Statistics corner: nonfarm payroll employment data: benchmark revisions and improvements in methodology

Business Economics, Oct, 1996 by Patricia M. Gregg

Because of recent sharp reactions in the financial markets to the nonfarm payroll employment data released early each month, we have asked the Bureau of Labor Statistics to provide us with some detail on the revisions and the reasons for the seemingly big adjustments. (The changes in the numbers of jobs may seem large, but they are less than 1 percent of total payrolls.)

Martin Fleming Editor, The Statistics Corner

On June 7th the Bureau of Labor Statistics released its May employment report along with annual benchmark revisions to previous years' data; the report signaled robust job growth, providing good news on the economy and touching off strong reactions on Wall Street. This monthly measure of nonfarm payroll jobs is one of the earliest and most closely watched economic indicators in the United States. The data are both an important economic indicator in their own right and input into other key series, including the national and state personal income estimates, the index of industrial production, and productivity measures.

The nonfarm payroll employment estimates are a primary product of the Current Employment Statistics (CES) program, along with measures of the average weekly hours and average hourly earnings; all are provided in abundant industry and geographic detail. The BLS surveys nearly 400,000 business establishments nationwide each month to form the basis of the CES series. The sample results along with a model-based "bias adjustment," intended to correct for absence of new business births and other biases in the current sample, are the key inputs to the published monthly estimates.

While the preliminary estimates of employment published on the first Friday of each month are the most eagerly anticipated and widely reported, the final economic time series produced for nonfarm payroll employment is the product of not only monthly sample survey results but also an annual recomputation process known as benchmarking. This process can have a significant impact on previously reported employment figures; this most recent benchmark added over a half-million jobs to the employment count for the March 1995 reference point, with additional upward adjustments in all the subsequent months.

The annual benchmark revision is based on comprehensive universe counts of employment, or benchmarks. They are derived primarily from information reported on unemployment insurance (UI) tax reports that nearly all employers are required to file with state employment security agencies. These data provide the opportunity to evaluate sample-based results against full population counts, on a lagged basis. In recent years the size of annual benchmark revisions, as well as the magnitude of bias adjustment required to the sample survey results, prompted BLS to initiate a comprehensive program of research and improvement for the CES sampling and estimation methodology. Initial implementation of improved techniques are scheduled to begin following the next benchmark revision, in July 1997, and be phased in over several years.

In addition to annual benchmark revisions, effective with the May 1996 release, the BLS also introduced refinements to its seasonal adjustment procedures. The new procedures are designed to control for the effects of varying survey intervals (also known as the four- vs. five-week effect), thereby providing for improved measurement of over-the-month changes and underlying economic trends.

The remainder of this article will provide more detail on the effects of the 1995 benchmark revisions as well as current and future improvements to CES methodology.

EFFECT OF THE 1995 BENCHMARK REVISIONS

The March 1995 benchmark revision for total nonfarm employment was an upward adjustment of 542,000, generating an employment level 0.5 percent above the previously published sample-based estimate. This marked the second year of substantial upward adjustment; last year's March benchmark revision was an upward adjustment of 747,000 or 0.7 percent. Benchmark revisions have averaged [ or -] 0.3 percent over the past decade with a range from zero to [ or -]0.7 percent (See Table 1).

[TABULAR DATA FOR TABLE 1 OMITTED]

Following standard methodology, the March 1995 UI-based benchmark employment level has replaced the March 1995 sample-based employment estimate; its incorporation resulted in recomputation of employment levels for both the year preceding and the year following the benchmark reference month. The previous year (April 1994 - February 1995) was adjusted by wedging back the difference between the March 1995 benchmark level and the March 1995 sample-based estimate (i.e., the 542,000). The wedging process is linear - 1/12 of the difference was added to April 1994, 2/12 to May 1994, and so forth up through February 1995, which received 11/12 of the difference. This linear wedge technique assumes that the sample-based measurement error accumulated at a steady rate since the last benchmark.

New estimates were computed for each month subsequent to March 1995, by linking from the new benchmark level. This postbenchmark revision raised still further the estimate of job growth in the recent past. By February 1996 the upward revision to previously published employment levels was 767,000 on a not seasonally adjusted basis; this further upward revision is mainly a function of changes to bias adjustment levels. Average monthly bias adjustment levels for the April 1995 to March 1996 time period were revised up an average of 20,000 per month from 110,000 to 130,000 per month, reflecting evidence that the original estimates were understating employment growth when compared to the most recent UI universe counts. In total, the benchmark and postbenchmark revisions raised the estimated employment growth approximately 15 percent higher than previously published figures for the two-year period.