Is a 10% sample always enough?

No; precision depends on sample size, variability, design, and the question.

Is the fraction the response rate?

No; response rate concerns participation among relevant sampled units.

Does f equal one prove accuracy?

It removes sampling variation for a fully observed finite population, not all error.

AI Glossary / Statistics and probability / Intermediate

Sampling fraction

Definition

The sampling fraction is the proportion of a defined finite population selected into a sample, commonly written f = n/N, where n is the sample size and N is the population size. In survey sampling, it helps describe the design and determine whether sampling without replacement materially affects uncertainty.

What it means

Sampling 100 units from a population of 500 gives a fraction of 20%. Sampling the same 100 units from 500,000 gives 0.02%. The sample sizes are equal, but the amount of the finite population directly observed is very different.
The fraction does not itself measure representativeness, response quality, or statistical power. Its interpretation requires a clear population, distinct eligible units, and an explanation of how the sample was selected. A percentage of records is not automatically a percentage of people.

How it works

Define N using the population relevant to the estimate. A mailing list, a customer database, and the target population may have different sizes. Duplicates, ineligible entries, and missing units should not be silently included in the denominator.
Define n consistently. In a sample selected without replacement, it normally counts distinct selected population units. The initial selected sample and the achieved responding sample can have different counts, so label which fraction is being reported.
Calculate f by dividing n by N. Multiply by 100 to express it as a percentage. This descriptive calculation is straightforward, but its connection to uncertainty depends on the design.
For simple random sampling without replacement, every population unit has inclusion probability n/N. In unequal-probability designs, individual inclusion probabilities can differ, so a single overall fraction does not describe every unit's selection chance.
For the sample mean under simple random sampling without replacement, the sampling variance is (1-f)S squared/n, where S squared is the finite-population variance calculated using denominator N-1. The factor 1-f is a finite-population correction for variance.
The corresponding correction for a standard error is the square root of 1-f. If population variance instead uses denominator N, an algebraically equivalent formula includes (N-n)/(N-1). State the variance convention before comparing formulas.
When the fraction is very small, the correction is close to one and may have little numerical effect. When the sample covers much of the finite population, omitting it can materially overstate sampling variability under the specified design.
Complex designs may require separate fractions by stratum or sampling stage. A large fraction of selected schools is different from a large fraction of pupils within those schools. Use design-aware variance estimation rather than applying one correction to every analysis.
Finally, decide whether inference concerns the observed finite population or an ongoing process producing future units. Enumerating today's complete customer list does not remove uncertainty about future customers or measurement errors.

Three practical examples

Inventory audits: distinguish reviewing 100 of 500 items from reviewing 100 of 500,000.
Customer surveys: report the selected fraction separately from the response rate.
AI evaluation: document how much of a finite labeled benchmark was randomly selected for a manual quality audit.

Think of it this way

Think of checking pieces in a fixed box. Examining nearly the whole box leaves relatively little uncertainty about that box's average, assuming accurate measurement and proper selection. It does not establish the average quality of boxes that have not yet been produced.

Worked example

Consider an original planning example with a finite population of 500 units. Select 100 distinct units by simple random sampling without replacement. The sampling fraction is 100/500 = 0.20, or 20%.
Suppose the finite-population variance S squared, defined using denominator 499, is 25. The exact design variance of the sample mean is (1-0.20) times 25/100 = 0.20. Its standard error is square root of 0.20, approximately 0.4472.
Ignoring the finite-population correction while using the same S squared would give variance 25/100 = 0.25 and standard error 0.5. The correction multiplies that standard error by square root of 0.80, approximately 0.8944.
If all 500 units were measured, f would equal one and the sampling variance of the finite-population mean would be zero. This statement concerns sampling only. Incorrect measurements, incomplete coverage, or inference about future units would still need separate treatment. The stipulated variance is supplied for illustration, rather than estimated from an invented survey.

Illustrative case study

Fictional case: Willow Audit reviews a random subset of a fixed library of AI-generated summaries. Because the review covers a substantial fraction of the library, analysts account for sampling without replacement when estimating its average quality. They separately record summaries that reviewers could not assess and do not treat nonresponse as a harmless reduction in sample size. When management asks about next month's summaries, the team explains that the current library's sampling fraction does not establish uncertainty for future production.

Common misconceptions and mistakes

Do not confuse sampling fraction with response rate: their denominators differ. A large fraction can still be biased if selection is inappropriate. Do not count repeated draws of one unit as distinct population coverage. Do not apply a finite-population correction merely because a dataset has a known row count; the target and design must justify it. A census does not remove non-sampling errors.

Business use

Use the fraction to document survey coverage by selection, plan audits of finite inventories, and check whether standard-error calculations reflect substantial sampling without replacement. Keep it alongside the absolute sample size, design, population definition, and achieved response information. It is one design descriptor, not a standalone quality score.

Check your understanding

What is the fraction for 100 of 500 units?

Answer: Twenty percent.

Which correction multiplies the variance in the stated convention?

Answer: One minus f.

Why is the standard-error correction different?

Answer: Standard error is the square root of variance.

Related terms

Sample size
Simple Random Sampling
Stratified sampling
finite population
finite-population correction
inclusion probability
response rate
sampling variance

Learn next

Study design-based variance estimation and inclusion probabilities. Then compare finite-population inference with prediction for future observations, where the relevant uncertainty can remain after a census.

Variance estimation

Sources

Statistics Canada, Survey Methods and Practices: https://www150.statcan.gc.ca/n1/en/pub/12-587-x/12-587-x2003001-eng.pdf Australian Bureau of Statistics, Errors in Statistical Data: https://www.abs.gov.au/websitedbs/d3310114.nsf/home/Basic%20Survey%20Design%20-%20Errors%20in%20Statistical%20Data The finite-population calculation is original.

Previous: Sampling error Next: Sampling probability