By Jeffrey Miller
Regular readers of "A Dash" may be surprised to see that I have objections to the most recent payroll employment report results from the Bureau of Labor Statistics. Ironically, my objections come at a time when many critics say it is a "clean" report. In addition, I think that the problem relates to the measurement of job creation.
Some Background
I hope that I have established some credibility on this subject. The BLS has a method for estimating the monthly job change, including job creation. For several years I have insisted that the right way to keep score was to look at the final results, which we eventually know from state employment data, and test the estimates against those results.
Until recently, the results were excellent.
Something happened. It did not happen at the onset of the recession, as many critics predicted. In fact, the BLS method had worked through the 2001 recession, something that everyone ignored.
It did not even happen in Q408, at least not very dramatically. The problem showed up in Q109, as I reported and discussed (here) and in my November preview.
Let's repeat my recent review of the BLS method for estimating job creation. If you take a moment to read this carefully, you will see why the critics were wrong before, and are also missing the problem now.
How the BLS Handles Job Creation
The BLS approach is to make an estimate of the total payroll jobs in one month, make another estimate for the next month, and subtract the two to determine the change. They use an excellent and sophisticated survey technique to do this. Their historical record, judged by the eventual count from the states, has been very good -- until quite recently.
The Survey Problem. Any time you do a survey, there will be non-respondents. When the question is something like "How many people favor health care with a public option?" the non-respondent problem takes a simple form. You need only ask whether the non-respondents are similar to those who actually answered. Most polls make this assumption.
The employment question is qualitatively different. We are not asking the opinions of non-respondents. We are asking whether they are even still in business. If the BLS were to assume that non-respondents had all ceased operations, they would seriously underestimate total employment. Historical data conclusively show that the non-respondents are split between those who did not answer and those who are out of business. The data also show that new job creation, running at about 2 million jobs per month even in recessions, are a predictable function of dying businesses.
Let me emphasize the difficulty. There are always non-respondents to the voluntary survey, despite the best efforts to get everyone. If the BLS assumed that the non respondents were all lost jobs, and that the impact was proportional, we would see a loss of 13 million jobs per month, a silly result. Instead they attempt to impute business deaths and births. At one point, they assumed a business birth for every death. This is the natural result from extrapolating the sample to the entire population.
This is not the +/- 100K jobs from sampling error; it is non-sampling error. This means that the non-respondents are different in an important way from those who answer the survey. We know this to be true, so the problem is how to compensate.
The Job Creation Estimation. Because of this, the BLS employs a two-step process. The imputation step forecasts job creation from job destruction, and includes a cyclical component.. The Birth/Death adjustment, (the only thing cited by most critics, who ignore the more important imputation step), is a residual. For many years this residual was stable. The most recent test against the state data indicated a significant error, showing that the BLS estimates have been wrong for nearly a year, especially since Q1 09.
Read the full article here.
No comments:
Post a Comment