The Haze Surrounding the Unemployment Data

Economic data are subject to error, and employment statistics are no exception. Each month the employment figures are reported and discussed in the media down to the last digit as though blessed with divine perfection. But they're not.

The U.S. Bureau of Labor Statistics (BLS) reports two monthly employment series: nonfarm payroll jobs collected from employers, and civilian employment collected from household interviews. Both are valuable in assessing the condition of labor markets and guiding policy decisions. But users would be better off, and gain better perspective, if they knew the limitations of the data, including the types of underlying errors. Some errors mainly affect the level of data and some mostly affect monthly or cyclical changes.

Let's start with the payroll employment numbers. They are incomplete as initially reported by employers because of an undercount in the net number of new businesses. The BLS, using a model of business births and deaths, temporarily fills the information gap by estimating and adding in the number of jobs missed. In subsequent months the data are revised as the business count of workers becomes more complete. And once a year there is a basic and often substantial benchmark revision, a comprehensive count based on unemployment insurance tax records.

As recently reported by the BLS, the benchmark correction amounted to an upward revision of over 200,000 jobs a month in the past half year and somewhat less going back to January.

The job count is also overstated when workers change jobs and appear on the payrolls of both their new and their old company in the same pay period.

Also, the person responding for a business can misunderstand instructions or definitions or err and provide the wrong job count. Besides collection and processing errors, seasonal adjustment factors are subject to error and later revision. The payroll data are also affected by sampling error, which is estimated and reported.

The household employment data also have their problems.

The sampling error for a monthly change in civilian employment is 430,000, more than four times that of payroll employment. Consequently, the last time a monthly change in civilian employment was statistically significant (not counting the beginning-of-year months when the data were distorted by the introduction of new population controls), was December 2009.

Each year the Census Bureau updates its population estimates, which the BLS uses as controls in its labor market measures. The population data include an estimate for net international migration, which is more a guesstimate subject to considerable error.

In its latest press release on the employment situation, the BLS reported the effect of the new population controls on December 2011 data. The population count was raised by 1.5 million, the labor force by 258,000, employment by 216,000 and unemployment by 42,000. The labor force participation rate and the employment-population rate were both reduced by 0.3.

Response error is a persistent problem in the household survey. For example, when the Census interviewer needs information about a householder who is not at home, a proxy respondent can answer for that person. The proxy respondent may not know that the absent person worked for pay sometime during the reference week - it could have been a temporary or a part-time job - in which case the information provided will be wrong and employment will be undercounted.

When the absent householder is unemployed but the proxy respondent did not know the absentee looked for a job during the past four weeks and answers wrongly, the official unemployment count becomes understated. Studies based on more probing periodic reinterview findings show that the weaker the labor market, the greater the understatement in unemployment. In recessions the unemployment rate is estimated to be understated by a half point or more.

Revisions to economic statistics improve their accuracy. Revisions are also a confession of error. Raw data are sometimes processed, adjusted, imputed, and modeled to the point of dominating the underlying signal.

Knowing about data shortcomings is critical to the interpretation of economic statistics. Unfortunately, some errors can't be measured (e.g., uncounted workers employed off-the-books). They have technical names seldom mentioned, yet cast a dark shadow. Worse, errors that are known are too often ignored.