In order to make the sample more representative of the sampling frame, we constructed post-stratification weights. The weights were calculated separately for Louisiana and Los Angeles and are based on the tract counts of people stratified by 1) gender, 2) age (<34, 35–44, 45–54, 55–65), 3) race/ethnicity (non-Hispanic white, non-Hispanic black, Hispanic, and other), and 4) median tract household income (<$27,000, $27–40,000, $40–55,000, >55,000). Because of the large variance in weights when we attempted to construct cross-classified weights, we opted to base the weights on the marginals (i.e. each dimension in isolation). In order to facilitate cross-site comparisons, the weights were then standardized to reflect the total number of people in the sample at each site.
In order to better understand within group regional variation in BMI and walking behavior as well as identify group-specific relationships between neighborhood characteristics and our outcomes, we stratified all of our analyses by race/ethnicity. First, using weighted t-test statistics, we tested whether BMI and frequency of walking behaviors varied by site for non-Hispanic whites and African Americans respectively. Then, we utilized weighted chi-square and t-test analyses to evaluate site differences for these two groups in terms of individual socio-demographic characteristics and neighborhood features.
We then modeled the frequency of recreational and utilitarian walking separately for non-Hispanic whites and African Americans. We used 2-level weighted hierarchical linear models with respondents clustered in census tracts to model the number of times a week participants reported engaging in walking for 1) utilitarian and 2) recreational purposes. Level-one (individual-level) predictors for these models included site, BMI, age, gender, household income, access to a car, the number of markets and parks within 1 mile of respondents’ residences, and respondents’ perception of neighborhood safety. Level-two (tract-level) predictors included neighborhood SES, the alpha index, median block length, and street density.
Finally, we modeled the BMI of non-Hispanic whites and African Americans. We ran models with all same level-one and level-two variables utilized in the earlier models. However, in the BMI models, we added in the two continuous variables for frequency of utilitarian and recreational walking.
In order to make the intercepts more meaningful, we centered respondent BMI, walking frequency, age, neighborhood SES, and all the street network measures around their grand means. Because the first-level residuals were not normally distributed, we also log-transformed the walking frequency and BMI variables making the coefficient estimates the percent difference in frequency for each unit change in the covariate.