Tuesday, December 08, 2009

November 2009 job losses: BLS vs. ADP

Via Mark Thoma, FT Alphaville questions the accuracy of the BLS payroll numbers released on Friday:
Just how amazing were the US payroll numbers released on Friday?

So amazing they’re verging on the (perish the thought) unbelievable, according to some analysts.

The consensus forecast among analysts for the November job loss had been -130,000, with even the relatively optimistic and sometime-clairvoyant economists at Goldman Sachs forecasting -100,000. The official data showed a fall of just 11,000 — about 90 per cent fewer than the consensus estimate.
Many months ago I began tracking the Automatic Data Processing payroll numbers in addition to those from the Bureau of Labor Statistics in order to satisfy those conspiracy theorist commenters who just don't trust the government's data. For November 2009, ADP and the BLS disagree by over 150,000 job losses.

Here's my graph of BLS data for this recession:

Here's my graph of ADP data for this recession:

Update: I just noticed that the BLS measures both private and government payrolls while ADP only measures private payrolls. However, even just comparing private payrolls, the disparity is roughly the same. ADP says there were 169,000 private nonfarm job losses in November, while the BLS says there were only 18,000 private nonfarm job losses—a 9-fold difference.


  1. Taking ratios of the numbers doesn't seem like it is entirely valid, and really only serves to artificially magnify the change.

    Taking the difference seems like a much more meaningful way to compare these things. In fact, you could plot the difference in the ADP and BLS numbers going back to Dec 07 to see if there are any interesting patterns.

  2. Yeah, I agree with JackRussel. You're creating a ratio of the 2 change in employment numbers, but since they're based on the change in a much larger magnitude number (total private sector jobs) this ration doesn't really tell you much. A very small difference in how you measure the jobs in the whole market and/or in your margin or error (since as I understand it these numbers are all based on relatively small sample sets) could result in huge multiples for this ratio. This is especially error prone as the numbers approach 0. Imagine BLA saying we lost 1 job and APD saying we lost 1000. That would mean a 1000x multiple. But really, they 1 numbers are extremely close when considered against the size of our employment population

  3. Good point. I'll point out that I posted both the difference in job losses and the ratio.

    My use of the ratio was a comparison to the source's "about 90 per cent fewer than the consensus estimate" statement, which is also a ratio.

  4. yeah, not really criticizing the post. JackRussel's post just spurred me to think the problems the ratio could have. On first glance it didn't even occur to me.