Seasonal Adjustment of Hybrid Economic Time Series

1. Introduction State industry employment is estimated monthly from the Current Employment Statistics survey, a sample of about 380,000 employers, and seasonally adjusted with X-11-ARIMA. An annual benchmarking process revises estimates to reflect universe counts available from administrative records of the Unemployment Insurance (UI) programs of each state. At any point in time, the current series consists of universe data through the latest benchmark month followed by sample data up to the current month. A straightforward application of X-11-ARIMA to this hybrid series gives projected seasonal factors which are heavily influenced by the universe data, but which are applied to sample data. Distortions can occur, because the two data sources historically have displayed different seasonal patterns. Beginning with January 1994 data, the U.S. Bureau of Labor Statistics (BLS) implemented an alternative method that separately adjusts each part of the series, an approach first carried out by Berger and Phillips (1993). The decision to implement the alternative method, which we refer to as the two-step method, was based on the evaluation reported in this paper. The major users of the employment statistics include the Federal Reserve Board, the President's Council of Economic Advisors, the Joint Economic Committee of Congress, and various other policy-oriented groups. Where economic statistics are used as the basis for their policy analysis, it is important that the preliminary estimates be accurate and that the economic information found in these data be discernible. Highly variable economic series make the interpretation of such data difficult. Furthermore, large annual revisions to the data may also impact the validity of policy analysis conducted on the original estimates, as Berger and Phillips (1994) suggest. Our analysis of the seasonal adjustment of state industry employment statistics compares the two-step method with the combined method formerly used (Shipp and Sullivan, 1992), i.e., a basic application of X-11-ARIMA to the hybrid series. Our findings are: