Online Program

Evaluation of Changes in Generalized Variance Function Estimators Following Revision of Publication Structure
*MoonJung Cho, U.S. Bureau of Labor Statistics 
John Eltinge, Bureau of Labor Statistics 
Julie Gershunskaya, U.S. Bureau of Labor Statistics 
Larry Huff, U.S. Bureau of Labor Statistics 
Lily Wang, University of Georgia 

Keywords: Degrees of freedom, Design-based inference, Generalized least squares, Model-based inference, Superpopulation model, Variance estimator stability

In work with data from establishment surveys, analysts have interest in estimation and inference for one or more population totals, means or ratios within relatively small subpopulations. For example, subpopulations can be defined by the intersection of industry and geographic area. Due to relatively high levels of sampling variability in standard direct design-based variance estimators at these finer levels of aggregation, analysts may need to consider use of generalized variance function (GVF) models. Development and evaluation of these models requires substantial effort, but it is important to consider periodic re-fitting of these models with more recent data. In particular, refitting may be important if there are substantial changes in (1) the inferential needs of the primary data users; (2) features of the population; (3) point estimation methods; and (4) the type and amount of available data. This paper reviews issues (1)-(4) within the framework defined by a class of GVF models for continuous variable mean estimators. We illustrate the primary idea with the U.S. Current Employment Statistics Program data from the years 2000 and 2010 respectively.