How should we interpret today's employment report, showing a growth in non-farm payrolls of 88K, a bit less than expected? There is also an increase in the unemployment rate and a different picture of job growth from the survey of individuals.
This is the real story, with as little technical stuff as we can manage. If traders or investors want to understand the data -- viewed by many as most important -- this is as good as we can make it for those without technical background.
Background: The BLS in Action
The most important thing for investors and traders to understand is that the reports are prepared by civil service employees using professional statistical methods. They do not change staff or opinions with shifts in partisan control. They are not biased in favor of a particular outcome.
Nearly all of those commenting on today's report, and particularly the Birth/Death Model, have never worked in a government agency and never done any forecasting using statistical methods. Do not listen to them! Since this is a subject where we can add value for the investment community, we have written extensively on this topic.
[Digression for the record -- I was "loaned" from the University of Wisconsin to the Wisconsin Department of Revenue for a couple of years a long time ago. The DOR did these forecasts for the state. I also taught graduate classes to mid-career government professionals. The courses covered public finance and research methods, among other things. So I know how these public officials think and what they try to do.]
Imagine that you are attempting to count all of the jobs in the U.S. economy, and you must do it EVERY MONTH. There is no way to count them. Compare this to voting, where in the 2000 election we had trouble actually counting the votes of people who showed up at the polls. (Remember the "hanging chad"?) There are 130 million jobs. The result you get is an estimate based upon surveys and statistical techniques.
The monthly job count is the result of a survey, using as your sampling frame the businesses that existed at the start of the year. This is subject to various sources of error, as follows:
- Sampling error. Even when you have a large sample size, there is always sampling error. It is small in percentage terms, but large in the aggregate. The sampling error for the jobs report has a standard deviation of about 60k jobs. (Check here for the best standard English description of this you will see online via my colleague Allen Russell). That means that 2/3 of the time the reported result will be +/- 60K. If you want a 95% confidence interval, you need to double that. This is the result after ALL businesses in the sample send in their forms.
- Revisions. As we discussed in our Employment Report Preview, the source of the revision is not any tinkering by the government. It is the result of businesses reporting late. Did you ever take an extension in filing your tax return? It has no meaning, unless you think there is a systemic bias to businesses that are late in filing.
- Non-sampling error. The original sample does not include new businesses, and it includes other businesses that have failed. Briefly put, this is a major problem unless there is an exact offset between the two.
So what does our earnest and professional government statistician do? The only major place to improve is the non-sampling error. The BLS discovered that they had serious errors without making some adjustment to the original survey. This was the reason for creating the Birth/Death model.
They state in the description of the methodology that there is a fairly constant relationship between births and deaths of businesses.
Earlier research indicated that while both the business birth and death portions of total employment are generally significant, the net contribution is relatively small and stable. To account for this net birth/death portion of total employment, BLS is implementing an estimation procedure with two components: the first component uses business deaths to impute employment for business births. This is incorporated into the sample-based link relative estimate procedure by simply not reflecting sample units going out of business, but imputing to them the same trend as the other firms in the sample.
It seems counter-intuitive to use business deaths to predict births, but they are looking at the data and we are not. None of the critics are either. Perhaps the BLS needs to share more about the underlying data.
The major point is that the BLS economists are making an honest and earnest effort to capture job changes missed by the survey.
They check their results with a benchmark revision. Eventually, they have an actual count of jobs from state data, and they update their model regularly to reflect this. They do an annual revision for the benchmark and a quarterly revision of their ARIMA time series modeling. (This is a standard method. Like other forecasters, we use it frequently. Readers should look with suspicion on the comments of anyone who does not seem to understand this method).
In short, the BLS staff is using standard professional methods to make the best possible estimate of job growth. It could be off by 60K on the sampling error, or more if the Birth/Death model is wrong. None of the critics offer a better method.
What the Birth/Death Model Contributes
The Birth/Death Model has improved the estimates. In fact, despite the constant criticism from those who claim that it adds "fictional jobs" the model actually underestimated job growth in recent years. That is why the benchmark revisions added jobs. It was an attempt to correct the non-sampling error, and it improved the estimate. The critics who said that the model over-estimated job growth were proven wrong by actual state data, used in the benchmark revision. The reaction of the critics was to ignore the actual count and criticize the BLS for "finding jobs." These critics should walk a mile in the moccasins of someone trying to provide useful data.
Please note that the Birth/Death model does not make a single jarring adjustment to the number of jobs in a particular month. The BLS goes back and spreads the jobs growth over the year, using the ARIMA model, in proportion to the year's data. Critics are challenged to offer a better approach....
The BLS warns that the methods are based upon the continuation of past trends. They do not look at any other sources of data about economic growth.
- Those drawing inferences from revisions do not understand the process, as we explained in the preview for this report. Doug Kass (we like him, we respect him, we read him, and we often profit from his short-term advice) made the point both on his daily blog and on Kudlow (video here) that there was the first downward revision in many months. This is meaningless and Doug should know better.
- There were various mis-statements about the Birth/Death Model. Some analysts tried to add the adjustment to the final NFP report. The BLS explicitly states that one should not add Birth/Death adjustments to seasonally adjusted data. Anyone doing this does not understand the method.
Q: Can I subtract the birth/death adjustment from the seasonally adjusted over-the-month change to determine what it is adding to employment?
A: No. Birth/death factors are a component of the not seasonally adjusted estimate and therefore are not directly comparable to the seasonally adjusted monthly changes. Instead, the birth/death factor should be assessed in the context of its effect on the not seasonally adjusted estimate.
- Some observers tried to focus on the construction job component, especially suggesting that the Birth/Death Model added an unrealistic number of jobs in this sector. The model does not specify the sector for job growth, so this statement is simply wrong. Many of those looking at the "internals" of the report lose track of a simple fact: This is a survey. There is an error band around any reported category. They tried to count all of 9 million jobs in construction and mining and such in one month and compare it to the count in the next month. There is a wide error band.
While we do think that the BLS is using a professional and technically sound method, it was obviously slow in catching growth during the expansion. We would not be surprised to see some modest downward benchmark revision when we have an actual job count for this year. This point was made by Tony Crescenzi today in his excellent blog on TheStreet.com's RealMoney (a paid site, but worth it for any serious investor). Tony uses the household employment survey to assist his analysis, since he thinks it might catch factors missed in the payroll report. We agree.
Econobrowser, a site that we read daily on RSS, takes a different approach. James Hamilton blends data from several report estimates to get a different projection. While this differs dramatically from our own forecast using Michigan Consumer Confidence, initial jobless claims, and the ISM manufacturing index, it is interesting and professional.
Meanwhile, the Fed is looking at these data in a long-term trend, in conjunction with other indicators. Chairman Bernanke, the Open Market Committee, and the 350 Fed economists all know how to interpret employment data. They do it much better than those on TV.
We would all like to have a monthly read on employment. I would also like to know how the Sox will do tomorrow or the Bulls will do versus the Pistons. It is important to realize the limitations of the data. There is a difference between the data we want and the data we need. It requires interpretation.
Many of the interpretive mistakes cited here will appear in Alan Abelson's column tomorrow.....