Authors:Xingyou Zhang (U. S. Census Bureau) (Speaker) Samuel Szelepka (U.S. Census Bureau)Blandine Bawawana (U.S. Census Bureau)Alfred O. Gottschalck (U.S. Census Bureau)
National demographic and health surveys have become routinely geocoded in federal statistical agencies in the United States. The geocoded surveys allow us to construct and fit appropriate unit-level multilevel models for small area estimation. We developed a unit-level multilevel model and poststratification (MRP) approach for small area estimation with geocoded ACS. The multilevel logistic model for health insurance coverage accounts for both individual characteristics of survey respondents (age, sex, race/ethnicity) and also their geographic contexts, specified by county-level and state-level random effects, to predict individual health insurance coverage status. The classic unit-level small area models are often criticized for ignoring survey design in model fitting. Thus, we compared small area estimates of health insurance coverage based on unit-level multilevel models with and without accounting for survey weights and design effects. Our preliminary results show that, as expected, unit-level survey-weighted models produced estimates more consistent with direct survey estimates than unweighted models; And unit-level models with sample sizes rescaled to reflect design effects generated estimates with larger mean squared errors (MSE) than models without a design effect adjustment. We further compared our MRP model-based health insurance estimates with those based on the current Small Area Health Insurance Estimates (SAHIE) area-level model. MRP provides a more flexible data linkage and modeling platform that incorporates survey design components and makes full use of geocoded ACS data and available geodemographic data to generate small area estimates of percentages of the population without health insurance coverage.