Robert Barro tosses in his $.02 about the divergence between payroll and household employment growth:

since the peak of payroll employment in March 2001, household employment has risen by 700,000, while payroll has fallen by 2.4 million, so that household did better by 3.1 million.

A less well-known fact, also shown in the graph, is that the two surveys diverged in the opposite direction in the 1990s. The discrepancy built up until July 2000, when payroll was 22.3 million and household 17.1 million above their respective levels in January 1993. Thus, payroll did better by 5 million jobs…

…In fact, the reported number of self-employed was smaller as a ratio to household employment at the start of 2004 than it was in the mid-1990s. In any event, when the BLS considers self-employment and other measurable differences between the two surveys, they explain only 200,000 to 400,000 of the extra three million jobs in the household survey.

…the main differences between the two surveys cannot yet be explained…

Barro has what sounds like a vector autoregression to predict payroll employment.

I use a forecasting equation for payroll employment growth that considers the history of employment, GDP, and unemployment-insurance claims. By chance, this equation was pretty accurate for the latest reading, an increase by 21,000 jobs in February. My forecast was for a rise by 6,000 jobs. I attribute this accuracy to chance because a reasonable error range for the forecast was very broad: -152,000 to +165,000.
…More reliable forecasts can be generated for longer periods, such as yearly average payroll employment growth. In this case, unemployment-insurance claims lack predictive power but the history of GDP is more important…The forecast for February 2004 to February 2005 is an employment growth rate of -1.1%, with a range of -1.6% to -0.7%. This growth forecast translates into a 12-month change in employment of -1.5 million, with a range of -2.1 to -0.9 million.

Hmmm. I wonder what his long-term model was saying in, say, June of 2000. I could imagine a VAR model at that time forecasting job growth of 3 million jobs a year, just because that was what all the then-current data would have suggested.

For Discussion. How much of the positive discrepancy of 5 million jobs that built up in the 1990’s between the payroll survey and the household survey might be due to the double-counting effect?