The final analysis was based on 74 ZIP code areas within King County. Forty-three ZIP code areas (n=93), were excluded from the analysis due to insufficiently small sample size within these ZIP code areas (n<10). A majority of the ZIP code areas excluded from the analysis had only one observation. A total of 8,803 individuals were included in the final analysis. The number of individuals surveyed in each ZIP code included in analysis ranged from 13 to 275. ZIP code areas used in the analysis had an average population size of 24,543 with a population range of 3,822–64,214. More than 95% of the King County adult population reside in the ZIP code areas included in the present study.

Binomial confidence intervals calculated by the score method were used to evaluate the statistical stability of the estimated crude obesity rates. Smoothed obesity rates were mapped using boundary files provided by the U.S. Census. ZIP code tabulation areas (ZCTA), a Census feature that approximates the geographic boundaries for ZIP code areas, was used for mapping and collecting contextual measures. The ZCTA boundary file was current for 2000.

Data from the 2000 U.S. Census included percent of adults living below the federal poverty line, median household income, median value of all owner-occupied housing units, percent African-American, and percent of Hispanic origin. The federal poverty line in the United States is based on household income and size. In 2000, the single-family household poverty threshold was $8,959. Contextual data were used to evaluate the bivariate association between area-based obesity prevalence and ZIP code area contextual characteristics. The smoothed obesity prevalence followed a normal distribution, justifying the use of linear regression models as opposed to Poisson models. The bivariate relationship between each predictor and smoothed obesity prevalence was assessed by estimating linear regression models (using a two-tailed test, α = 0.05). Multi-collinearity between predictors of interest was a concern. The tolerance for each bivariate covariate combination was assessed in linear regression models (1-r²). In no instance did the tolerance indicate serious collinearity problems. The smallest tolerance between predictors was between percent of adults below poverty and median household income (r²=0.70; tolerance=0.30), however percent of adults living below the federal poverty line was not a significant independent predictor of obesity prevalence, and was not included in the final models.

Significant parameters from bivariate analysis were included in an exploratory ordinary least squares multivariate linear regression model. Any analysis of spatial data is complicated by potential spatial autocorrelation or dependence, where contiguous or nearby areas have correlated response or predictor values. The Moran’s I statistic, a measure of spatial autocorrelation, was calculated to evaluate spatial autocorrelation for the smoothed obesity prevalence. A queen weights contiguity matrix was used for all spatial analyses. The queen weights define neighbors as ZIP code areas that share a boundary or a corner. One assumption of ordinary linear regression is independence of predictors, so the identification of significant spatial autocorrelation in our data necessitated the use of a spatial linear regression model, using a spatial error term, which corrects for spatial dependence (Anselin & Bera, 1998).

The observed disparities in obesity rates across King County ZIP code areas suggest a strong relationship between obesity rates and some area-based indices of SES. Those geographic disparities were much higher than those traditionally ascribed to race or ethnicity. Analyses of local 2000–2005 BRFSS data indicate that obesity rates among African-Americans in King County (26.5%) were 63% higher than among whites (16.3%) (Public Health - Seattle & King County, 2005). Obesity rates among persons with incomes <$15,000 were 20.0% as compared to 15.1% among persons with incomes >$50,000, a 32% difference (Public Health - Seattle & King County, 2005). At the national level, Hispanics are 1.2 times more likely to be obese than are whites (23.7% compared to 19.6%) (Mokdad et al., 2001). Nearly three-fold differences in obesity rates, based on geographic location, have not been reported in the literature.

Traditionally, race/ethnicity as well as education and incomes have been the focus of health disparities research (Isaacs & Schroeder, 2004). Given recent concerns that individual education and incomes may not adequately reflect social class (Marmot, 2000), there is a growing emphasis on the roles of occupation, social capital, and social context, all difficult parameters to capture in epidemiologic studies. Area-based measures of SES can provide additional information on poverty and wealth that is only rarely collected in United States health surveys (Krieger, Chen, Waterman, Rehkopf & Subramanian, 2003). In past studies, percent of residents living below the federal poverty level, based on census data, was identified as the best predictor of health outcomes (Krieger et al., 2003); whereas the present analyses point to median house values, a shorthand measure of wealth, as a strong predictor of obesity rates. The present findings are thus wholly consistent with past research on social determinants of health (Marmot, 2000).

The present study is subject to some serious limitations. First, heights and weights, used to calculate BMI values, were based on telephone self-report. Both men and women under-report weight, and men may over-report height in telephone surveys (Nawaz, Chan, Abdulrahman, Larson & Katz, 2001). However, the same BRFSS data are said to provide the best picture of the obesity epidemic in the U.S. and are the basis for national policy decisions (Mokdad et al., 2001; Mokdad et al., 2003). Second, our exclusive use of aggregate measures of SES and health outcomes does not allow us to generalize the effect of SES on obesity risk among individuals.

In addition, the standard CDC cut-points for obesity may not be appropriate for 18–19 year olds. We did not adjust BMIs for 18–19 year olds because approximately 1.5% of the sample was 18–19 year olds (n=140), and failure to adjust BMIs in the present study would not appreciably bias the results. While the current study evaluated the prevalence of obesity by area, an alternative approach would map mean BMI per area. Such an analysis would reduce potential bias due to misclassification of individuals as obese or not obese; however, data on mean BMI per ZIP code area was not available.

Perhaps more important is the issue that the ZIP code area is a problematic scale for spatial analysis. Because population counts per ZIP code area can vary widely, many ZIP code areas were too small to provide area-based prevalence estimates despite the use of an Empirical Bayes tool. Even for ZIP code areas with a large population, confidence intervals can be quite large as shown in Figure 1. We were unable to analyze the prevalence of obesity at the census tract level because respondents were only asked to report their ZIP code. Another challenge associated with the use of ZIP code areas is that they are designed to efficiently deliver mail, and the boundaries change subtly on a regular basis (Krieger, Waterman, Chen, Soobader, Subramanian & Carson, 2002).

Finally, obesity rates per ZIP code area were calculated with unweighted responses. Although the CDC provides weights for use in county- and state-level calculations, weights created for larger geographic areas are unlikely to adjust for non-response and differential probability of selection at the ZIP code level. Using unweighted data greatly simplified our analyses, since survey design effect for the BRFSS prevalence rates did not have to be taken into account when calculating the significance level of correlations and regression models. Potential sources of bias in state-specific BRFSS data include low response rates and non-response biases within certain demographic groups. Such biases may also occur at the neighborhood level, which might make the present population samples non representative. At state level, BRFSS demographics have been compared to other data sources to determine potential sources of response bias. Fewer external data sources are available in small area studies. This is an important caution, especially since the BRFSS design weights were not devised with small area studies in mind or used in the present analysis.

The present disparities at ZIP code area level stand in contrast to the well-known CDC maps, where the differences in obesity rates between the richer and the poorer states are only weakly apparent (Mokdad et al., 2001). Whereas the CDC maps have been used to support the argument that obesity rates in the U.S. are unrelated to social class, the present data show – to the contrary - that the obesity problem is concentrated in the most disadvantaged areas.