While covering the Employment Situation Summary from the BLS on Sunday, we neglected to take an in depth look at the unadjusted employment data from which the widely reported seasonally adjusted numbers are derived. 

As you should all know, the summary employment report that's released on the first Friday of most months is actually two reports generated from two surveys; the first, the Current Employment Statistics (CES), which is more commonly known as the establishment survey, is a monthly polling of roughly 145,000 businesses and government agencies, or roughly 26% of all US employers, is conducted by the BLS and it gives us a reasonably close estimate (+/- 90,000) of how many jobs they've added each month, how many hours were worked, and how much the workers were paid.  The second report, the Current Population Survey (CPS), more commonly known as the household survey, is a telephone poll of roughly 60,000 households typically representing 110,000 working age individuals conducted by the census bureau for the labor department, which despite being quite detailed as to the reasons each individual's employment status, is subject to sampling and other errors that result in an 90% confidence level of +/- 300,000 in the monthly change in the number unemployed, and +/- 0.2% in the unemployment rate. Queries to either businesses or individuals for both of these reports focus on employment conditions as of the week that includes the 12th day of the surveyed month, but that's where the similarity ends.  The establishment survey only covers those workers who are on a business or government payroll, so it doesn't include the employed farmers or farm workers; it also doesn't include the self employed or individual proprietors, while the household survey includes all of these as employed, and in fact anyone over the age of 16 who is reported as working for pay even one hour during the reference week, ie, even a teenager who's mowing lawns for neighbors…

Now, after the employment data is collected by both of those surveys, the BLS runs each of them through a program which essentially compares that month's data to the changes for the same months from each survey over the past several years and adjusts the results for seasonal factors, holidays, and other unusual factors, and ultimately generates the press release and summary using only that adjusted data.  The reason this is done is so we can compare one month to another on a similar basis, without unusual factors such as the beginning and end of the school year or holiday hiring that would create large and not very meaningful changes in employment.  So there is no mention of the original actual employment numbers or other survey data in the report as released, every statement in the summary fairly much assumes that readers know it's been seasonally adjusted, and only a few of the tables at the end of the summary include unadjusted data.  Thus, when it's reported by the business press and most economic blogs, none of the actual raw employment data from which the report is generated is even mentioned.

In a similar manner, we ourselves typically report just the adjusted data and only mention the raw, unadjusted data from the two surveys in passing, when it seems that it might be useful for some perspective.  It was in so doing that we first noticed a major discrepancy between the number of new payroll jobs indicated by employer responses, and the number of employed as reported by households while covering the August report, which we headlined as a seasonal adjustment discrepancy, because the seasonal adjustment subtracted more than 200,000 jobs from the establishment survey and added nearly 500,000 to the household survey in August.  What happened in the August report was that the unadjusted data from the establishment survey indicated payrolls jobs increased by 378,000 from 135,583,000 in July to 135,961,000 (it’s since been slightly revised) while the not seasonally adjusted household data indicated that the count of the employed dropped 604,000, from 145,113,000 in July to 144,509,000, and the seasonal adjustments brought them into approximate alignment.  Even though we thought this to be an aberration, after consulting with several economists, we felt confident that the August misalignment would subsequently reverse itself in the next month or two, and the two surveys would be brought back into sync.  However, the September report compounded the discrepancy, as it again showed a payroll job gain of 612,000, which was lowered by the seasonal adjustment to 148,000, while the unadjusted household survey count of 142,000 employed was little changed by the seasonal adjustment.  The October report was even more alarming, in that there was a difference of 1,162,000 jobs in the seasonal adjustments between the the two surveys; the establishment survey seasonal adjustment subtracted 682,000 payroll jobs, while the household survey seasonal adjustment added 480,000 to the count of employed.  Thus, after three months we found that the unadjusted payroll job count had gone up by 1,963,000, while the unadjusted count of the employed went down by 967,000, only to have them brought into approximate alignment by the seasonal adjustments.  Still, because that data was affected by the shutdown, we expected this to be corrected in subsequent months, but as of the November and December reports it has not to any degree.  So this post is to update where the two unadjusted surveys stand as of the end of December, taking the minor revisions of previous months since into account.  We'll start by looking at the recent data as it’s available from FRED:

First, we have the unadjusted non farm payrolls as reported by employers for each month from July through December (000's):
2013-07    135577
2013-08    136002
2013-09    136612
2013-10    137523
2013-11    137999
2013-12    137753

Then, we have the raw unadjusted count of employed extrapolated from those who reported they were employed in the household surveys over the same time frame (000's):
2013-07    145113
2013-08    144509
2013-09    144651
2013-10    144144
2013-11    144775
2013-12    144423

It’s clear from the above that from July to December, the unadjusted count of the employed from the household survey fell by 690,000, from 145,113,000 in July, to 144,423,000 in December.  Over the same time frame, the unadjusted non-farm payrolls rose by 2,176,000, from 135,577,000 in July to 137,753,000 in December. After seasonal adjustments, the household survey was changed to show an increase of 301,000 employed over those same months, while the increase in non-farm payrolls was cut to 928,000.. For a visualization of what has been happening, we are including below a FRED graph which shows the historical track of the jobs or employed count from two surveys, before seasonal adjustments, since the beginning of the last decade. In blue, we have the unadjusted count of those self-reporting as employed from the household survey, with the count in thousands of employed on the graph’s right margin.  In red, we have the unadjusted count of payroll jobs as reported by employers taking part in the establishment survey, with that count on the left margin, such that the two tracks tend to overlap.  There’s obviously a lot of noise in the unadjusted data series.  What we want to focus on is the last five monthly changes, where we see the payroll jobs in red moving almost straight up until December, while the count of the employed in blue has been trending downward. Notice the red line is relatively well above the blue one for the first time in years. Ominously, the two other times that there was such a large increase in payroll jobs that was not accompanied by an equal increase in the number of employed were in 2008, and 2000-01, just preceding the recessions which are shown as grey bars on this graph.  We are not suggesting that this pattern predicts a recession, however; we're just noting that it’s inconceivable for these two lines, which are in effect measurements of the same function of employment, to continue moving in different directions indefinitely..

(cross-posted from MarketWatch 666)