Dependent Latent Effects Modeling for Survey Estimation with Application to the Current Employment Statistics Surveys

Julie B. Gershunskaya and Terrance D. Savitsky

Abstract

The Current Employment Statistics Survey, administered by the U.S. Bureau of Labor Statistics, publishes total employment estimates for thousands of domains at detailed geographical and industrial levels. Some of these domains do not have adequate sample size for the direct probability sample-based estimates to be reliable. Small area estimation methods are used to integrate information from historical sources and correlated domains to improve estimation efficiency. In this paper, we explore alternatives to the Fay-Herriot two-stage hierarchical model that relax distributional and independence assumptions among random effects indexed by domain and month in order to more fully borrow strength to improve the efficiency of published employment estimates. We compare the performances of our alternative models on both synthetic data and in application to estimates from the Current Employment Statistics survey.