Hours Paid Drive 44 Percent of Industrial Production, Labelled Hours Worked

The Federal Reserve’s industrial production release carries a small table that readers of the headline index rarely open. Its title is “Availability of Monthly IP Data in Publication Window”, and its stub column is headed “Type of Data”. One row is labelled Production-worker hours. Across the six columns of that row the value is 44, then 44, then 44, then 44, then 44, then 44. The row above it, Physical product, climbs from 33 to 41 to 48 to 51 to 54 and stops at 54.

Before reading those values, read the unit. The note under the table says “Percent of value added in 2024; the numbers may not sum because of rounding”. This is not a count of series or of industries. It is the share of the index’s weight. Forty-four percent of 2024 value added is estimated from hours in the first monthly estimate, and the same 44 percent is still estimated from hours in the sixth. For nearly half the index, by weight, hours are not a stopgap that better data later displaces.

That makes one question worth following to the end: whose hours? The Federal Reserve Board, G.17 Explanatory Notes on Industrial Production and Capacity Utilization call them hours worked. The survey those notes name measures hours paid. The annual survey those notes name as the benchmark source carries its own hours line, and that line throws three paid-leave categories out by name. Three documents, one payroll series, two definitions.

Read the Column Header Before the Value

The spanner over the six value columns reads “Month of estimate”, with sub-headers running first through sixth. The nearby text gives the clock: the first estimate of output for a month is published around the fifteenth of the following month, and it is revised in each of the subsequent five months. Each column is a vintage of the same month, not a different month.

Two more rows sit below. IP data received rises from 78 to 86 to 92 to 95 to 98 and 98. IP data estimated falls from 22 to 14 to 8 to 5 to 2 and 2. Add the two source rows in the first vintage and the arithmetic lands near the received row: 33 plus 44 is 77 against 78 received, a one-point gap the table’s own rounding note covers.

One Survey, Two Words for Its Own Hours

The notes are explicit about when hours get used and where they come from. On the first point: “When suitable data on physical product are not available, estimates of output are based on production-worker hours by industry”. On the second: “Data on hours worked by production workers are collected in the monthly establishment survey conducted by the Bureau of Labor Statistics”.

That monthly establishment survey is the Current Employment Statistics program, and the first line of its hours definition settles the matter. The BLS Handbook of Methods, Current Employment Statistics — National: Concepts state: “Aggregate hours are the total hours for which employees are paid”. The next sentence anticipates the confusion and names it: “The concept of aggregate hours differs from those of scheduled hours and hours worked”.

The definition then itemises what sits inside the total. “For the reference pay period, aggregate hours include all hours worked (including overtime hours), ... equivalent hours for which employees received pay directly from the employer for sick leave”, and the list continues through holidays, vacations and other leave. The ellipsis covers a third inclusion, “hours paid for standby or reporting time”, also paid and not worked. The derived measure inherits the same base: “Average weekly hours (AWH) are the aggregate weekly hours divided by the number of employees paid for those hours”. Paid, not worked, in numerator and denominator alike.

So the G.17 notes attach the phrase hours worked to a series whose own documentation says the concept differs from hours worked, and whose itemisation puts paid sick leave, paid holidays and paid vacation inside the total.

What the Annual Benchmark Counts

The monthly indicator is not the end of the chain. The notes describe a conversion step: “The factors used to convert inputs into estimates of production are based on historical relationships between the inputs and the comprehensive annual data used to benchmark the IP indexes; these factors also may be influenced by technological or cyclical developments”.

Then the anchor: “The annual data used in benchmarking the individual IP indexes are constructed from a variety of source data, such as the quinquennial Censuses of Manufactures and Mineral Industries and the Annual Survey of Manufactures, prepared by the Bureau of the Census”. Which sends the trail to a Census questionnaire item.

That line is Item 7D, “annual number of hours worked by the production workers”. The U.S. Census Bureau, Economic Census Manufacturing FAQ, Item 7D answers what belongs in it: “All hours worked or paid for, except hours paid for vacations, holidays or sick leave. If an employee elects to work during their vacation period, report only actual hours worked by the employee”. A note in the same answer adds that overtime is to be reported as actual hours worked and not as straight time equivalent hours.

Set the instructions side by side. The monthly series counts paid vacation, paid holidays and paid sick leave inside the hours total; the annual questionnaire removes those three by name. Two instruments measure two different quantities, and both label the quantity hours worked. The notes fit the conversion factors to the comprehensive annual data used to benchmark the indexes, and Item 7D is one hours question on a survey that supplies that data, which is narrower than being the benchmark for hours.

The One Program That Builds the Bridge

One federal program does the arithmetic the gap requires and publishes the steps. The BLS Productivity and Costs, Technical Notes name the input, then undo it in two moves. The input: “The primary source of hours and employment data is the BLS Current Employment Statistics (CES) program, which provides monthly survey data on the number of jobs held by and hours paid to wage and salary workers in nonfarm establishments, counting a person who is employed by two or more establishments at each place of employment”.

The first move removes paid-and-not-worked time: “Hours of paid time off are removed from hours paid using data from the National Compensation Survey (NCS) for 1996 forward and data from the BLS Hours at Work survey, conducted for this purpose, prior to 1990”. The seam between instruments is documented rather than glossed: “Between 1990 and 1995, hours of paid time off are removed using a combination of NCS and Hours at Work survey data”.

The second move adds worked-and-not-paid time: “Off-the-clock hours are added, yielding hours worked, using data from the Current Population Survey (CPS)”. The coverage is wider, too: “Hours worked data for the labor productivity and cost measures include hours worked for all persons working in the sector—wage and salary workers, the self-employed and unpaid family workers”.

The reasoning behind the subtraction sits in the methodology rather than in inference. The BLS Handbook of Methods, Major Sector Productivity: Concepts define the target: “This includes all hours spent on actual production of goods and services, as well as time that is incidental to production, such as paid time for traveling between job sites, coffee breaks, and machine downtime”. Then the exclusion: “Paid vacations and other forms of paid leave are best viewed as benefits rather than as time available for production and should not be included for productivity analysis in measured hours”.

When the productivity program prints hours worked, two adjustments stand between that number and the payroll survey. When the industrial production notes print the same words, they point at the payroll survey itself.

Where the Wedge Moves

How much is at stake depends on how fast the wedge moves, and the productivity release is where the adjusted series appears with a growth rate attached. The BLS Productivity and Costs: Second Quarter 2026, Revised reported: “Nonfarm business sector labor productivity increased 1.4 percent in the second quarter of 2026, the U.S. Bureau of Labor Statistics reported today, as output increased 1.7 percent and hours worked increased 0.3 percent”. Over a longer window: “From the same quarter a year ago, nonfarm business sector labor productivity increased 2.2 percent in the second quarter of 2026”.

That 0.3 percent is an hours-worked number, which is to say an hours-paid number with paid time off subtracted and off-the-clock time added. The industrial production index, for 44 percent of its 2024 value added, moves with the unadjusted version of the same payroll hours. A stable level difference is harmless: the conversion factors absorb it. A change in the wedge is not, because the fitted factor carries last year’s relationship into this month. What the notes date is the seasonal adjustment, not that fit: for series based on production-worker hours the current seasonal factors were estimated with data through August 2025, and individual series are seasonally adjusted with Census X-13 ARIMA.

Three windows move the wedge: one where paid holidays fall differently against the reference pay period, shifting paid hours without shifting worked hours; one where firms cut scheduled hours but keep paid leave intact, shifting worked hours by more than paid hours; and a year where paid-leave incidence changes, shifting the ratio the productivity program applies while leaving the industrial production input untouched until the next benchmark. The same mechanism appears whenever a monthly indicator is pinned to an annual anchor, as it did when benchmark revisions land on one month and when capacity utilization is a ratio whose denominator moves.

Concrete Framework

Five steps, in order, before treating a published hours figure as a labour-input measure.

One. Name the survey, not the adjective. A document saying hours worked may point at an hours-paid series. Follow the citation to the collecting program and read that program’s own definition of its hours total.

Two. Ask where paid leave sits. Paid vacation, paid holidays and paid sick leave are the named part of the wedge, not all of it: standby and reporting time are also paid and not worked. Current Employment Statistics puts them in; Census Item 7D takes those three out; productivity hours remove paid time off as a class with National Compensation Survey ratios.

Three. Ask whether off-the-clock time was added. Removing paid leave is half the bridge. A series that stops after the subtraction is not the productivity program’s series.

Four. Read the unit on the share. The 44 in that table is a share of value added in a base year, not a share of series.

Five. Separate level from change. A fitted conversion factor absorbs a stable definitional gap, not a moving one. Ask what would have to move for the gap to widen this quarter, then check whether it moved.

What to Watch Next Week

Three dated items turn this from a definitional note into something testable.

First, the next industrial production release and its revision table. If the manufacturing index revises toward payroll hours while the physical-product share for that month rises, the hours input did the work in the first vintage.

Second, the establishment survey’s aggregate hours and average weekly hours for the same reference month. Those are hours-paid measures by their own definition, and they are published before the industrial production estimate that leans on them.

Third, the next productivity and costs release, the only one of these printing an hours series with paid time off removed and off-the-clock time added. Comparing its hours growth with aggregate hours growth for the same quarter measures the wedge in the agencies’ own published numbers.

What Would Invalidate This

The industrial production notes answer the conversion question rather than leaving it open, and the answer weakens the sharpest version of this piece. The factors turning hours into output are fitted to the annual benchmark and may be influenced by technological or cyclical developments. A fitted factor absorbs a definitional gap that holds still. If the paid-leave wedge is close to constant at monthly frequency, the mislabel costs the index nothing.

The holiday channel is addressed in those same notes rather than ignored. The notes state that series are pre-adjusted for the effects of holidays or the business cycle when appropriate, with the seasonal factors for hours-based series dated through August 2025. Holiday timing is therefore handled where the notes judge pre-adjustment appropriate.

Four things stayed unverified. The closing list of the establishment survey’s equivalent-hours sentence was confirmed twice to run through holidays, vacations and other leave, but its exact wording was not captured verbatim, so it is described rather than quoted. The year-over-year hours and output cells in Table A1 of the productivity release were read once, through a summarizer, and not returned as verbatim text, so they are left out of the argument. Whether every hours-based industrial production series draws on the establishment survey rather than an industry-specific hours source was not confirmed series by series. And the Census hours instruction quoted here is the Economic Census manufacturing answer for Item 7D, the current annual survey questionnaire being unreachable.

What would settle it is narrow and public. Take a month in which paid holidays fell differently against the reference pay period, compare aggregate hours growth with the productivity program’s hours-worked growth, then watch whether the hours-based part of the index revised toward the worked measure once physical product arrived. If it did not, the fitted factor absorbed the wedge. If it did, 44 percent of 2024 value added moved on a word.