What is a shadow payroll
Employment and Unemployment
August 24th, 2004
"GOVERNMENT ECONOMIC REPORTS: THINGS YOU'VE
SUSPECTED BUT WERE AFRAID TO ASK!"
A Series Authored by Walter J. "John" Williams
"Employment and Unemployment Reporting"
(Part Two in a Series of Five)
August 24, 2004
The Bureau of Labor Statistics (BLS), U.S. Department of Labor, conducts two monthly surveys of U.S. employment and unemployment. Results usually are released on the first Friday of the month following the survey:
Household Survey (also Current Population Survey ) -- The household survey generates the unemployment rate from a statistically designed monthly sampling of roughly 60,000 households. Other surveys, such as the annual poverty survey, often are piggybacked on the employment questions. The survey measures the number of people who have jobs.
Payroll Survey (also Establishment or Current Employment Statistics Survey ) -- The payroll survey generates an estimate of the number of nonfarm jobs in the U.S. economy, based on a monthly non-random sampling of payroll tax filings of about 160,000 U.S. corporations and government agencies. The survey measures the number of jobs (some individuals hold more than one job).
The household survey is conducted during the week that includes the 12th of the month. The payroll survey is conducted as of the payroll period that includes the 12th of the month. Other than for seasonal factors, the household survey gets revised only with series or population redefinition. The payroll series is revised for two months following the initial release and then again in an annual benchmark revision.
Where the household survey includes farm workers, the self-employed and workers in private homes, the payroll survey does not. The payroll survey counts jobs, making no adjustment for multiple jobholders. Yet, adjusting for all differences, the BLS never has been able to reconcile the two series within one million jobs.
Conventional wisdom in the financial community is that the payroll survey is more accurate, given its larger sampling base. To the contrary, the household is scientifically designed, and the error can be estimated to any degree desired. The payroll data are haphazard at best, and the BLS has no idea of potential reporting error.
The BLS estimates a 90% confidence interval for a change in the unemployment rate of ±0.22%, and a 90% confidence interval for the monthly change in payrolls of ±108,000. The BLS, however, admits the payroll survey's confidence interval is not solid, given built in biases and the lack of randomness in the monthly sample.
The payroll survey used to include a regular monthly bias factor of about +150,000 jobs. Those jobs were added each month for good measure, as an estimate of jobs created by new companies. Companies that went out of business generally were assumed to be employing the same number of people as before they went out of business.
In the last couple of years, the BLS has modeled and seasonally adjusted its bias factor; there is no more guesstimation. Accordingly, new monthly bias factors have ranged from -321,000 to +270,000 during the last year. This, combined with continuous seasonal adjustment revisions, has added to the volatility of recent monthly reporting.
Suggesting that the household survey is more accurate than the payroll survey, however, does not mean household survey accurately depicts unemployment. While its measures have definable statistical accuracy, the accuracy is related only to the underlying questions surveyed and to the universe of people surveyed.
The popularly followed unemployment rate was 5.5% in July 2004, seasonally adjusted. That is known as U-3, one of six unemployment rates published by the BLS. The broadest U-6 measure was 9.5%, including discouraged and marginally attached workers.
Up until the
Clinton administration, a discouraged worker was one who was willing, able and ready to work but had given up looking because there were no jobs to be had. The Clinton administration dismissed to the non-reporting netherworld about five million discouraged workers who had been so categorized for more than a year. As of July 2004, the less-than-a-year discouraged workers total 504,000. Adding in the netherworld takes the unemployment rate up to about 12.5%.
The Clinton administration also reduced monthly household sampling from 60,000 to about 50,000, eliminating significant surveying in the inner cities. Despite claims of corrective statistical adjustments, reported unemployment among people of color declined sharply, and the piggybacked poverty survey showed a remarkable reversal in decades of worsening poverty trends.
Somehow, the Clinton administration successfully set into motion reestablishing the full 60,000 survey for the benefit of the current Bush administration's monthly household survey.
While the preceding concentrates on the numbers that tend to move the markets, the household survey also measures employment. The payroll survey also surveys average hourly and weekly earnings and average workweek.
Addendum to Installment One (Published 9/7/04)
Bureau of Labor Statistics' Correction to Payroll Survey Description
In response to my comments on the "non-random" and "haphazard" nature of the payroll employment survey in Installment One, the Bureau of Labor Statistics (BLS) advised that my information was outdated, that the payroll survey used scientifically designed probability sampling, which had been phased in over several years and completed as of June 2003.
I was aware of the changes to the system, but did not think they improved the quality of the reported results much. I have just reviewed the BLS's current sampling methodology and have not changed my mind. While I may have used inaccurate terminology in describing the sampling method for the series, my general comments remain, and I still believe the household survey to be the more accurate of the two.
The household survey is proactive in nature and designed and sampled so its results can be determined with measurable statistical confidence.
While the payroll survey sampling approach may be sounder statistically than it was several years ago, it still is responsive, in nature, subject to whatever is reported or not reported by U.S. corporations. While individual companies are selected at random for following, the universe they are selected from still is not random and can vary meaningfully with changing times. An element of haphazardness is inherent in the universe of reporting companies.
During a recession, for example, firms go out of business and stop reporting, but the BLS does not know whether a company is out of business or did not report for some other reason. This supposedly is accounted for by the business birth (creation)/death (going out of business) modeling of companies, which replaces the old bias factor system.
There is no way to model these numbers with any meaningful accuracy, and the monthly swings in the birth/death data now often are greater than the reported monthly changes in total payrolls.
The BLS has a Herculean task in trying to measure monthly payrolls with meaningful results, and it has expended significant effort to improve its system. Nonetheless, it is difficult to see noticeable improvement in monthly reporting quality. Contrary to BLS expectations of improved results, I would be extraordinarily surprised if revisions to the series don't get larger, as opposed to smaller, as a result of what now is probably over-modeling of the series.
This already is evident in the monthly revisions to some individual industry series that I follow closely. It will be interesting to see how large the next several annual benchmark revisions are for the new system.Source: www.shadowstats.com