Identifying areas of high diabetes prevalence can have an impact on public health prevention and intervention programs. First, an elevated risk of diabetes incidence is associated with low socioeconomic status (SES) (1). Second, studies based on Centers for Disease Control and Prevention (CDC) estimates of diagnosed diabetes prevalence at the county level link high diabetes risk with obesity and sedentary lifestyles (2). Those county-level studies helped to identify a geographic region of the United States, mostly in southern states, as the diabetes belt (3).

Geographic disparities in diabetes prevalence at both the state and county levels have been observed in national health surveys, such as the Behavioral Risk Factor Surveillance System (BRFSS) (3–6). Geographic variations in diabetes prevalence were also found in metropolitan statistical areas (7), ZIP code areas (8), and census tracts in Canada (9). However, even more detailed data are needed to document disease burden and its associated health care costs at the local level. Such estimates are beyond the scope of CDC county-level estimates or national BRFSS surveys that are typically coded at the county and ZIP code levels only.

Small-area estimates, often based on modeling, are useful to local health agencies for targeting scarce resources toward areas with greatest need. The prevalence of heart disease in ZIP Code Tabulation Areas (ZCTAs) were predicted by using multilevel small-area estimate modeling that linked BRFSS data with county-level covariates (10). More recently, childhood obesity prevalence at the census block–group level was predicted by using geocoded data from a national survey and multilevel modeling (11). Our study used clinical data to examine the geographic distribution of diabetes at the census-tract level in King County, Washington. Census tracts are small areas that have an average population of 4,000, are smaller than ZCTAs, and are larger than census block groups.

Using geo-localized health and anthropometric data from insurance companies for small-area studies is a new direction in diabetes and obesity research. The objective of our study was to explore spatial relationships between area-based measures of SES, obesity, and diabetes for subcounty areas to arrive at a better understanding of the impact of local environmental factors on disease rates.

The final analysis was based on 371 census tracts and 59,767 insured adults. The number of participants per census tract ranged from 22 to 413; the average was 161. Of the sample, 22% were aged 55 to 64, and 30% were aged 65 or older. By comparison, 13% of the King County population were aged 55 to 64, and 13% were aged 65 or older. The crude prevalence of diabetes was 12.7% (men, 15.4%; women, 10.9%). The unsmoothed prevalence of diabetes ranged from 4.0% to 42.1%. EB-smoothed prevalence of diabetes ranged from 6.9% to 21.2%.

Census-tract–level prevalence of diabetes showed a nonrandom spatial pattern (Figure 1 and Figure 2); census tracts with a high prevalence of diabetes tended to cluster together (Moran’s I = 0.46, P < .001) (Table 1). High prevalence of diabetes was concentrated in south and southeast King County (Figure 2). We found significant spatial autocorrelations for the SES variables of interest. Census tracts with a high median household income, a high percentage of adults with a college degree, a high median home value, a high percentage of black residents, a high percentage of Hispanic residents, and high population density tended to cluster together. The extent of spatial clustering was particularly strong for the percentage of adults with a college degree (Moran’s I = 0.81) and the percentage of black residents (Moran’s I = 0.74) (Table 1).

In bivariate analyses, each area SES variable was significantly associated with an EB-smoothed prevalence of diabetes; we found the strongest negative association between the percentage of residents with a college degree and median home value (Table 1, Table 2, and Supplemental Figures 1–4). Percentages of non-Hispanic black and Hispanic residents were positively associated with the prevalence of diabetes, and high population density was negatively associated.

The spatial-error model showed that when holding other factors constant, for each 50% increase in median home value (eg, from $200,000 to $300,000), the prevalence of diabetes was 1.2 percentage points lower (95% CI, −1.6 percentage points to −0.9 percentage points) (Table 2). For each 10% increase in the percentage of adults with a college degree, the prevalence of diabetes was 0.9 percentage points lower (95% CI, −1.1 percentage points to −0.7 percentage points). For each $10,000 increase in median household income, the prevalence of diabetes was 0.3 percentage points lower (95% CI, −0.5 percentage points to −0.2 percentage points). The spatial dependence of diabetes data was fully accounted for by area-SES variables. We no longer found positive spatial autocorrelation after accounting for the covariates and the spatial dependence of the data (Moran’s I statistic between −0.05 and 0 for each model).

The association between area SES and diabetes was mediated to a large extent by the prevalence of obesity (Table 2). For the association between median home value and diabetes prevalence, 58% (95% CI, 41%–81%) was explained by differences in obesity prevalence after adjusting for age, race/ethnicity, and population density. For the association between diabetes prevalence and percentage of adults with a college degree, 47% (95% CI, 29%–67%) was explained by obesity. Similar effects were found in models that adjusted only for age.

In bivariate analyses, 24% of the variance in diabetes prevalence was explained by median home value. Adding population density, race/ethnicity, age, and spatial dependence to the model increased the variance explained to 48%. Similarly, 45% of the variance in diabetes prevalence was explained by college education; 52% was explained after adjusting for covariates. The 3 area-SES measures accounted for 39% of the variance in diabetes prevalence.

An examination of crude and adjusted linear regression models with standardized β-coefficients confirmed that college education had the strongest negative association with diabetes, followed by property value and median household income. By contrast, the principal metric of the food environment, the mRFEI, was not associated with diabetes prevalence either independently or after accounting for demographic covariates. In models adjusted for age, race/ethnicity, and population density, none of the SES measures were associated with the mRFEI. The mRFEI was not a mediator of the association between area-SES measures and diabetes prevalence.

An evaluation of the Moran’s I statistic for the residuals showed that much of the spatial dependence of diabetes was accounted for by area-SES variables (Table 3). We found strong spatial clustering in crude, age-adjusted, and age/race-adjusted models. Adding population density reduced the strength of clustering. Adding home value and college education variables further reduced clustering, although modest residual spatial dependence remained.

We used data from the Group Health Cooperative health care system to conduct spatial analyses of diabetes prevalence at the census-tract level in King County, Washington. Although the estimated prevalence of self-reported diabetes from the 2006–2010 BRFSS in King County was 6% (20), our estimates among insured adults ranged from 6.9% to 21.2%, depending on geographic location. The geographic distribution of diabetes was nonrandom, showing a high concentration in south and southeast King County. This “diabetes corridor” is a microscale version of the diabetes belt in the southeastern United States, consisting of 644 counties in 15 states (2).

Our ecological analyses thus support and extend previous research on the geographic distribution of adult diabetes and its SES correlates in the United States. In many such studies, in which data were aggregated to the county level, obesity and diabetes were regionally concentrated and associated with low levels of education and unemployment (21). However, county-level data are inadequate for health planning purposes; more granular data are required. Although King County is considered one of the healthiest counties in Washington State (22), we observed a smoothed diabetes prevalence of more than 15% in 78 of 371 (21%) of census tracts. By contrast, the county-level estimate of diabetes prevalence, according to BRFSS self-reports, is approximately 6% (20).

Studies have emphasized the contribution of local geographic factors to obesity and diabetes prevalence (23,24). Although state- and county-level maps of obesity and diabetes are best known (25), local health authorities have turned to smaller-area studies. Public Health–Seattle & King County analyzed BRFSS data by 26 health reporting areas (ie, aggregations of ZIP codes) to determine whether obesity and diabetes are concentrated in disadvantaged areas. Our census-tract–level analyses showed that diagnosed diabetes was strongly linked to area SES and in particular to the objective measure of residential property value. By contrast, the spatial clustering of diabetes was not related to the mRFEI, a widely available measure of the food environment.

In previous studies, we established an association between individual residential property values obtained directly from the county tax assessor and self-rated health: the higher the property value, the lower the likelihood of reporting fair or poor health (26) and the lower the prevalence of obesity among women (27). Analyses in adjacent Snohomish County showed a similar relationship between the prevalence of diabetes and home values, using data from BRFSS (28). For most Americans, home equity is among the most important components of accumulated wealth, suggesting the potential utility of this measure in future studies (29).

Census-tract–level maps of obesity and diabetes may provide unique insights into the social and environmental determinants of these conditions. The causal relationship between obesity and type 2 diabetes is thought to be mediated by weight-related changes in insulin resistance. Although a complex interaction among many genes and environmental factors may be involved, we found that the ecologic relationship between SES and adult diabetes was mediated in large part by obesity. These findings have policy implications for obesity-focused interventions to prevent diabetes, with efforts targeted toward neighborhoods and communities at risk.

Our study has numerous strengths. First, height and weight were measured during a clinic visit, not through self-report. Similarly, diabetes prevalence was estimated on the basis of medical records, rather than self-report. Third, access to clinical data from health insurers for a large and geographically diverse sample of nearly 60,000 adults allowed us to estimate the prevalence of diabetes at the census-tract level. To our knowledge, census-tract–level estimates of diabetes prevalence have not been published in the United States.

This study has several limitations. First, it was based on a sample of insured persons. Uninsured and Medicaid participants were excluded or underrepresented. The insured population, whether working or retired, may be in better health than the general population of King County. If so, the true disparities in diabetes prevalence by SES may be even more dramatic than those observed here. Second, the data on height and weight may have been subject to bias. Group Health members for whom data on height and weights were available tended to be older, had a higher RxRisk score (30), and were more likely to be women. Ecologic analyses, even at the census-tract level, do not allow for examination of cross-level effects. The paucity of individual-level measures of SES, race/ethnicity, or other factors was an additional limitation that precluded a multilevel analysis. Last, the mRFEI metric may not be the best summary measure of the availability of healthful or unhealthful food. In addition, the association between the food environment and disease in one place does not preclude the observations of different associations in other places.

Nonetheless, our data show that the prevalence of diabetes among insured adults can be reliably predicted by small-area SES, particularly education level and home value. Together, the area-based measures of median household income, median home value, and percentage of residents with a college education accounted for 39% of the variance. Further studies should evaluate these measures in predicting area-level diabetes; such evaluations could result in the creation of a powerful tool for policy making and assessment in public health. The use of health data from insurers to map the geographic distribution of diabetes could be a particularly valuable tool.