Revisions to jobs data: What you need to know
During the dog days of August — as commentators were already wondering whether the nearly decade-long economic recovery was on its last legs — government number-crunchers made an announcement that jolted economists.
The Bureau of Labor Statistics announced that, according to preliminary data, 501,000 fewer jobs had been created between March 2018 to March 2019 than officials had initially estimated.
No one lost their job over it, because adjusting the numbers isn’t unusual — in fact, the government does it routinely. Sometimes the numbers go up, and sometimes they go down. In this case, the government realized that it had been too generous in its attempt to quantify the number of people employed over that year-long period.
When this happens, economists throw out the previous employment figures and replace them with newer, more accurate measurements.
The process can be confusing if you’re not a credentialed labor economist.
Here’s how it works.
The initial numbers are released on Friday mornings after the start of a new month. They are usually met with significant media coverage, because they estimate how many net jobs were created or lost during the previous month.
The count is generated by BLS’s Current Employment Statistics survey, a study based on payroll data reported by employers. (The other familiar statistic released every month, the unemployment rate, comes from a different survey that asks questions of a much smaller sampling of individuals. It is not revised like the job numbers.)
According to BLS, Current Employment Statistics survey includes a random sample of about 142,000 businesses and government agencies. All sizes of businesses are included, with companies with fewer than 20 employees accounting for about 41% of the entities surveyed. Participating companies are asked to provide data for the pay period that includes the 12th of the month.
By the standards of surveys, the Current Employment Statistics survey is quite large, so the initial numbers announced are generally considered pretty solid.
But they are not perfect, and that’s the reason why revisions are needed.
The revised numbers are more reliable because they include more thorough data.
The data for the initial announcement is collected over a window of just nine to 15 days, and some businesses have trouble getting their data to BLS that quickly. So BLS continues to accept information over the next two months, and these figures are factored into the revised numbers.
To explain, let’s look at some actual recent numbers.
The change in non-farm payroll employment between May and June was initially reported as an increase of 224,000 during the July report. A month later, the June gains were lowered to 193,000. Two months later, the number was downgraded again to 178,000.
All in all, there was still an increase in jobs between May and June. But BLS’s measured increase shrank by 46,000 between the initial and final estimates.
Over time, initial BLS estimates have become more accurate.
Between 1979 and 2003, the average variation between the initial and final numbers was 61,000 jobs.
Since 2004, the average variation has been 40,000 jobs.
Economists say they pay some attention to the monthly revisions, though they urge against obsessing over them.
"I watch revisions with some concern, particularly when there's a string of negative revisions," said George Washington University economist Tara Sinclair.
As it happens, the past few months have produced exclusively downward revisions. Every revision in March, April, May, June, and July has been downward.
However, the size of the final estimate — rather than the nature of the revisions along the way — is probably most important, said Jesse Rothstein, a University of California-Berkeley economist.
Those have been a bit lower, but for now, at least, the change in net jobs has been positive every month, rather than falling below zero. Negative job numbers haven’t occurred since the depths of the Great Recession.
"Negative revisions to payroll employment are a concern, but the revised job growth numbers aren't dropping below zero, so it's not signaling a recession as of yet," Sinclair said.
Potentially more worrisome is the annual benchmark revision.
There’s a second type of revision that happens once a year.
This revision, known as the "benchmark" revision, relies on a different data source: unemployment insurance tax records. These are considered more accurate than the main survey, because they include employer records for about 10 million establishments. However, they are produced too late to be included in the first two rounds of monthly revisions.
This revision is particularly helpful in filling in the gaps caused by "births" and "deaths" of companies. BLS won’t know whether a non-response to its survey from a firm means that the firm simply didn’t send its data that month or whether it went out of business. The unemployment insurance data helps ameliorate that problem.
Over the past 20 years, the largest revisions have added almost 800,000 net jobs to the prior year’s total or subtracted about 900,000 jobs from that year’s total. In most years, however, the revisions are much smaller.
While the final benchmark revision of the 2019 employment statistics will be released in February 2020, BLS released a preliminary revision in August — and it wasn’t good.
The preliminary report found that the economy created 501,000 fewer jobs than had initially been reported for the year between March 2018 and March 2019. If losses on that scale hold up in the final numbers, that would represent the second largest downward revision in the past two decades.
Martha Gimbel, the former research director at jobs site Indeed.com, said revisions can be large "any time the economy is more volatile. These revisions were particularly large in the retail sector, which likely reflects the convulsions that sector has been undergoing."
Dean Baker, senior economist at the liberal Center for Economic and Policy Research, wrote that "it is unlikely that we will find that we were actually in a recession between March of 2018 and March of 2019, but add this to the list of worrying data points."