Business Employment Dynamics (BED) data are captured from a near-census of all businesses. Because these data are gathered for the entire population of businesses, there is no need to conduct a sample estimate; thus, the published data are not subject to sampling error. BED has published multiple articles on methodology, including the creation of BED, measuring establishment employment flows, size-class methodology, dynamic sizing, and establishment birth and death data.
Prior to the measurement of gross job gains and gross job losses, QCEW records are linked across two quarters. The linkage process matches establishments' unique State Employment Security Agency identification numbers (SESA-IDs). Between 95 percent and 97 percent of establishments identified as continuous from quarter to quarter are matched by SESA-ID. The rest are linked in one of three ways. The first method uses predecessor and successor information, identified by the states, that relates records with different SESA-IDs across quarters. Predecessor and successor relations can come about for a variety of reasons, including a change in ownership or the restructuring of a firm or a UI account. If a match cannot be attained in this manner, a statistical programming–based match is used. This match attempts to identify two establishments with different SESA-IDs as continuous. The match is based upon comparisons such as the same name, address, and phone number. Third, analysts examine unmatched records and matched records where appropriate.
In order to ensure the highest possible quality of data, SESAs verify with employers the industry, location, and ownership classification of all establishments and update any misinformation if necessary. Verification and updating are done on a 4-year cycle. Changes in establishment classification codes resulting from the verification process are introduced with the data reported for the first quarter of the year. Changes resulting from improved employer reporting also are introduced in the first quarter.
The sole estimation technique applied to the BED data series is that of seasonally adjusting the aggregated components. Seasonal adjustment is a statistical method used to facilitate the analysis of a data series across time by removing typical movements found at the same point of time from year to year. For example, we expect to see employment levels increase in June because most secondary and tertiary academic institutions have an extended break during the summer. The smoothing of seasonal fluctuations through seasonal adjustment highlights the true shift in the demand or supply of labor by removing well-behaved, predictable patterns from the series. The BED data are produced quarterly and are based on two of the series’ components: gross job gains and gross job losses. Gross job gains normally rise in the second and fourth quarters, whereas increases in gross job losses are expected in first and third quarters. All of the BED components are seasonally adjusted with the airline model, one of the most recognizable statistical methods in economic time series modeling.
The airline model was developed in the 1970s by the statisticians, George Box and Gwilym Jenkins. As the name suggests, the model arose from a project intended to forecast airline ticket sales, using past sales data as the only predictive variable. Because of its functionality and universality, the original autoregressive integrated moving average (ARIMA) (0,1,1)(0,1,1) airline model is applied to all of the quarterly BED time series. Also, because of limited resources and the vast number of BED series that are seasonally adjusted, the X-13-ARIMA program, originally constructed by the Census Bureau, is utilized through call functions in macros coded in the associated Statistical Analysis System (SAS) software. Finally, in order to maintain the additive properties of each series, gross job gains and gross job losses are indirectly seasonally adjusted by summing the seasonally adjusted expansions and openings components and the seasonally adjusted contractions and closings components, respectively. Net changes are calculated on the basis of the difference between seasonally adjusted gross job gains and losses.
The method of dynamic sizing is used in calculations for the BED size-class data series. Dynamic sizing allocates each firm’s employment gain or loss during a quarter to each respective size class in which the change occurred. For example, if a firm grew from 2 employees in quarter 1 to 38 employees in quarter 2, then, of the 36-employee increase, 2 would be allocated to the first size class (1 to 4), 5 to the size class 5 to 9, 10 to size class 10 to 19, and 19 to size class 20 to 49.
Dynamic sizing provides symmetrical firm-size estimates and eliminates any systematic effects that might be caused by the transitory, back-and-forth changes in firms’ sizes over time. In addition, it allocates each job gain or loss to the actual size class in which it occurred.
As mentioned earlier, because the BED data series are based on administrative data rather than a sample estimate, no issues related to sampling error arise. Nonsampling error, however, still exists. Nonsampling errors can occur for many reasons, such as the employer submitting corrected employment data after the end of the quarter or businesses making typographical errors when providing information. Such errors, however, are likely to be distributed randomly throughout the dataset.
Changes in administrative data sometimes create complications for the linkage process. These complications can result in overstating openings and closings while understating expansions and contractions. BLS continues to refine methods for improving the linkage process in order to alleviate the effects of such complications.
The BED data series are subject to periodic minor changes based on corrections in QCEW records, updated information on predecessors and successors, and seasonal adjustment revisions.
Annual revisions are published each year with the release of the first-quarter data. These revisions cover the last four quarters of data that are not seasonally adjusted and 5 years of seasonally adjusted data.
A primary input to the QCEW program is derived from employers’ mandatory reports of UI taxation information. Enforcement of these UI reporting requirements is substantial in order to ensure the solvency of the UI system. Thus, coverage for all UI-covered businesses is very high. Other sources of statistical error are controlled through significant editing and review of employment and wage data and of industry and geographic codes. Imputation for employment is very low, about 2 percent. The magnitude of the changes produced by revisions also is very low. Over the course of the yearly refinement cycle, during which analysts revise four quarters of data that are not seasonally adjusted and 5 years of seasonally adjusted data, the average magnitude of total revisions for the last 5 years, at the U.S. total level, has been less than ±50,000 jobs. The average magnitude of net revisions has been less than +25,000 jobs. For more information on BED methodology and the reliability of BED data, see technical note published with every BED news release.