Time Series Modeling of Official Statistics with Two-Stage Benchmarking

• Activity Number: 495
• Type: Invited
• Date/Time: Wednesday, August 4, 2010 : 10:30 AM to 12:20 PM
• Sponsor: Business and Economic Statistics Section
• Author(s): Richard Tiller*+ and Danny Pfeffermann
• Companies: Bureau of Labor Statistics and Hebrew University
• Keywords: Auto-correlated measurement errors ; Generalized least squares ; Recursive filtering ; Sampling errors ; Small Area Estimation

Abstract

The U.S. Bureau of Labor Statistics uses state-space models to produce monthly unemployment estimates for 9 census divisions (CD) and 50 states grouped within the CDs. The models are fitted to the time series of direct estimates produced from the Current Population Survey (CPS). Each model combines a model for the population values with a model for the sampling errors. The models, however, are slow to adapt to external shocks. To provide protection and for publication consistency, the CD model estimates are benchmarked each month to the CPS national estimate and the state estimates are benchmarked to the respective benchmarked CD estimate. The benchmarks in both steps are serially correlated with the model estimates. Our estimates account for errors in the benchmarks and their correlations and are shown to be design consistent. Application of the approach is illustrated using CPS data.