US non-farm payrolls and the employment report

One report, two independent surveys 

The employment situation report is built from two surveys that share a reference period and share almost nothing else. The establishment survey asks employers how many people were on their payrolls, and it produces the payroll count, the average workweek and average hourly earnings. The household survey asks people about their own labour market status, and it produces the unemployment rate, the participation rate and the employment to population ratio. They are conducted by different means, they count different populations, and they are reconciled by nobody.

The differences are definitional rather than accidental. A person holding two jobs is counted twice by the establishment survey and once by the household survey. The self-employed, unpaid family workers, agricultural workers, private household employees and the armed forces are outside the payroll count entirely but inside the household survey's employment measure. A person who worked for pay for even an hour in the reference week is employed on the household definition. Each survey is internally coherent; neither is the other's cross-check.

The sample sizes differ by more than an order of magnitude. The establishment survey covers a very large panel of business and government establishments; the household survey covers a comparatively small sample of households. The consequence is that the payroll count is statistically the more precise of the two on a month to month basis, while the unemployment rate carries a wider sampling error than its one decimal place presentation suggests.

What the payroll count includes 

Non-farm payroll employment counts jobs, not people, on the payrolls of non-agricultural establishments during the pay period that includes the twelfth of the month. Government employment at every level is included. Anyone who received pay for any part of that pay period is counted, so a person on paid leave is in the count and a person on unpaid strike is not.

The count is published as a change from the previous month rather than as a level, which is the form in which it is universally quoted. Because the level is very large and the monthly change is small relative to it, the reported change is a small difference between two large estimates, and that is the structural reason the series is revised as much as it is.

Revisions, and the birth-death model 

Each month's payroll figure is revised twice in the following two reports as late responses arrive, and is then re-benchmarked annually against unemployment insurance tax records covering nearly the entire universe of employers. The annual benchmark can move a year of monthly figures materially, and it is published with a preliminary estimate several months before it is incorporated.

The survey also cannot observe businesses that opened or closed too recently to be in its sampling frame, so the Bureau estimates their net contribution with a statistical model, commonly called the birth-death model. The model is estimated from historical patterns, which means it performs least well precisely at turning points, when the rate of business formation and failure departs from its history. This is a known limitation, published by the Bureau itself, and it is the usual explanation offered when a benchmark revision is large.

The lines that are not the headline 

• The unemployment rate. The share of the labour force that is without work, available for work and actively searching. Its denominator is the labour force, not the population, so it falls when unemployed people stop searching and rises when discouraged people resume searching, in both cases without any change in the number of people holding jobs.
• The participation rate. The share of the working age population either employed or actively searching. It moves with demographics, education patterns and caring responsibilities as well as with the cycle, and it is the denominator that makes the unemployment rate interpretable.
• Average hourly earnings. Average pay per hour across the payroll panel. It is affected by composition: if job losses fall disproportionately on lower paid roles, the average rises without any individual receiving a raise.
• Average weekly hours. Hours per worker, which employers typically adjust before headcount, making it a component practitioners watch for changes in labour demand that have not yet reached the payroll count.
• The broader underutilisation measures. A published family of alternative rates that progressively include discouraged workers, others marginally attached to the labour force, and people working part time who want full time work.

Cadence and the surrounding calendar 

The report is published monthly, conventionally on the first Friday of the month, in the morning Eastern Time, covering the preceding month. The schedule is published a year in advance and the first Friday convention is a strong tendency rather than a rule, because the release date depends on the position of the reference week in the calendar.

Several other labour releases sit around it and are frequently used as partial previews. Weekly initial claims for unemployment insurance are an administrative count rather than a survey, published every week, and cover a different concept: filings for benefit, not the level of employment. The job openings and labour turnover survey reports vacancies, hires, quits and layoffs, and is published with a longer lag. A private payroll processor publishes its own employment estimate days before the official report; it is a different sample with a different methodology, and the correspondence between it and the official count in any given month is not reliable.

How the report is conventionally read 

The most common convention reads the headline change against consensus, then immediately adjusts it for the revisions to prior months, then examines earnings for wage pressure and participation for the composition of any change in the unemployment rate. A second convention compares the payroll change against an estimate of the pace required to hold the unemployment rate steady given population growth, a threshold that is itself estimated rather than published and that different analysts calculate differently.