed fruit and vegetable intakes and variation in recommended servings by sex and total energy intake.29–30
Individual-level baseline characteristics included age (grand mean centered), race (white, black), and study center. Education (≤high school, some college, college graduate) at Year 7, after most individuals attained their highest education level, was examined as a time-constant variable; Year 0 education was used if Year 7 education was missing. Time-varying individual-level characteristics included income (continuous), marital status (married, not married), and children or stepchildren ≤18 years living in the household (any, none). Income was not collected in year 0 or 2, so the closest measurement (year 5) was analyzed; each year was inflated to 2001 U.S. dollars using the Consumer Price Index. Missing income (n=897 observations; 5.6%) was imputed based on individual-level age, race, sex, education, and study center; and residence within or outside of an urbanized area, census tract-level median household income, and county-level cost of living index.
Because neighborhood socioeconomic status correlated with food resource availability in prior research4
and is independently related to diet,31
we controlled for percent of persons <150% of federal poverty level (1.5*federal poverty level32
) within the respondent’s census tract of residence at the time of examination. Spearman correlations with neighborhood poverty were 0.40 for grocery stores within <1km and 1–2.3km, but otherwise smaller than ±0.15; associations adjusted and unadjusted for neighborhood poverty were similar.
Effects of food resource availability on corresponding diet measures throughout young to middle adulthood were estimated in a series of sex-stratified longitudinal models. We focused on the most theoretically direct relationships: fast food consumption in relation to fast food availability, and diet quality and fruit and vegetable consumption in relation to supermarket and grocery store availability. Most interactions between sex and each independent variable were significant (Wald p<0.10) so models were sex-stratified.
We used fixed effect longitudinal models, which exploit the repeated measures of environment and diet in the CARDIA study by conditioning on each individual, thereby analyzing variation observed within person, over time. In this way, fixed effect models control for time-constant unmeasured variables (e.g., diet preferences that remain constant over time).34–36
In essence, each individual serves as his/her own control in fixed effect models. In contrast, random effects models (random person-level intercept) analyze variation both within and between individuals; they do not control for possible correlation between observed and unmeasured characteristics and are therefore more comparable to cross-sectional associations reported in prior research. The Hausman specification test indicated systematic bias with respect to the independent variables (p<0.001), so we report the more robust fixed effects estimates; corresponding random effects estimates are reported in Appendix Tables e6–e8
Models were fit using Stata 10.1 xt longitudinal functions (xtpoisson for fast food frequency, xtreg for diet quality, xtlogit for meeting fruit and vegetable recommendations), using the “fe” option.37
As described elsewhere,38
we treat neighborhood poverty as an individual-level exposure.
Natural-log transformation of food resource variables linearized relationships. All models controlled for time-varying age, income, marital status, children, and neighborhood poverty; because fixed effects models rely on within-person variation, coefficients for time-constant variables (study center, education, race, sex) are not estimated. To test the hypothesis that food resources within a shorter distance from home influence diet in low-income groups, we tested interactions by individual-level income (3 categories with adequate counts in whites and blacks: low, <$20,000; medium, $20,000–89,900; and high, ≥$90,000); income-specific associations calculated from estimated main effect and income interaction coefficients are presented for models containing significant (Wald p<0.10) income interactions. Due to unstable estimates, income interactions are not reported for fruit and vegetable recommendations. P-values for income interactions and Bonferroni-corrected comparison of estimates for food resources within different concentric areas are reported in Appendix Tables e4 and e5
Due to differences in diet measures collected across CARDIA exam year, fast food availability in relation to fast food consumption (fast food model) was examined using exam years 7, 10, and 15; and supermarkets and grocery stores in relation to diet quality and fruit and vegetable intake (food store models) were examined using exam years 0 and 7. Study retention and exclusions are presented in ; analytical samples included 10,975 (fast food model) and 8,652 (food stores models) person-exam observations. Food resource data were complete for all observations, so exclusion was unrelated to the study exposures. Additionally, our fixed effects models may mitigate selection bias (attrition and missing data) related to unobserved fixed individual-level characteristics.
Summary of study retention and exclusions