Lies, Errors, Statistics

Barrons Gene Epstein makes an interesting observation regarding the recent jobless stats. He notes that extrapolating these numbers from partial data — as the Bureau of Labor Statistics does — is a fools errand, subject to frequent revisions:

“Employment gain may be 126,000 — or not. The Jobless Rate dropped to 6%,” read the AP wire Friday morning, “as payrolls grew by 126,000 last month.”

Narrow variations on that summary of October’s employment report could be found on the wires, in the dailies, and on the cable shows all through the weekend.

All of it wrong.

For starters, no one can say whether the unemployment rate “dropped” to 6.0% in October, even though the rate for September printed at 6.1%. That’s because, given the range of error on this number, a change of 0.1 of a percentage point is too close to call. Put another way, there’s even a fair chance the joblessness rate didn’t drop at all in October, but actually rose.”

That raises an interesting query: How reliable is the data we hear from the BLS each week and month? :

“The monthly unemployment numbers are taken from a door-to-door survey of about 50,000 households nationwide. Since this is supposed to be statistically representative of the more than 100 million households in the U.S., it’s possible to estimate the range of error on the rate derived from this sample.

From 6.1% in September, October’s result had to vary by plus or minus 0.2 of a point — to 5.9% or 6.3% — to be considered meaningful. And that only takes care of “sampling error.” Human error, too, can distort the results: wrong answers from respondents, wrong information recorded by the interviewers, incorrect processing of the data, and so on.

All of which explains why the Bureau of Labor Statistics itself called the unemployment rate for October “essentially unchanged.” Indeed, the only reason to believe 6% is about right is that it’s been essentially unchanged since August.”

But we got such terrific numbers yesterday. Is that why the market didnt respond? I say no, it was likely previously anticipated, and baked into the cake already. But Gene makes some compelling arguments the other way:

“If joblessness didn’t fall in October, neither did payroll employment rise by 126,000. That number is just a preliminary estimate, to be revised next month, to be revised again the following month — and to eventually go through a “benchmark” revision, based on a nearly complete count from unemployment insurance data.

In this very same report for October, the September rise in payroll employment is now recorded as 125,000, revised up from the 57,000 originally recorded in September’s report.”

Interestingly, Gene expects additional upwards revisions of employment numbers going forward. That stands in contrast to the past 5 weeks upwards revision in “New Unemployment Claims.” Or does it? Is the methodological bias towards extrapolation undercounting both new unemployment claims and new jobs? I’m not a statistician, but I do know a few (including a Nobel Winner >grin<). Perhaps the number wizards can share some insight into how reliable these numbers are . . .

Sources:

Barron’s
Lies, Errors, Statistics
Gene Epstein, Barron’s November 10, 2003
http://online.wsj.com/barrons/article/0,,SB10682385149224700,00.html

Bureau of Labor Statistics
http://www.bls.gov/bls/employment.htm