Use of Small-Area Estimates to Describe County-Level Geographic Variation in Prevalence of Extreme Obesity Among US Adults

Key Points Question What is the county-level prevalence of extreme obesity in the United States? Findings In this cross-sectional study of adults in the United States, county-level prevalence of self-reported extreme obesity ranged from 1.3% to 15.7%. Several prominent clusters of high prevalence were identified, including in the Mississippi Delta and the Southeast. Meaning Prevalence of extreme obesity appears to vary considerably by county; heterogeneity is obscured by available state-level prevalence estimates.

eMethods 1. Description of Geographic Aggregation Aggregation was conducted using the Geographic Aggregation Tool (GAT) R version 1.33.A BRFSS dataset with complete cases to be used in the model was created; respondents missing BMI, age, race/ethnicity, or county indicator were removed.Variable summaries by county (FIPS5) were run for total observations and ultimately saved as a shapefile, which was imported into the data aggregation tool.Several levels of aggregation were explored using the tool: total count = 5, 10, 25; allowing and restricting counties to merge across states.A determination was made by visual inspection of the new geographies to combine areas only when needed to ensure representation for each area.Because a small area estimation approach was used after this step to create prevalence estimates, it was not necessary to obtain geographical regions with enough observations to produce stable estimates.The final aggregation setting used required each area contained 5 observations.The determination to restrict counties from merging across states was made both because of the conceptualization of counties nested within states in the multilevel models and to facilitate the internal validation technique that compared state-level estimates.
In 2012, there were 3,109 counties within the 48 states in the contiguous United States and Washington, D.C.After county aggregation using the Geographical Aggregation Tool, there were 2,215 county or county-like areas.Of the 430 aggregations, 57.4% included only two counties and 90% of aggregations contained four or fewer counties.

Reference
Talbot TO, LaSelva GD.After weighting the predicted value by the actual frequency of each sex-age-race/ethnicity cell within each county for each of the aforementioned cells, all cells within a county were summed to produce the county-level prevalence estimates.The Moran's I is defined as: for observations at  = 1, … ,  counties,   and   represents the deviations from the country-wide mean, (  − ̅ ) and (  − ̅ ) respectively, and   is the weight assigned to all  counties, relative to their spatial adjacency to county , as detailed in eFigure2. 0 represents the aggregate of all spatial weights and is given by: Using notation similar to the global Moran's I, the local indicators can be depicted as follows: Such that   refers to the local Moran's coefficient at county  and deviations from the regional mean are summed over  neighboring values.We determined significance for counties within geographical clusters based on how far the local Moran's Index deviates from the Monte Carlo simulated distribution under the null hypothesis of complete spatial randomness -that is, significantly higher or lower prevalence of extreme obesity than would be expected, using a threshold of p<0.05.A weights matrix defines the spatial relationships such that closer areas receive a greater weight in calculations than those that are further away.In the queen matrix, all areas contiguous to the observed location are incorporated into the weighting scheme (as in a queen's movement in chess).In a rook matrix, regions are spatial neighbors if they share a side but not if they meet at only one vertex.For a k-nearest neighbors' matrix, centroid distances between polygons are calculated to determine the closest k neighbors for incorporation into the weighting scheme.A first-order queen contiguity weights matrix was chosen, though a sensitivity analysis was run utilizing both a rook and k-nearest neighbors (k=4) matrix.
contiguity, order of contiguity=1 Moran's I: 0.376, z-value: 29.504, p value: 0.001 Weight: Rook contiguity, order of contiguity=1 Moran's I: 0.381, z-value: 30.328, p value: 0.001 Weight: k-Nearest Neighbors, number of neighbors=4 Moran's I: 0.394, z-value: 25.915, p value: 0.001 Geographic Aggregation Tool.Albany, NY: New York State Health Department; 2015.https://www.albany.edu/faculty/ttalbot/GAT/GAT_vR13_guide.pdf.Description of the Multilevel Regression Model Final adjusted models were specified as follows: �  = 1� =  −1 ( 000 +  1   +  2   +  3   +  01   +  00 * Allowing  to refer to the individual,  to the county, and  to the state, the probability of extreme obesity is modeled by the inverse logit of intercept  000 , with  1 representing the regression coefficient for individual age group,  2 representing the coefficient for individual sex,  3 representing the coefficient for individual race/ethnicity, and  01 representing the regression coefficient for the county education quartile.The adjustment to the error term due to state as a random effect (state-dependent deviation) is referred to by  00 * and the adjustment due to county as a random effect by (county-dependent deviation)  0 * .A similar model was constructed modeling the probability of moderate obesity.indicates county and   is the predicted probability of extreme obesity for an individual in age , sex , race/ethnicity category  in county c and state .
Weighted Prevalence of Extreme and Moderate Obesity by Individual-Level Model Covariates Used for Estimating Predicted Risk, United States, BRFSS 2012* eTable.*Includes data from all 50 states and Washington, D.C.