Small Area Estimation for High-Dimensional Multivariate Spatio-Temporal Count Data


Accounting for Dependence in Small Area Estimation

Scott Holan (Department of Statistics, University of Missouri) (Speaker)
Jonathan R. Bradley (Department of Statistics, Florida State University)
Christopher Wikle (Department of Statistics, University of Missouri)

Small area estimation of count data has become a research topic of widespread interest due to the ever-increasing need to produce more precise estimates for undersampled/unsampled geographies. This problem becomes more exacerbated when one acknowledges that many data sources also report related variables of interest that are referenced at different levels of spatial aggregation and by time. Moreover, the resulting multivariate spatio-temporal (areal) dataset can be extremely high-dimensional. Thus, to provide a coherent set of small area (domain) estimates in a computationally efficient manner, we propose a model that takes advantage of this highly complex dependence structure. Specifically, we propose a Poisson multivariate spatio-temporal mixed effects model that uses extremely effective dimension reduction to model high-dimensional multivariate spatio-temporal count data. We illustrate our method through simulation as well as an analysis of Quarterly Workforce Indicators (QWI) published by the US Census Bureau’s Longitudinal Employer-Household Dynamics (LEHD) program.