September’s increase in payroll jobs of 110,000 was good news, and equity markets hit records. But the upward revision in the August employment numbers from an initial report of -4,000 to the current estimate of +89,000 shows that you can’t base economic policy on the first employment estimate, even if you like the number.

The September 7 announcement of the first decline in payroll jobs in four years from the Bureau of Labor Statistics generated headlines such as “Weak Jobs Data Boosts Likelihood of Fed Cut,” “Dollar Plummets After Jobs Data,” and “Recession 2008?”

Turns out BLS was wrong after all. Last Friday BLS revised up not only the August payroll numbers, but also the July numbers, which turned out to be+93,000, 25,000 more jobs than was thought on September 7.

At the same time, as part of its annual benchmark revision, where the jobs estimate is compared to the actual administrative count of jobs reported by employers to state unemployment insurance agencies, BLS announced that 297,000 fewer jobs had been created over the period April 2006 to March 2007.

The last month saw not only havoc in the financial markets, but also a half point cut in the Federal funds rate. That cut was based on the view that the economy was weakening, and the -4,000 number greatly contributed to the perception. Credit market conditions made a cut likely, but the employment data gave the Fed additional “evidence” because inflation dangers could be put on hold.

What to do?

Measuring payroll jobs is not easy. Payroll employment averages over 138 million, and the monthly jobs change that BLS is trying to measure amounts to a 1% variation in that total. Estimating a small change against such a large scale is like trying to count the weekly change on a crowded beach from an airplane flying over 30,000 feet.

But it may be possible to do better. Here are three suggestions to improve the accuracy of BLS data and thereby to reduce the extent of policy-making fiascos.

Americans get their unemployment data faster than residents of any other country. In today’s fast-changing labor market, we may be asking BLS to do the impossible if we expect both 100% accuracy and speedy reports on the first Friday of each month.

BLS could delay the release of the data by two weeks to allow more of the sample to be collected. The initial estimate would then be more accurate. Rather than being able to analyze 60% to 70% of the 400,000 firm payroll sample, BLS then would have 70% to 80%.

With a two-week delay, we’d get our information on employment about the same time that we get readings on retail sales, housing starts, and inflation. No one seems to complain that those data come out late.

Second, BLS could keep the same early release date, but improve the quality of the data by calling those firms who send their employment data to BLS through the mail or by using touch tone phone service rather than filing electronically. This would result in about 8% more of the sample included in the first monthly data reading, at a cost of approximately $3 million.

Third, BLS could fit the estimated employment count to the actual job count every quarter rather than just once every year. The current method is to find the annual discrepancy, divide it into 12 monthly amounts, and revise the data back a year through addition or subtraction, a process known as benchmarking. Last Friday’s revised benchmark figure of -297,000 is a preview of the final number which will be announced in February 2008.

The turnover that is characteristic of our modern, competitive, and flexible labor market has added new challenges to the task of estimating monthly total payroll jobs changes.

Given the competition for employers’ time and resources, BLS does a remarkable job of achieving their survey results. Yet it may be possible to make a few small changes and do even better than we’re doing now. We should give it a try.

Ms. Furchtgott-Roth, dfr@hudson.org, is a senior fellow at the Hudson Institute and former chief economist at the U.S. Department of Labor.