213 – Poverty Mapping with Complex Survey Data
Hierarchical Bayes Estimation of Poverty Rates
Sam Hawala
U.S. Census Bureau
Partha Lahiri
University of Maryland
In practice many applications of small area models use a `Normal-Normal-Linear' assumption, i.e., a normality assumption for the design-based survey estimates and for the area-level random effects and a linear regression function relating the true parameters to available covariates. We compare the performance of rate models by slightly changing the assumptions and using internal and external checks. when area sample sizes are in the hundreds, empirical analyses using a 'Normal-t-Linear' to protect against outliers, or a seemingly reasonable `Beta-logistic' assumption for rates, show no gain over the `Normal-Normal-Linear' type model. However, the same type of analyses show additional benefit from including historical data through a cross-sectional and time series model. We use Monte Carlo Markov Chain (MCMC) to implement the proposed models, posterior predictive checks, as well as external checks for model comparisons.
