Multilevel Regression and Poststratification (MRP) for Small Area Estimation with Geocoded FoodAPS Data

• Activity Number: 292 - Small Area Estimation with Small Samples
• Type: Topic Contributed
• Date/Time: Tuesday, July 31, 2018 : 8:30 AM to 10:20 AM
• Sponsor: Survey Research Methods Section
• Author(s): Xingyou Zhang* and Mark Denbaly and John Kirlin and Elina T. Page and Elizabeth Larimore and Shelly Ver Ploeg
• Companies: Economic Research Service, USDA and Economic Research Service, USDA and Economic Research Service, USDA and Economic Research Service, USDA and Economic Research Service, USDA and Economic Research Service, USDA
• Keywords: FoodAPS; Small Area Estimation; Multilevel Regression and Poststratification; Parametric Bootstrapping; External Validation; Geographic Linkage

Abstract

National social, economic, and health survey direct estimates are often not reliable for small geographic areas. This study explores a multilevel regression and poststratification (MRP) approach for small area estimation with the first National Household Food Acquisition and Purchase Survey (FoodAPS-1) that involves four major steps: 1) construct and fit multilevel models that link an outcome with both individual characteristics and area-level factors; 2) select a final multilevel model that could explain major geographic variations; 3) apply the final model to make a prediction with the census 2010 population; and 4) summarize small area estimates (SAEs) at census tract, county, state, and national levels. Preliminary results suggest that a multilevel logistic model that can explain more than 80% county-level variations in adult obesity status. Model-based estimates of obesity prevalence ranged from 14.1% to 54.7% with a median of 34.5% at county level and from 3.9% to 63.1% with a median of 31.3% at census-tract level. Additionally, FoodAPS-1 model-based estimates were correlated and consistent with BRFSS model-based estimates at both county and census-tract levels.