One report, two units of measurement
The establishment survey, known as Current Employment Statistics or CES, counts jobs on nonfarm payrolls. The household Current Population Survey, or CPS, measures people’s employment and labor-force status. A person with two payroll jobs is counted twice in the establishment survey and once as employed in the household survey. Self-employment and some other work included in the household survey are outside CES coverage. The unemployment rate comes from the household survey, not from subtracting payroll jobs from a population total. [1]
An increase in jobs alongside an increase in unemployment is therefore possible. More people can enter the labor force looking for work even as employers add positions. The two surveys also use different reference periods: generally the week containing the 12th for households and the pay period containing the 12th for employers. [1] Neither headline is a complete description of every hiring or layoff event during the calendar month.
How a sample becomes an estimate
Census administers the household survey using a probability-selected sample of about 60,000 occupied households. The normal rotation keeps households in for four months, out for eight, and back for four. That provides continuity while refreshing the sample; it does not make the survey a census of every person. Census documentation also distinguishes eligibility to participate from the BLS labor-force statistics generally published for people aged 16 and older. [2]
CES applies weighted changes from establishments reporting in adjacent months to the previous employment estimate. A model-based component addresses net employment from business openings and closures missed by the sample. New firms cannot appear instantly in a survey frame, and a closed firm may initially resemble a late respondent. The birth–death component addresses this measurement problem rather than directly counting all new companies that month. [3]
The analytical limitation is that a precisely printed number can remain uncertain. Sampling error concerns which respondents happened to be selected; nonsampling error includes missed responses, reporting mistakes and imperfect coverage. A confidence interval for sampling error alone cannot describe every way an estimate might later change. [1] Greater industry detail can reveal concentration while also making sampling noise more important.
Routine revisions and benchmarks are different operations
BLS routinely revises the previous two CES months as additional employer reports arrive and seasonal factors are recalculated. Seasonal adjustment attempts to distinguish recurring calendar patterns from other changes; it is not a deletion of actual employment. [1] A revision in this process usually reflects a different information set for the same reference period.
Annual benchmarking instead re-anchors estimated employment levels to more comprehensive counts, primarily unemployment-insurance tax records in the Quarterly Census of Employment and Wages. Those administrative data arrive with a lag and have their own errors. Benchmarking is a comparison of measurement systems, not an audit establishing a perfectly error-free total. [3][4]
BLS cautions that the unadjusted birth–death contribution should be compared with unadjusted employment changes. Subtracting it directly from the seasonally adjusted headline mixes measurement bases and can exaggerate the model’s apparent role. [3] A critique of an estimate can be legitimate while the arithmetic used to illustrate that critique is not.
Worked example: a downward level revision can raise a monthly gain
Consider an invented employment series, shown in thousands of jobs. The first published levels are 100,000 in the base month, 100,120 in month one and 100,220 in month two. The monthly gains are 120,000 and 100,000 jobs. Later information lowers month one to 100,080 and month two to 100,200, with the base unchanged.
The revised gains are 80,000 and 120,000. Month two’s employment level was revised down by 20,000, yet its monthly gain was revised up by 20,000 because month one was revised down even more. Across the full two-month interval, the total gain falls from 220,000 to 200,000. A story describing only the latest level revision, only the latest monthly gain, or only the combined revision is answering a different question.
This is pure arithmetic, not an estimate of U.S. employment. It shows why revisions to levels and revisions to changes cannot be used interchangeably. The starting matters as much as the ending vintage when constructing a comparison.
Scroll horizontally to see all columns.
| Invented period | First level, thousands | Revised level, thousands | Revised monthly change, thousands |
|---|---|---|---|
| Base | 100,000 | 100,000 | Not applicable |
| Month one | 100,120 | 100,080 | 80 |
| Month two | 100,220 | 100,200 | 120 |
A dated example of preliminary versus incorporated data
On August 28, 2026, BLS released a preliminary March 2026 benchmark revision of −79,000, or −0.1%, for total nonfarm employment and −178,000 for private employment. The release explicitly states that these preliminary estimates do not update the official establishment series. It schedules incorporation of the final benchmark with the January 2027 Employment Situation release in February 2027. These are the status and figures in that specific release, checked October 4, 2026. [4]
The March level comparison cannot be described as 79,000 jobs lost in August. Nor is dividing the number by twelve a verified reconstruction of every month’s future revision. The preliminary comparison pertains to a benchmark level; the finalized monthly path also depends on the benchmarking procedure, updated information and seasonal factors. [3][4]
For a separate hypothetical illustration, a March level of 100 million revised to 99.6 million is a 400,000 downward level correction. If the prior March level stayed unchanged and the original twelve-month increase was 1.2 million, the revised increase would be 800,000. That conditional calculation does not establish how the 400,000 difference was distributed among the intervening months.
Why payroll growth and unemployment can diverge
In another hypothetical population, 950 people are employed and 50 are unemployed, giving a labor force of 1,000 and a 5% unemployment rate. If ten previously inactive people begin looking for work and no other status changes, employment remains 950, unemployment becomes 60 and the labor force becomes 1,010. The unemployment rate rises to about 5.94% without any employed person losing a job.
Likewise, one employed person taking a second payroll position increases the payroll-job count without increasing the number of employed people. These examples isolate definitions. In actual data, participation, population estimates, hours, multiple jobholding and sampling variation can all matter, so a divergence alone cannot identify one cause.
What revisions do and do not establish
Repeated downward revisions can be evidence that an earlier picture was too strong. They warrant explanation of response patterns, business formation, seasonal effects and later administrative counts. A single revision does not by itself establish political interference or prove that all monthly information is useless. Conversely, the existence of a routine revision process does not make a large error economically unimportant.
The useful distinction is between the labor market and knowledge about it. A contemporaneous estimate may help describe conditions before a complete count exists; later evidence can change that description. Comparisons across the same data , broader movements in employment and hours, and the eventual benchmark can clarify whether an apparent turning point was persistent or statistical noise. No one release resolves every dimension of labor demand.
Sources
- BLS; Employment Situation Technical Note; September 2026 reference-period version checked October 4, 2026Official releaseBack to text: ↑1↑2↑3↑4
- U.S. Census Bureau; Current Population Survey Methodology; June 4, 2024Official sourceBack to text: ↑
- BLS; Current Employment Statistics–National Handbook of Methods, Calculation; current methodology checked October 4, 2026Official sourceBack to text: ↑1↑2↑3↑4
- BLS; Current Employment Statistics Preliminary Benchmark (National), March 2026; August 28, 2026Official releaseBack to text: ↑1↑2↑3