Abstract:
|
Unit-level multilevel models can generate small area estimates at low geographic levels, such as census blocks. Because detailed census cross-tabulation population counts are available only by age, gender and race/ethnicity, multilevel models can only use these three variables in prediction. By using bootstrapping, we fitted multilevel logistic regression models with individual age, gender, race/ethnicity, and education for two outcomes -- current smoking and obesity -- from the 2011 Behavioral Risk Factor Surveillance System. We introduced a parametric bootstrapping method to assign education status for Census 2010 block-level population by age, gender, race/ethnicity in model prediction using 2007-2011 American Community Survey 5-year estimates. We compared county-level estimates with Missouri County-level Study direct estimates: the inclusion of individual education in model fitting and then prediction increased the correlation coefficients from 0.40 to 0.45 for current smoking and from 0.27 to 0.34 for obesity. Thus, multilevel small area estimation could include additional individual variables via bootstrapping.
|