Smoothed-Seasonally-Adjusted Estimates (SSA) Questions and Answers

  1. What are Smoothed-Seasonally-Adjusted (SSA) Estimates?

    SSA estimates are seasonally-adjusted estimates that have incorporated a long-run trend smoothing procedure. This results in estimates that are less volatile than those currently produced by the LAUS estimation methodology.

  2. Why is BLS replacing official seasonally-adjusted estimates with SSA Estimates?

    State users of BLS data have expressed concerns regarding the month-to-month volatility in their official seasonally-adjusted employment and unemployment series since the third generation LAUS models were introduced in January 2005. The SSA procedure is effective in reducing the number of spurious turning points in the current estimates. In addition, the SSA methodology represents an innovative alternative to the current annual revision process that incorporates an annual historical benchmark for seasonally-adjusted estimates. This new approach will address longstanding issues related to end-of-year revision and will enhance the analytical capability of the estimates.

  3. When will BLS introduce the SSA Estimates?

    We will implement the SSA methodology with annual revision of 2009 estimates. Seasonally-adjusted labor force estimates from January 1976 through December 2009 will be replaced with the SSA estimates for all four census regions, nine census divisions, all States, the District of Columbia, New York City, the Los Angeles-Long Beach-Glendale metropolitan division and the respective balances of New York and California. Estimates from January 1983 through December 2009 will be replaced for five additional substate areas (the Cleveland-Elyria-Mentor and Detroit-Warren Livonia metropolitan areas and the Chicago-Naperville-Joliet, Miami-Miami Beach-Kendall, and Seattle-Everett-Bellevue metropolitan divisions) and the respective balances of Ohio, Michigan, Illinois, Florida, and Washington. These data will be made available to States on February 17, 2010 and available on the BLS website on February 26, 2010.

    The SSA methodology will be reflected in monthly State labor force estimates beginning with January 2010. State estimates for January 2010 will be issued on March 10, 2010.

  4. What are the sources of volatility in the current estimates?

    Various sources of volatility in the month-to-month movement of the LAUS labor force estimates have been identified. These range from CPS sampling error and potential outlier observations to more obscure artifacts of the model-based estimation methodology. The latter includes the structure of the trend component model and the real-time benchmark effects. Research focused on ways to reduce the size and variability of the real-time benchmark adjustments and the variability of the seasonally-adjusted estimates themselves.

  5. How does the trend component model influence volatility?

    The structure of the current variable coefficient trend component model, which allows both the intercept and the slope terms to change over time, contributes to the month-to-month volatility of the LAUS estimates. Changes in the intercept term may cause up and down fluctuations in the normal trend level that can overwhelm the gradual changes in the slope term.

    In addition to the variable coefficient formulation of the trend model, there is another characteristic of these models that contributes to volatility. It is the fact that the “trend” component model is really a “trend-cycle” component model, where “cycle” refers to the irregular variation in the trend intercept term and the “trend” part represents the long-run smooth (slope) term of the model. When the trend-cycle component model is too noisy, the resultant seasonally-adjusted estimates contain fluctuations of little interest to the data user. These fluctuations come from the shifting intercept level as well as volatility associated with extreme CPS observations.

  6. What occurs during annual processing that introduces volatility into the estimates?

    There are three sources of revisions to estimates during annual processing that can introduce volatility. These are revisions of input data, model re-estimation, and controlling estimates to historical control totals.

  7. What is meant by “revision of inputs”?

    After completion of preliminary December estimates, States have the opportunity to revise their inputs to the models. These include the Current Employment Statistics (CES) payroll employment estimates and the unemployment insurance (UI) claims. Also, revised population controls are applied to the CPS estimates of employment and unemployment.

  8. What occurs during model re-estimation?

    During the year, a concurrent estimator is used to produce real-time estimates. This means that the estimate for the latest month uses all available CPS data, but previous months’ data are not revised to reflect CPS data available in later months. During the annual revision process, each of the 12 monthly model estimates for the previously completed year is revised, using all available data through December of that year. A seasonally-adjusted (SA) series is developed by removing the seasonal component from the not-seasonally-adjusted (NSA) series.

    The switch from a concurrent to historical estimator generally has a larger effect on estimates made early in the year than toward the end of the year. At the end point of the series, in December, the concurrent and historical estimators use the same amount of CPS data and would produce identical estimates if there were no revisions to the data. The overall effect of historical re-estimation is to produce estimates that look much smoother than the previous estimates.

  9. What is meant by “benchmarking to historical control totals”?

    Historical re-estimation of the models results in revisions to both the NSA component and the SA component. Two different benchmarking processes are used. NSA estimates are benchmarked to monthly Division model controls (which have been controlled to the monthly NSA national CPS estimates). (See #9, below.) SA estimates are controlled to the annual average of the NSA estimates.

  10. Why are different control totals used for the NSA and SA series?

    The use of the monthly Division estimates of employment and unemployment as the control totals for the NSA estimates ensures that the monthly State NSA estimates sum to the national CPS estimate. As in current estimation, pro-rata factors are used to do the adjustments. The pro-rata factors fluctuate from month to month. This introduces additional variability in the monthly benchmarked NSA estimates. Since monthly variability in the benchmarked NSA estimates is dominated by seasonality, the variability added by benchmarking has a relatively small effect.

    This is not the case with historical SA model estimates, which are much smoother than the NSA estimates. Since month-to-month change tends to be small in the model estimate when the seasonal component is removed, monthly benchmarking would dominate the change in the historical benchmarked SA estimates. To avoid this problem, annual average controls are used to benchmark the historical SA series.

  11. How is the annual average control used for the SA historical series?

    The annual averages of the re-estimated and re-controlled NSA estimates are used to control the monthly SA model estimates. In this process, the Denton procedure is used. It adjusts the monthly data to exactly satisfy the annual benchmarks while preserving the month-to-month movements in the original model estimates as much as possible. This process preserves the underlying smoothness in the model estimates that would be lost by monthly benchmarking.

  12. Are there issues associated with using different benchmark controls for the historical NSA and SA series?

    Yes, using a different benchmark procedure for historical SA estimates can produce breaks from December to January of the current year.

  13. Did particular issues arise with the 2008 annual processing?

    Yes. The annual revision process resulted in a general lowering of seasonally-adjusted unemployment rates for the latter months of 2008. Benchmarking to control totals had the most impact on the estimates in the latter months of 2008. The downward revisions in the SA series in December, during a period of steep economic decline, impacted comparisons between December 2008 and January 2009, which saw a marked rise in unemployment.

  14. What is BLS proposing to address the historical benchmarking issue for SA estimates?

    The first two steps in the annual revision process (revising model inputs and re-estimating monthly estimates) are unchanged.

    The last step—benchmarking to control totals—will be revised for SA estimates. We plan to drop the use of the annual average of the NSA series as the control total. Instead, as in concurrent estimation, the SA series will be adjusted by the same pro-rata factor used in adjusting the NSA estimates to the national controls totals of employment and unemployment. Since this process will introduce volatility into the historical SA series, we will smooth the SA series following the application of the pro-rata factors.

  15. What smoother is used for the SA series?

    The estimates are smoothed using the Henderson Trend Filter (H13). It suppresses irregular variation in real time. The H13 has a long history of smoothing seasonally-adjusted estimates in the X-11 and X-12 seasonal adjustment software to produce trend estimates. The H13 uses a filtering procedure, based on moving averages, to remove the irregular fluctuations from the seasonally-adjusted series, leaving the trend. Symmetric moving averages are used to smooth the historical series while asymmetrical averages are used in real time. The symmetric filter sees both the future and the past, has 13 terms, and does not over smooth the trend cycle. The asymmetric filter is used in real time as each current month’s estimate is an end point. It has seven terms and does not introduce lags in identifying turning points.

  16. What series will be affected and replaced?

    The entire series of seasonally-adjusted estimates for all States, the District of Columbia, New York City, the Los Angeles metropolitan area, and the respective balances of New York and California will be replaced from January 1976 through December 2009. Seasonally-adjusted estimates for the Cleveland-Elyria-Mentor and Detroit-Warren-Livonia metropolitan areas and the Chicago-Naperville-Joliet, Miami-Miami Beach-Kendall, and Seattle-Bellevue-Everett metropolitan divisions and the respective balances of Ohio, Michigan, Illinois, Florida, and Washington, will also be replaced from January 1983 through December 2009.

    Not-seasonally-adjusted estimates will not be impacted by this change in methodology.


Last Modified Date: March 3, 2010