Menu
Book a demo
Work Sampling: Meaning, Sample Size Formula and a Worked Example

Work Sampling: Meaning, Sample Size Formula and a Worked Example

How work sampling estimates where time goes from random observations, how many you need for a given accuracy, and when a machine signal does it better.
Work Sampling: Meaning, Sample Size Formula and a Worked Example

In short: Work sampling is a way to find out how people or machines spend their time without timing them all day. An observer records what is happening at many random moments, and the share of observations in each activity estimates the share of time spent on it. 36 "working" readings out of 40 means about 90% of the time was spent working.

  • It is also called activity sampling, ratio delay study or the snap-reading method.
  • The number of observations sets the precision: n = z² x p x (1 - p) / E². Halve the error and you need four times the readings.
  • For machines, a signal that logs run and stop time replaces sampling. Keep sampling for what people do.

Guide, updated 14 November 2026. Sources are listed at the end.

What is work sampling?

Work sampling is a work measurement technique based on probability. Instead of watching one person with a stopwatch, an observer makes short, instant observations at random times, often across many people or machines, and notes the activity at that moment: working, waiting, walking, setting up, idle. Course notes from NPTEL, India's national engineering course programme, describe it as getting facts about the use of machines or people through many instantaneous observations at random intervals. The ratio of observations of an activity to the total approximates the share of time spent on it.

The method comes from the textile industry. L. H. C. Tippett described a snap-reading method of making time studies of machines and operatives in the Journal of the Textile Institute in February 1935. The names have multiplied since: activity sampling, ratio delay study, observation ratio study. They all mean the same idea.

Work sampling meaning, in one example

An example in Eastern Mediterranean University's work study course (IENG 301) shows the whole idea. An observer records 40 observations of an operator: 36 working and 4 idle. Idle time is 4 / 40 = 10%, working time 90%. Over an 8-hour day of 480 minutes, that is about 48 minutes idle and 432 minutes working. Nobody held a stopwatch.

Forty observations are far too few to trust. The rest of this guide is about how many you need, how to take them and what the answer is worth in money.

Track OEE and downtime in real time
Book a demo

What work sampling is used for

The EMU course lists three uses:

  • Activity and delay sampling: the share of the day people or machines spend working, waiting or idle. This is the common use on a plant floor.
  • Performance sampling: working time plus a performance rating, which gives a performance index for a person or a team.
  • Work measurement: a time standard for an operation, from total time, the share spent working, a performance rating and the pieces produced.

It suits work that a stopwatch handles badly: long or varied cycles, indirect jobs like material handling, maintenance or quality checks, and teams spread over a hall. NPTEL lists its advantages over time study: several operators studied at once, no stopwatch, fewer observer hours, and results that span days or weeks, so one odd day does not skew them.

Observations needed, 95% confidencen = 1.96² x p x (1 - p) / error²Share p± 5 pts± 3 pts± 2 pts10%13938586520%2466831,53730%3238972,01750%3851,0682,401Halve the error and you need four timesthe observations.Sampling vs a machine signal~50observationsa day (NPTEL)Censusa machine signallogs it allPeoplestill needsamplingSample what people do;count what machines do.
Observations needed at 95% confidence, from the NIST/SEMATECH confidence interval for a proportion, with NPTEL's figure of about 50 observations a day. See the source.

How many observations? The sample-size formula

Each observation is a yes or no: was the person waiting at that moment or not? The share you count, p, is a sample proportion, so the standard statistics for a proportion apply. The NIST/SEMATECH e-Handbook of Statistical Methods gives the usual confidence interval: p plus or minus z x √(p(1 - p) / n). Set the half-width of that interval to the error you can accept, E, and solve for n:

n = z² x p x (1 - p) / E²

  • z is 1.96 for 95% confidence (2 is a common round figure) and 1.645 for 90%.
  • p is your first guess of the share. A short pilot study of a day or two gives it. If you have no idea, use 0.5, which gives a safe upper bound for n.
  • E is the absolute error in percentage points, as a decimal: 0.03 for plus or minus 3 points.

Absolute or relative accuracy?

Older work study texts state accuracy as a share of p, not in points. The EMU course writes it as S x p = 2 √(p(1 - p) / N), with S the relative accuracy, and notes that plus or minus 5% is often considered satisfactory. Rearranged: N = 4 (1 - p) / (S² x p). Its example: a pilot of 100 observations finds machines idle 25 of them, and the analyst wants plus or minus 5% relative accuracy. N = 4 x 0.75 / (0.0025 x 0.25) = 4,800 observations.

Plus or minus 5% of 25% is plus or minus 1.25 points, which is very tight. The same p with plus or minus 3 absolute points needs only 3.84 x 0.25 x 0.75 / 0.0009 = 801 observations. Before you start, write down which kind of accuracy you mean. Mixing them up can make a study six times bigger than it needs to be, or six times too small to trust.

The confidence table

Observations needed at 95% confidence (z = 1.96), absolute error, rounded up:

Expected share p± 5 points± 3 points± 2 points± 1 point
5% or 95%732034571,825
10% or 90%1393858653,458
20% or 80%2466831,5376,147
30% or 70%3238972,0178,068
50%3851,0682,4019,604

Two rules fall out of the table. Halving the error needs four times the observations. And the closer p is to 50%, the more observations you need, so the activity you care about decides the size of the study. At 90% confidence, multiply any figure by about 0.70 (1.645² / 1.96²).

How long does it take? NPTEL suggests about 50 observations a day as a good figure for one observer and one study. If one round of the hall records 6 people, 12 rounds a shift give 72 observations, so 720 take 10 shifts.

How to run a work sampling study, step by step

  1. Define the question and the categories. Write each category so two observers would record the same thing: "working" (adding value to the product), "waiting for material", "waiting for the machine or line", "walking or searching", "personal". Five to eight categories is plenty.
  2. Tell people why. Both course notes include this step: get the supervisor's approval and explain the purpose to the operators. NPTEL warns that operators may change their behaviour when observed. A study that is seen as a hunt for idle people will measure fear, not work.
  3. Run a short pilot of one or two shifts to estimate p for the category that matters.
  4. Size the study from the formula or the table, and spread it over enough shifts and days to cover normal variation: both shifts, Mondays and Fridays, changeover days.
  5. Set random times. Use a random number table or a spreadsheet's random function to set the round times, and vary the route too. Fixed times line up with breaks, deliveries and shift routines.
  6. Observe at the instant. Record what each person is doing at the moment you see them, not what they did a minute ago or are about to do.
  7. Check precision at the end. Recalculate E from the real p and n. If it is wider than you wanted, extend the study.
  8. Turn shares into hours and money, and report the confidence range with the result.

From work sampling to a time standard

For work measurement, EMU gives the time standard per piece as total time x the share of time working x the performance index, divided by the pieces produced, plus allowances. Using the inputs of NPTEL's example: a 1,500-minute study, 80% working, a 90% performance rating and 360 pieces gives 1,500 x 0.80 x 0.90 / 360 = 3.0 minutes of normal time per piece. With a 13% allowance added to normal time, the standard is 3.0 x 1.13 = 3.39 minutes. Allowance conventions vary between companies, so state yours.

Worked example: a packing hall

Example data. Every figure below is assumed, except the labour cost, which is from US government data as marked.

A packing hall runs two 8-hour shifts, 250 days a year, with 6 operators a shift. The CI lead suspects operators lose time waiting for components and cartons. A one-day pilot suggests waiting for material is around 15%. The study takes 12 random rounds per shift over 10 shifts, observing all 6 operators each round: 6 x 12 x 10 = 720 observations.

The results

ActivityObservationsShare
Working43961.0%
Waiting for material10114.0%
Waiting for the line (stopped or starved)8611.9%
Walking or searching588.1%
Personal, outside scheduled breaks365.0%
Total720100%
Where operator time wentExample data: 720 random observationsWorking61.0%Waiting for material14.0% ± 2.5Waiting for the line11.9%Walking or searching8.1%Personal, outside breaks5.0%Waiting for material: about 3,150 h a yearx $36.70 an hour = about $115,605 a year
Example data. Labour cost per hour worked from BLS Employer Costs for Employee Compensation, June 2026.

How precise is it?

For waiting for material, p = 101 / 720 = 14.0%. The 95% error is 1.96 x √(0.140 x 0.860 / 720) = 0.025, or plus or minus 2.5 points. The true share is likely between 11.5% and 16.5%. To get to plus or minus 2 points, the study would need 3.84 x 0.14 x 0.86 / 0.0004 = 1,157 observations, about six more shifts.

What it is worth

The hall pays 2 shifts x 6 operators x 7.5 working hours x 250 days = 22,500 operator hours a year (example data, half an hour of each shift is a scheduled break). The BLS Employer Costs for Employee Compensation, Table 4, gives $36.70 per hour worked in June 2026 for private industry production, transportation and material moving occupations, wages and benefits together.

  • Point estimate: 22,500 x 0.140 = 3,150 hours x $36.70 = $115,605 a year paid while waiting for material.
  • Low end: 22,500 x 0.115 = 2,587.5 hours x $36.70 = $94,961 a year.
  • High end: 22,500 x 0.165 = 3,712.5 hours x $36.70 = $136,249 a year.

Report the range, not just the middle. A plant manager can fund a fix on the low end with a clear conscience. If kitting components to the line halves the waiting, the saving is 1,575 hours, or $57,803 a year at the point estimate, whether it shows up as less overtime, more output from the same crew or one fewer temporary worker.

The hours are the same unit that units per labor hour divides by, so a study like this explains why that figure moves. Combined with performance and quality, the share of time working is also the availability part of overall labor effectiveness.

Book a demo and see downtime per line and shift.

What continuous counting changes

Look at the third row of the results: waiting for the line, 11.9%. That is not really a fact about people. It is a fact about the machine, seen through people. And for machines, sampling is the second-best tool.

A machine with a PLC signal or a sensor can log every start and stop. That is a census, not a sample: there is no sampling error, no observer and no need for random times. A machine signal also catches what random snapshots miss. A study by NIOSH researchers in the Journal of Applied Behavior Analysis simulated momentary time sampling, the same instant-observation idea, and found large relative errors when the interval between observations was longer than the events being measured. Error was small or negligible when the interval was about equal to or shorter than the event. A 40-second jam between rounds 30 minutes apart is exactly that case: you see some of them, you miss their number and their cause.

So split the job:

  • Machines: count. Run time, stop time, stop count and output come from the machine. OEE monitoring in OEE software turns them into availability, performance and quality per shift, with stop reasons added by operators.
  • People: sample. What operators do while the line runs, and while it is stopped, still needs eyes. Work sampling is the cheap, honest way to get it.
  • Join them. With machine stops logged, the observer can drop "waiting for the line" as a category and split people's time more finely, or sample only during stops to see who waits and who helps.

The two sets of numbers check each other. If the study says operators wait for the line 11.9% of the time but the machine log says the line stopped 25% of planned time, people are filling the stops with other work, which is good to know before anyone cuts the crew.

Work sampling vs other ways to measure time

MethodWhat it givesCostPrecisionBlind spot
Work samplingShare of time per activity, for many people at onceLow: one observer, short visitsKnown, set by nNo element times; short events between rounds
Stopwatch time studyElement times for one repetitive jobHigh per jobHigh for that jobOne operator, one cycle type at a time
Continuous video or observationEvery activity in order, with durationsVery high to reviewHighSmall samples of time; privacy and consent
Machine signal (OEE monitoring)Run, stop, speed and count for every minuteLow once connectedCensus, no sampling errorSays nothing about people unless they log reasons
Timesheets and self-reportingHours booked to jobsLowUnknownRounded, booked after the fact, idle time rarely booked

Common mistakes

  • Rounds at fixed times. Every 30 minutes on the half hour catches the same routines every day. Randomise times and routes.
  • Too few observations. With 100 observations and p at 15%, the error is about plus or minus 7 points. That cannot tell 8% from 22%.
  • Confusing relative and absolute accuracy. State it as points, and give the confidence level.
  • Vague categories. If two observers would record the same moment differently, the shares are noise. Test the definitions in the pilot.
  • Studying only the day shift. Nights and weekends often run with fewer support staff and different waits.
  • Recording what someone "is doing generally". The method only works if you record the instant you see.
  • Sampling machines that could be counted. If the line has a signal, use it, and spend the observer's time on people.
  • Using it to judge individuals. The method estimates a group's shares. NPTEL notes it is not economical for one operator, and naming people kills trust in the next study.

For plant, production and CI managers

This week

  1. Pick one question that costs money, such as "how much operator time goes to waiting for material in packing?"
  2. Write five to eight categories with one-line definitions, and walk the area with the team leader to test them.
  3. Explain the study to the crew: what it is for, that it is about the process, and that no names are recorded.
  4. Run a one-shift pilot of about 50 observations and estimate p for the category that matters.
  5. Size the full study with the table above and set random round times in a spreadsheet.

This quarter

  1. Run the full study over at least two weeks, covering every shift.
  2. Report shares with their confidence range, turned into hours and money at your labour cost.
  3. Fix the biggest category with one change, such as kitting or a material call signal, and repeat the study to measure the effect.
  4. Connect the line's machines so that stop time and output are counted, not sampled, and drop "waiting for the line" from the people study.
  5. Repeat the study once or twice a year, or after any big change in layout or staffing.

A light MES helps on the counting side: it ties machine stops and output to orders, shifts and products, so the observer can focus on people.

How Fabrico fits

Fabrico is a cloud OEE and light MES platform. It shows OEE and losses by machine, line and shift, from the PLC or a Fabrico IoT device, with the reasons operators add on a tablet or phone. That covers the machine side of a work sampling study, so the study can focus on what people do.

On manual stations, AI cameras (an add-on) count output.

Book a demo and see OEE and losses per line and shift.

What people ask next

What is the work sampling technique?

An observer records what people or machines are doing at many random moments. The share of observations in each activity (working, waiting, walking, setting up) estimates the share of time spent on it. It needs no stopwatch and can cover many people at once.

What is the meaning of work sampling?

Work sampling means estimating how time is spent from a random sample of instant observations instead of timing everything. If 36 of 40 observations show an operator working, the estimate is that about 90% of the time is spent working.

What is activity sampling in manufacturing?

Activity sampling is another name for work sampling. In a plant it is used to find how much operator time goes to waiting for material, waiting for the line, walking, searching or setting up, so the biggest loss can be fixed first.

How many observations do you need for work sampling?

Use n = z² x p x (1 - p) / E². At 95% confidence (z = 1.96), an activity near 20% of the time needs 246 observations for plus or minus 5 points, 683 for 3 points and 1,537 for 2 points. Halving the error needs four times the observations.

Who invented work sampling?

L. H. C. Tippett described it as a snap-reading method for time studies of machines and operatives in the Journal of the Textile Institute in 1935, from work in the textile industry. It was later also called ratio delay study.

What are the disadvantages of work sampling?

It gives no element times, it can miss short events that happen between rounds, operators may change their behaviour when observed, and it is not economical for one operator or machine. For machines with a signal, continuous counting is more precise.

Sources

All sources accessed on 14 November 2026. Dates are the publication dates shown on each page.

Track OEE and downtime in real time
See where lines lose time
Book a live demo.
See where lines lose time
See where lines lose time

Real Fabrico tiles, sample data.

By clicking the Accept button, you are giving your consent to the use of cookies when accessing this website and utilizing our services. To learn more about how cookies are used and managed, please refer to our Privacy Policy and Cookies Declaration