Work Sampling in Garment Factories: Measure Activity Without Timing Every Cycle

A quiet machine does not explain the work around it

An industrial engineer walks past a sewing station and sees that the machine has stopped. The operator may be checking a seam, moving a bundle, waiting for panels or taking an agreed break. One glance cannot show how often each state occurs across the week.

Work sampling builds that picture through planned observations. At selected moments, an observer records the activity of a selected person or machine. The share of observations in each category can estimate its share of time. That estimate needs a sound sampling plan and clear records.

It is useful when the question concerns how time is distributed. It does not require timing every sewing cycle. Nor does it turn a brief visit into a judgement about a worker's effort.

The factory, observation counts and calculations below are fictional. They demonstrate a study design and its limits. No percentage is offered as a target for garment factories or a finding from an actual workplace.

Start with one decision the study can support

Suppose a production team thinks operators spend much of the week waiting for work. It wants to change how panels reach the line. First, it needs to know where delays occur and how much time they take.

The question is how much of the defined operator time involves waiting for work to proceed. A second question asks which causes need a closer look. Neither question requires ranking people from best to worst.

Define the target group before watching it. This example covers ten named operator positions over five normal production days. Each has 7.5 hours within the study window per day, including agreed short breaks. Meal periods outside that window are excluded.

For the teaching case, all ten positions remain staffed throughout those windows. The total is 10 × 5 × 7.5 = 375 worker-hours. One hour with ten workers present is ten worker-hours, not one.

A real study needs rules for absence, relief staff, training and overtime. If those conditions change, update the frame and disclose the change. One selected line and week may not represent the whole factory.

Choose the right observation unit

Here, one observation means one selected operator position at one selected instant. It is not a completed garment, a whole round of the floor or a full minute of continuous work.

The frame must cover all eligible position-time combinations. If only the first row of machines is easy to see, those positions must not silently become the entire sample. Solve the access problem during fieldwork planning. If the study must cover a smaller group, state that change.

Sampling a machine answers a different question from sampling its operator. A machine can run automatically while the person performs another task. Conversely, the person can be doing useful work while the machine is stopped.

Choose which activity the record describes. If both are required, use separate defined fields and preserve their relationship. Do not mix machine-running observations with worker-activity observations in one percentage.

The IIT Kanpur notes on sampling proportions explain how a defined class is estimated from sample counts. Decide what belongs to each class before collecting the data.

Give each activity a clear meaning

“Working” and “not working” are often too broad. The team needs categories that distinguish useful support work from a blocked process. It also needs a rule for activities that seem to fit more than one label.

These illustrative categories suit the fictional study. They would need adaptation for another task or production area.

The categories should cover the intended work and avoid overlap. If an unplanned but valid activity appears, retain a description. The team may need to change the codes. It must then check whether old and new records still mean the same thing.

Do not call all non-sewing time waste. Checking, replenishment, safe setup and rest can be necessary. A category records what happens; it does not decide whether the activity should be removed.

A 2025 construction study in Purdue's CIB collection used random observations to study work flow. The setting differs from a garment line. The method still shows why it helps to separate activity types. Its reported results should not be imported as apparel benchmarks.

Explain the study and protect normal work

Workers should know why the study is taking place, what will be recorded and how the results will be used. Discuss the category definitions with them. They may spot a task that an outside observer would mistake for waiting.

The HSE guidance on consulting workers describes consultation as a two-way process. Its legal duties apply within its jurisdiction. Here, workers should have space to explain the process and raise concerns about the study.

Agree safe observation positions. The observer must keep aisles clear and avoid distracting workers at critical moments. A planned sample time does not permit entry to a restricted area. A missed observation is preferable to creating a hazard.

Use position codes where names are unnecessary. Only those who need personal records should have access to them. Apply the site's relevant data rules. A public board should explain process findings without exposing private notes about individuals.

Observation can change behaviour. Familiarisation may reduce the novelty, but it does not prove that the observer has no effect. Record unusual visits, coaching or changes during the study. Ask whether these events could have changed the observed work.

Select moments without following the production rhythm

For a simple design, draw position-time pairs at random from the complete study frame. Each eligible pair should have the same chance of selection. Generate the plan before seeing the activity outcomes and keep the original schedule.

The observer then records the selected position at its planned instant. A convenient walk every hour is a different design. It may keep catching the same bundle delivery while missing the work between visits.

Sort the random schedule by time to arrange fieldwork. Check whether enough observers can cover nearby appointments. Do not keep redrawing inconvenient times until the plan contains only easy visits. Any practical restriction should be designed and accounted for in advance.

If an observation is late, record the actual time and the deviation. A snapshot taken five minutes later describes a different instant. The study plan needs a rule for accepting such deviations or treating them as missed.

Never wait until the operator resumes sewing and then record that as the selected observation. Equally, do not stay beside a stopped machine to collect extra waiting records. Both choices let the outcome influence selection.

Choose precision before choosing the number of visits

Sample size depends on the uncertainty the decision can tolerate. Under a simple independent-observation model, a common approximate planning formula is:

n = z² × p × (1 − p) ÷ e².

Here, n is the number of valid observations. The planning proportion is p. The desired absolute margin is e, written as a decimal, and z sets the confidence level. For an approximate 95% interval, z is 1.96.

The formula uses the variance of a sample proportion and a normal approximation. It needs a suitable random design with independent observations. A design that closely meets those conditions may also support its use. It does not account for biased missing records or misclassification.

If there is no sound prior estimate, p = 0.5 gives the largest value of p × (1 − p). With a desired margin of five percentage points:

n = 1.96² × 0.5 × 0.5 ÷ 0.05² = 384.16.

Round upward to 385 valid observations. Five percentage points is an absolute margin. It does not mean five percent of whichever proportion the study eventually finds.

Halving the planned margin to 2.5 percentage points gives 1,536.64, rounded upward to 1,537 observations. The cost rises quickly as the desired interval narrows. These are approximate planning figures, not guaranteed final interval widths.

The fictional team schedules 420 observations to allow some fieldwork loss. It does not assume that this buffer will solve every problem. The final valid count and reasons for missing data still need review.

Test the definitions before the main study

A short pilot tests both access and the codes. Can observers reach each position on time? Do they agree on difficult cases, such as waiting for a check versus checking a piece?

In a separate fictional pilot, two observers independently code the same thirty moments. They agree on twenty-six. Raw agreement is 26 ÷ 30 × 100 = 86.7%.

That figure alone does not prove that the system is reliable. Agreement may be high because one activity is common. Review the four disagreements and note which codes were confused. Were the definitions unclear, or did the observers have different views of the work?

The two codes for a single moment are a check on measurement. They are not two independent moments to add to the main sample. Keep the pilot records separate unless a defined design supports their later use.

After revising a definition, try fresh examples. An observer should not merely remember which answer the other person gave. The aim is a shared rule that also works on unfamiliar situations.

Keep the record that sits behind each total

A final tally should lead back to the individual observations. Give each planned position-time pair a unique record number. Keep its scheduled time, actual observation time, position, observer and activity code together.

A short note can explain a difficult classification. “Hands on seam, checking against the approved sample” is more useful than “looked busy.” Record what could be seen, rather than a guess about the person's attitude or intention.

An unknown activity differs from a missed visit. With an unknown, the observer was present but could not classify the state. With a missed visit, the selected instant was not observed. Both create gaps, but they may need different remedies.

The fictional results below contain twenty missed visits and no additional unknown codes. If unknown codes had occurred, their count would also need disclosure. They should not disappear simply because the percentage table needs a clean denominator.

Keep a later cause check separate from the original snapshot. The observer may record waiting at 10:23, then learn from the production log that a hold was in force. Add the linked explanation without pretending it was known during the observation.

Review totals by day and position before pooling them. This can reveal a copied row, an omitted afternoon or one observer's unusual use of a code. It does not justify deleting a difficult day merely because its result differs from the others.

Reconcile the observations before making a percentage

Suppose the main study produces the following teaching results. There are 400 classified observations and twenty planned observations that could not be obtained. The missing records remain in a separate log.

The waiting estimate among classified records is 72 ÷ 400 = 0.18. Its denominator is visible. Do not divide by 420 while silently treating every missing record as not waiting.

Check duplicates, invalid codes and timing deviations before accepting the total. Preserve corrections with a reason so another reviewer can reproduce the result. A total that adds to 100% can still contain poor observations.

The 54% sewing share is not sewing efficiency. It describes activity at selected moments. Efficiency based on standard minutes and output asks another question. ApparelDiary's SMV calculation guide explains the basis of that different measure.

Show what the missing records could change

Twenty out of 420 planned observations are missing, or 4.76%. Whether that matters depends partly on why they were missed. For example, material deliveries might block the view. The missing records could then depend on the activity being studied.

For the selected set of 420 moments, consider two extreme completions. If none of the twenty missing moments involved waiting, the full-set proportion would be 72 ÷ 420 = 17.14%.

If all twenty involved waiting, it would be 92 ÷ 420 = 21.90%. This range is a sensitivity check for the missing selected records. It is not a confidence interval for all 375 worker-hours.

Review missingness by time, location and observer. Similar overall missing percentages can hide a whole shift or inaccessible area. Extra observations made later do not reveal what happened at the original missed moments.

The ONS guide to uncertainty separates sampling error from other errors. Missing observations and wrong records can cause these other errors. A statistical interval does not, by itself, cover them.

Add a confidence interval with its assumptions attached

For this calculation, assume that the 400 classified observations are independent draws that represent the target time. Also assume that the missing observations do not bias the sample that remains. Those assumptions would need review in a real study.

With 72 waiting observations, the estimated proportion is 0.18. The NIST guide to proportion intervals gives the Wilson method. Using n = 400 and z = 1.96 gives an approximate 95% interval from 14.54% to 22.06%.

For readers checking the arithmetic, the Wilson centre is:

[p + z² ÷ (2n)] ÷ [1 + z² ÷ n].

Its half-width is:

z × √[p × (1 − p) ÷ n + z² ÷ (4n²)] ÷ [1 + z² ÷ n].

Here, p is the observed proportion, 0.18. The centre is about 0.18304 and the half-width about 0.03759. Subtracting and adding the half-width gives the two limits.

The interval is not exactly centred on 18%. That is expected with this method. Nor does “95%” mean that 95% of the observations are correct. Imagine repeating the sampling process many times under those assumptions. About 95% of the resulting intervals would contain the true proportion.

This is one interval for the waiting proportion. It is not a joint guarantee for every activity in the table. If the design is clustered, weighted or otherwise more complex, use an analysis suited to that design.

Do not treat a whole observation round as unrelated snapshots

A different field method might record ten workers during each floor visit. If one material shortage affects all ten, their states in that visit may be closely related. A hundred recorded cells are not automatically equivalent to a hundred independent draws.

The IIT Kanpur notes on cluster sampling show why the grouping affects variance. The analysis needs the visit or group identifiers, not just the final activity totals.

This does not mean that every repeated observation of a worker is invalid. The issue is the selection design and dependence relevant to the estimate. Drawing each position-time pair on its own differs from taking many nearby snapshots during a few convenient rounds.

Spread the intended study over the periods it aims to represent. If the design uses rounds, plan enough independent rounds. Keep the records grouped by round for analysis. Do not invent a correction factor after seeing an inconvenient interval.

Weight unequal groups to match the question

Separate shifts or product groups can need their own samples. The IIT Kanpur notes on stratified sampling explain how group estimates are combined using the groups' population shares.

Consider a separate illustration. Day-shift time forms 80% of the target worker-hours and evening time forms 20%. Random samples within those shifts estimate waiting at 10% and 30%, respectively.

The combined time-share estimate is:

(0.80 × 0.10) + (0.20 × 0.30) = 0.14, or 14%.

Suppose each shift supplied 100 observations. Pooling the counts would give (10 + 30) ÷ 200 = 20%. That gives the smaller evening shift too much influence for the stated worker-hour question.

Keep the sampling plan, group sizes and weights with the result. A confidence interval for this weighted estimate also needs the stratum design. The simple Wilson calculation for the earlier unweighted example cannot be transferred unchanged.

Convert a share into hours without calling every hour a saving

Suppose the 18% estimate reasonably represents the full 375 worker-hours. Projected waiting time is then 0.18 × 375 = 67.5 worker-hours. This result depends on the study frame, activity codes and assumptions about missing data.

It is not a stopwatch total. The observer did not directly record the start and finish of every delay. Event records can show how long delays lasted. A closer study of the affected process can help explain their causes.

The 67.5 hours also do not establish removable labour or an avoidable wage payment. Some delays may overlap across people. Solving one may expose another bottleneck, and freed time may be too scattered to use for extra output.

Verify causes separately. An operator may lack panels for several reasons. Cutting or transport may be late, an approval may be missing, or production may be on hold. The snapshot shows the state; it does not establish which department caused it.

Remember that time share and event frequency differ

Work sampling is more likely to encounter a long activity than a brief one. That is useful when estimating time share. It becomes misleading if the same snapshots are described as a count of separate incidents.

Consider a separate, fully logged 100-minute example. There is one thirty-minute wait, ten one-minute waits and sixty minutes in other states. Total waiting is forty minutes, or 40% of the period.

The long wait accounts for 30 ÷ 40 = 75% of waiting time. Yet it is only one of eleven waiting events, or 9.1% of the event count. A random-time study can often see that long wait without having observed many distinct failures.

The distinction changes the next action. One long interruption may need a different remedy from frequent small stops. Use event logs when the frequency and length of delays matter. The number of waiting snapshots cannot establish those details.

Very brief or rare events can also be missed entirely. No recorded event in a modest sample is not proof that the event never occurs. Work sampling cannot certify that a hazard is absent. It also cannot replace required checks for a critical process.

Use the result to choose a focused next step

Bring the activity pattern back to the people who do the work. Check whether it fits their experience and investigate differences. Their account can suggest causes, while the records help test those explanations.

If late panels appear to be the main issue, trace a few affected handovers. Test a specific improvement and keep product mix, study scope and category definitions comparable in the follow-up. A second lower percentage alone does not prove that the action caused the change.

Report the question, frame, planned and valid counts, missing reasons, point estimate and uncertainty together. Include the limitations that affect the decision. ApparelDiary's factory KPI dashboard guide can help keep those definitions visible.

Work sampling should leave the factory with a clearer process question and evidence for its next action. It should not turn essential checks or agreed breaks into unexplained losses, or a few snapshots into permanent labels for workers.

References

• IIT Kanpur, Shalabh: sampling for proportions and percentages.

• Fiirgaard, Namage and Toca Pérez: location-based work sampling research, 2025.

• HSE: consulting workers about health and safety.

• ONS: uncertainty, sampling error and confidence intervals.

• NIST: Wilson confidence intervals for a proportion.

• IIT Kanpur, Shalabh: cluster sampling.

• IIT Kanpur, Shalabh: stratified sampling and weighted estimates.