Is There Actual Bias In the BLS's Unemployment Reports?

On Friday the U.S. Bureau of Labor Statistics (BLS) reported that nonfarm payroll jobs, based on data collected by the employer survey, rose by a modest 114,000 in September. The BLS also reported that total employment, based on data collected by the household survey, rose by 873,000 in September - though involuntary part-time employment accounted for two-thirds of the increase. This was the biggest monthly gain in total employment in 28 years (not counting the beginning of each year when the introduction of new population controls into the data series distorts the monthly change). With such a huge leap in reported total employment, the unemployment rate, also based on data from the household survey, fell from 8.1 percent in August to 7.8 percent last month.

Traditionally, in its monthly press release on the employment situation, the BLS features the change in payroll jobs in its write-up. And with good reason. The payroll survey covers 440,000 establishments while the household survey sample size is a much smaller 60,000 households. Consequently, the statistical error range in the payroll survey is much smaller than in the household survey, as reflected in the latter's greater month-to-month volatility. Seasonal adjustment, response bias, and various measurement errors can also be distorting factors.

It would hardly be surprising if household-measured employment and the unemployment rate returned to their more normal pre-September levels in next month's BLS report, which will be released four days before the presidential election.

A look at recent trends in payroll and household employment can help evaluate the September data. From January to August of this year, payroll jobs rose by an average 150,000 a month, with all months showing an increase, reflecting the relative stability of the data resulting from the payroll survey's larger sample size. By comparison, employment from the household survey in the same period averaged 164,000 a month, with monthly changes ranging from minus 195,000 to plus 422,000 (excluding beginning of the year changes affected by the introduction of new population controls.) Both the size and instability of monthly changes in the household employment series raise serious questions about the reliability of any one-month change in the data, and in particular the September jump.

If data in subsequent months show that the September leap in household-measured employment was an aberration and the series returns to its earlier trend, then the unemployment rate can also be expected to return to its recent 8-plus percent trend. If household employment had increased in September by its average monthly change in January to August, last month's unemployment rate would have been unchanged at 8.1 percent.

When last month's 7.8 percent unemployment rate was reported, it strained credulity for many data users and even raised questions about political tampering with the numbers. But those who are familiar with Census Bureau and BLS methodologies in collecting and constructing the data, including yours truly, know better. The labor market data are objectively and honestly collected and calculated.

But there is a subtler question. Is the way in which the data are reported wholly objective and neutral? Are written descriptions unbiased? Are there periodic omissions or inconsistencies in reporting? Here one can find fault.

In its September press release on the employment situation, the BLS says "Total employment rose by 873,000 in September..." And that indeed is what the data show. But if one were to look back over prior press releases in months where total household-measured employment rose by enough to pass the BLS test of statistical significance, you would be hard put to find any month that uses language mentioning the rise in the level of total employment. For example, in April 2010 measured employment rose by a sizeable 550,000, yet there was no statement in the text of the press release mentioning that increase. Then again, September 2012 is not April 2010.

Nor were significant declines in total employment written about in past releases. In December 2009 household-measured employment fell as then reported by 589,000 at a time when payroll jobs dropped by only 85,000. In September of that year total employment declined by 785,000 compared to a 263,000 drop in payrolls. The text of the press releases mentioned only the smaller declines in payroll jobs.

I have argued in an earlier column that BLS should give equal treatment in their textual write-ups to statistically significant monthly changes in both total employment and payroll jobs, consistently and in the context of their trends. If the latest press release is the beginning of that practice, fine. But given the upcoming election, the timing is terrible.