AI Glossary / Statistics and probability / Intermediate
Variable sampling fraction
Definition
A variable sampling fraction means that different proportions of units are selected from different parts of a population, commonly from different strata. In stratified simple random sampling, stratum h has fraction nh/Nh, and these fractions vary across strata. This can deliberately oversample small or important groups. It does not mean that the sample size changes unpredictably, and it is distinct from variables sampling, an acceptance-sampling method based on quantitative measurements.
What it means
A small subgroup may need more observations than proportional allocation would provide if it must be reported separately. A high-variability group may also justify a larger sampling fraction. Such choices can improve selected objectives, but the resulting sample proportions no longer mirror population proportions automatically. Overall estimates usually need weights that account for unequal inclusion probabilities. Unequal fractions are not inherently biased; ignoring the design when estimating population quantities is the problem.
How it works
Define strata and their population sizes, choose a sample size within each stratum and record the resulting fractions. Under simple random sampling within strata, a unit's inclusion probability is nh/Nh and its basic design weight is Nh/nh. Estimate a population mean or proportion by combining stratum estimates with population shares, or use an equivalent weighted estimator. Variance estimation should reflect stratification and finite population sampling where applicable. More complicated selection schemes require their actual inclusion probabilities rather than assuming that a simple count ratio describes every unit's chance.
Three practical examples
Example 1: A company samples half of a small specialist workforce but only one-tenth of its larger general workforce. Example 2: An AI usage survey oversamples a rare user group to support a separate estimate. Example 3: An audit allocates a higher fraction to records in a high-variability stratum. These designs can serve legitimate objectives, but overall summaries must restore the intended population representation and account for the sampling design.
Think of it this way
Think of examining many items from a small box and a smaller share from a large box. The inspected items help you understand both boxes, but simply pooling them gives the small box too much influence on an overall estimate. Weights describe how much of the original collection each sampled item represents.
Worked example
Suppose stratum A contains 100 units and stratum B contains 900. Select 50 from A and 90 from B by simple random sampling, giving fractions 0.50 and 0.10. Their basic weights are two and ten. If 30 sampled units in A and 18 in B have a characteristic, the stratum sample proportions are 0.60 and 0.20. The population estimate is 0.10 times 0.60 plus 0.90 times 0.20, equal to 0.24. Equivalently, the weighted characteristic total is 30 times two plus 18 times ten = 240 out of a population of 1,000. The unweighted pooled fraction is 48/140, approximately 0.3429, which overrepresents the heavily sampled A stratum. The example assumes complete response and the stated probability design.
Illustrative case study
In this fictional case study, Beech Analytics surveys a small group of expert AI users and a much larger group of occasional users. It deliberately samples a larger fraction of experts to estimate their needs separately. The reporting team retains both strata's population sizes and applies design weights for the company-wide result. It also reports subgroup uncertainty rather than assuming that oversampling makes every comparison precise. The design serves both subgroup learning and overall estimation when its weighting and variance calculations are handled correctly.
Common misconceptions and mistakes
Common mistakes include describing unequal fractions as inherently unrepresentative, pooling observations without weights for an overall population estimate, and confusing this term with variables acceptance sampling. Another mistake is using planned selection fractions after substantial unequal nonresponse without considering appropriate adjustments. Weights cannot automatically repair an incomplete frame, inaccurate population counts or nonresponse related to unobserved outcomes. Extreme weights can also increase uncertainty, so allocation choices should consider precision as well as representation.
Business use
Businesses can use variable fractions when rare groups, different costs or unequal variability justify disproportionate allocation. Document why each stratum receives its share and which estimates the design prioritizes. Give analysts the stratum identifiers, population sizes, selection information and response details needed to reproduce estimates. Evaluate precision using the actual design. A weighted overall result and separate subgroup results can answer different useful questions; neither should be substituted for the other without explaining the target population.
People also ask
Does variable fraction mean the same count from every group?
Not necessarily; equal counts produce unequal fractions only when group sizes differ.
Is oversampling always wrong?
No.
Are weights generally needed for overall estimates under unequal inclusion probabilities?
Yes.
Is variable sampling fraction the same as variables sampling plan?
No.
Can nonresponse require further adjustments beyond basic design weights?
Yes.
Check your understanding
What fractions are used in the example?
Answer: 0.50 for A and 0.10 for B.
What are the basic weights?
Answer: Two and ten.
What is the weighted population proportion?
Answer: 0.24.
Why is the unweighted result larger?
Answer: It gives the high-rate A stratum more influence than its 10% population share warrants.
Related terms
- Related terms include stratified sampling
- disproportionate allocation
- inclusion probability
- survey weight
- oversampling and uniform sampling fraction
Learn next
Learn next how allocation rules balance population size, variability, cost and subgroup reporting needs. Then compare design weights with nonresponse and calibration adjustments. Practice computing both weighted estimates and design-appropriate uncertainty so that representation and precision are assessed together.
More in Statistics and probability
- Uplift modeling - Uplift modeling estimates how an intervention changes an outcome, often as a function of a person's or item's characteristics.
- Upper confidence bound - An upper confidence bound is an upper limit intended to cover an unknown quantity with a specified confidence under stated assumptions.
- Van der Waerden test - The Van der Waerden test is a rank-based test that compares independent groups by converting pooled ranks into standard normal quantiles, called normal scores.
- Variables sampling plan - A variables sampling plan is an acceptance-sampling procedure that uses quantitative measurements from sampled items to decide whether to accept a lot.
- Variance - Variance measures spread by averaging squared deviations from a mean.
- Variance components - Variance components are parameters that represent distinct sources of random variation in a statistical model.
Sources
Source: OECD, Glossary of Statistical Terms, Variable sampling fraction, https://www.oecd.org/content/dam/oecd/en/publications/reports/2008/09/oecd-glossary-of-statistical-terms_g1gh9ad7/9789264055087-en.pdf. This primary institutional glossary defines varying stratum selection fractions. The two-stratum weighting example and comparison with the unweighted result are original and independently checked. The AI-user survey case is fictional, and complete response is an explicit simplifying assumption.
Previous: Variable-order Markov model Next: Variables sampling plan