How BLS employment data is built, from the household survey to the payroll count

What to take away

• BLS employment data comes from two independent surveys: the household Current Population Survey and the establishment Current Employment Statistics program.
• The household survey counts people and produces the unemployment rate; the payroll survey counts jobs and produces the monthly change in nonfarm payrolls.
• Both are seasonally adjusted before publication, so the headline figure differs from the raw count.
• Payroll figures are re-based each year to near complete tax records through benchmark revisions.
• The two surveys can move in opposite directions in the same month because they measure different things.
• One month of data is a noisy signal, and the Employment Situation release carries more context than the headline number.

The two surveys behind one monthly jobs headline

The Employment Situation release lands on the first Friday of most months and carries two job numbers. One is the unemployment rate, drawn from a survey of households. The other is the change in nonfarm payrolls, drawn from a survey of employers. They answer different questions and come from different samples.

The household survey is the Current Population Survey, run for the Bureau of Labor Statistics by the U.S. Census Bureau. It asks people whether they worked, looked for work, or neither. It produces the unemployment rate and the labor force participation rate.

The payroll survey is the Current Employment Statistics program. It collects job counts from employer records, mostly through a mandatory survey of businesses. It produces the payroll number that most headlines lead with.

Both programs sit inside the Bureau of Labor Statistics, an agency of the U.S. Department of Labor. Neither is a single national headcount. Each is a sample, weighted and adjusted to represent the country.

That structure explains most of what confuses readers about the monthly report. Two samples, two definitions, two sets of revisions. When they disagree, the disagreement is usually arithmetic rather than error.

Reporters who treat the two numbers as competing versions of one fact will misread the release. They are better understood as two instruments pointed at the same labor market from different angles. To compare indicators side by side, learn to build a news timeline.

Inside the Current Population Survey: households and the unemployment rate

The Current Population Survey samples about 60,000 households each month, drawn to represent the civilian noninstitutional population. It is the source of the official unemployment rate and much of what is known about who is working and who is not.

Interviewers ask about the week containing the 12th of the month, called the reference week. A person counts as employed if they did any paid work in that week, even one hour. A person counts as unemployed if they had no work, were available, and actively looked in the prior four weeks.

Everyone else is outside the labor force: retirees, full time students, people caring for family, and discouraged workers who stopped looking. The unemployment rate is unemployed divided by the labor force, not divided by the whole population.

That denominator matters. The rate can fall because people found jobs or because people gave up looking. Both show up as a lower number, but only one is good news. The survey also publishes the U-6 measure, which adds marginally attached workers and part time workers who want full time hours.

The household survey counts people, not jobs. A person holding two jobs counts once. A person who lost one job and found another counts as employed in both months. The CPS home page carries the definitions, sample design, and the monthly tables behind the rate.

Because the sample is smaller than the payroll sample, the household estimates carry wider margins of error. Month to month moves of a tenth or two in the unemployment rate are often inside the noise band. The full methodology for both surveys sits in the BLS Handbook of Methods.

Inside Current Employment Statistics: payroll counts from establishment reports

The Current Employment Statistics program surveys roughly 119,000 businesses and government agencies covering about 629,000 worksites. It asks a narrower question than the household survey: how many people were on the payroll in the pay period that includes the 12th.

The unit is the job, not the person. Someone working two payroll jobs is counted twice. That is why payroll employment can exceed household employment even though both describe the same labor market.

The survey is a probability sample stratified by state, industry, and size. Most responses arrive electronically, and BLS collects them through a secure website and a touchtone phone system. Reporting is mandatory for sampled firms, which supports high response rates.

BLS publishes the payroll number alongside average hourly earnings, average weekly hours, and the diffusion index showing how many industries added jobs. Those details often explain a headline that looks strange on its own. A solid payroll gain with falling hours tells a different story than the same gain with rising hours.

State and metropolitan estimates come from the same program, published a few weeks after the national figures. The Current Employment Statistics program publishes the national series, the sample design, and the definitions behind each measure.

The payroll count is a count of jobs, so it says nothing directly about unemployment. A month can add payroll jobs while the unemployment rate rises, if more people enter the labor force looking for work. That combination is common at turning points and routinely misread.

Seasonal adjustment and why the raw numbers look different

Hiring follows the calendar. Retailers staff up before the winter holidays and cut in January. Construction ramps in spring and slows in winter. School districts hire in September and lay off in June. These patterns repeat closely enough to be measured.

Seasonal adjustment removes the expected seasonal component so that consecutive months can be compared. BLS estimates seasonal factors from past years and applies them to the current data. The result is the seasonally adjusted figure that appears in headlines.

The unadjusted number is published too, and it can look alarming next to the adjusted one. A large unadjusted payroll gain in November is normal, and the adjusted figure strips out most of it. A large unadjusted loss in January is also normal.

Seasonal factors are revised each year as new data accumulate. When a factor changes, previously published adjusted figures change with it, even though the underlying unadjusted data did not move. That is one reason a number can shift without any new information about the economy.

A seasonal data desk case walks through how a headline can be revised for this reason alone.

Month	Typical unadjusted pattern	What seasonal adjustment does
January	Large job losses after holiday hiring	Removes the expected drop, often turning it positive
April	Construction and leisure hiring picks up	Trims the spring gain to isolate the underlying trend
July	Auto plant retooling and school layoffs	Offsets the recurring summer dip
September	School hiring lifts government payrolls	Reduces the education driven jump
November	Retail and transport hiring for holidays	Removes most of the seasonal gain

Adjustment is not manipulation. It is a standard technique used across economic statistics, and BLS publishes both versions so readers can check the work. The choice of which to use depends on the question. Year over year comparisons of unadjusted data are often the cleanest way to sidestep seasonal noise entirely.

Benchmark revisions, errata, and the annual re-basing of payroll data

Every year, BLS re-anchors the payroll series to near complete employment counts from state unemployment insurance tax records. That process is the benchmark revision, and it corrects the cumulative drift between the sample based estimates and the administrative universe.

The preliminary benchmark estimate is published with the January Employment Situation release, and the final revision appears the following February. The level of payroll employment can move by a few hundred thousand jobs in either direction. The revision is not a correction of an error, it is a recalibration to better data.

Monthly revisions are smaller and more frequent. Each month, BLS revises the prior two months as late survey responses arrive. The first print of a payroll number is an early estimate, and the second and third prints are usually closer to the final figure.

A guide to court reporting explains why separating the record from the narrative matters when those monthly changes happen.

Errata are a separate category. When BLS finds a processing error or a coding mistake, it documents the correction and republishes the affected series. The errata page lists documented corrections to published data series, with dates and affected tables.

Errata are rare relative to the volume published, but they matter to anyone who cited a figure that later changed. The BLS frequently asked questions explain how the agency collects, produces, and revises employment data, including what triggers a correction.

For reporters, the practical rule is to date every figure. A payroll number is a first print, a second print, or a benchmarked level, and each has a different status. Saying which one you are using prevents most corrections after publication. Before filing, run the figure through basic statistical claim checks.

Why the household and payroll numbers diverge in the same month

Divergence is normal, not a sign that one survey failed. The two programs measure different populations, use different definitions, and carry different sampling error. In any given month they can point in opposite directions.

The clearest source of divergence is the unit of measurement. Payrolls count jobs, including multiple jobs held by one person. The household survey counts employed people once, no matter how many jobs they hold. Growth in multiple job holding widens the gap.

Another source is coverage. The payroll survey excludes agricultural workers, private household employees, unpaid family workers, and the self employed. The household survey includes them. In a month when self employment rises, household employment can grow while payrolls stall.

Age and population controls also matter. The household survey is weighted to independent population estimates that are updated annually. When those controls change, the level of household employment shifts without any change in the underlying responses.

Sampling error cuts both ways. The household sample is smaller, so its month to month changes are noisier. A payroll gain of 150,000 with a household decline of 100,000 is not a contradiction. It is two estimates with overlapping confidence intervals.

A worked example makes the point. Suppose a state adds 20,000 payroll jobs in a month, mostly in warehousing, while 15,000 residents leave the labor force after retirements. Payrolls rise, the labor force shrinks, and the unemployment rate can fall even though household employment barely moved. All three statements are consistent.

Over longer periods the two surveys track each other closely, because they are measuring the same economy. The divergence is a monthly phenomenon, and it tends to shrink when averaged over a year. Reporters who compare twelve month changes rather than single months avoid most false alarms.

Reading the Employment Situation release without overreading one month

The release contains two news releases in one document, plus dozens of tables. The headline payroll change and the unemployment rate get the attention, but the establishment survey also carries earnings, hours, and industry detail that often explain the headline.

Start with the payroll change and its confidence interval, not the point estimate. Then check the prior two months for revisions, which can change the story from a slowdown to a steady pace. Then look at the household survey for the unemployment rate, participation, and the U-6 measure.

Next, check whether the move is broad or narrow. A gain concentrated in one industry, such as health care or government, means something different from a gain spread across many. The diffusion index and the industry table answer that question quickly.

Finally, compare the month to the trend. Three month and twelve month averages smooth the noise and are published in the release. A single month that breaks a trend is worth reporting, but it should be labeled as one month.

BLS publishes the release on a fixed schedule, and the underlying data are available through the agency's public databases. Anyone can reproduce the headline figure from the published tables, which is the strongest check available. Knowing how official record types differ before writing the headline is the difference between reporting a number and reporting what happened.

Common questions

Why do the household and payroll surveys give different job numbers? They measure different things. The household survey counts employed people once, including the self employed and agricultural workers. The payroll survey counts jobs at sampled establishments, so one person with two jobs counts twice.

What is seasonal adjustment in the jobs report? It removes recurring calendar patterns, such as holiday retail hiring and summer school layoffs, so consecutive months can be compared. BLS publishes both adjusted and unadjusted figures, and the seasonal factors are revised each year.

How large are benchmark revisions to payroll data? They can move the level of payroll employment by a few hundred thousand jobs in either direction. The preliminary estimate comes with the January release and the final revision the following February.

Does a falling unemployment rate always mean more people are working? No. The rate can fall because unemployed people found jobs or because they stopped looking and left the labor force. Participation and the U-6 measure help separate the two cases.

How much should one month of jobs data change a story? Less than the headline suggests. Sampling error, revisions, and seasonal factors all affect a single month. Three month and twelve month averages are more reliable for describing a trend.