Small Area Estimation with Small Sample Size National Household Surveys: a Multilevel Regression and Poststratification Approach with FoodAPS

• Activity Number: 138 - Modeling Applications for Backcasting, Nowcasting and Forecasting
• Type: Contributed
• Date/Time: Monday, July 29, 2019 : 8:30 AM to 10:20 AM
• Sponsor: Survey Research Methods Section
• Author(s): Xingyou Zhang* and Alisha Coleman-Jensen and Shelly Ver Ploeg and Mark Denbaly
• Companies: Economic Research Service, USDA and Economic Research Service, USDA and Economic Research Service, USDA and Economic Research Service, USDA
• Keywords: Small Area Estimation; Multilevel Regression and Poststratification; FoodAPS; Food Security

Abstract

National surveys have been designed to produce reliable national-level social, economic and health indicators. However, national survey direct estimates often cannot meet the rapid growing data needs from state and local data users. Small area estimation techniques have been developed to address this critical data need. Current small area estimation methods are mainly based on very large sample size national surveys, such as ACS, CPS, BRFSS, and NHIS. Small sample size national surveys, such as NHANES and FoodAPS, have limited sampling geographic coverage: not all states are sampled, and less than 10% of counties are sampled. FoodAPS has sampled 4,826 households from only 27 states and 107 out of 3,143 US counties (about 3.4% counties sampled). We extended a Multilevel Regression and Poststratification (MRP) approach with FoodAPS to generate small area estimates at state, county and census tract levels. We illustrate this methodology using USDA’s household food security measure, a critical economic outcome collected by FoodAPS. Our final model explained 89% of county-level variation and 91.4% of state-level variation in the prevalence of food insecurity with FoodAPS data.