Dollars and Jens
Friday, February 05, 2010
The unemployment rate for January, based on the household survey, has been reported at 9.7%, down from 10% last month, with the participation rate somewhat higher, and an implication of 540,000 more employed people, though the payroll number, based on the employer survey, showed a statistically insignificant change in the number of jobs. The employer survey is popular with economists because of its smaller statistical error bars, but there are a couple of reasons I prefer the household survey (and none of them is because it's more optimistic).
Each set of numbers is, as indicated, based on a survey; it's not possible to go around finding out exactly who is employed by whom and tallying up all the numbers precisely. A certain number of known employers are surveyed, and a certain number of random individuals are surveyed. To get aggregate numbers of jobs created and so on, assumptions have to be made as to how many people there are, how many new businesses were created, how many businesses went out of business, and so on. It is particularly near turning points in the business cycle when the estimates used in the employer survey seem to be at their worst; they tend to be based on factors from the previous year or two, and thus the payroll survey usually looks bad for a while after jobs are being created (as the BLS underestimates the number of new businesses starting up) and looks good for a while after jobs are being lost. If what we're interested in is new news, instead of old news, this is a big flaw. Further, the biggest assumption that goes into the interpretation of the household survey is the change in U.S. population; if the population is presumed to be bigger than it is, it will appear that there are more employed people than there are, but it will also appear that there are more unemployed people than there are. This biggest assumption exactly cancels itself out in the calculation of the unemployment rate. Further, the unemployment rate, being scaled to the size of the economy, is a better indication of labor market slack; if the number of jobs is growing at the same rate as the population and labor force are growing, labor market conditions aren't changing.
The reasons for preferring the payroll survey do suggest that an unexpected 0.3% drop in the unemployment rate should be taken with a grain of salt, but — here I get all econometric — there is a positive serial autocorrelation between changes in the unemployment rate from month to month, which is to say that reported drops tend to be followed by reported drops and reported rises are followed by reported rises. A simple atheoretic ARIMA fit to the last 10 years of data predicts a reading of 9.7% again next month, with a slow decline after that, but the margin of error is comparable to the rate of decline; if I use a symmetric 50% confidence interval, the top of the interval stays at 9.8% well into the fall, though the optimistic end drops below 9% in the seventh month (the August rate, released in September).