One of the defined subgroups; strata is the plural.

Can a stratum be fully enumerated?

Yes. A design can include every unit in a small or important group.

Does stratification guarantee representativeness?

No. Frame coverage, selection, response, and measurement still matter.

AI Glossary / Statistics and probability / Beginner

Strata

Definition

Strata are the distinct subgroups into which a population is divided for a statistical purpose, especially before sampling. The singular is stratum. In a stratified sampling design, the groups are defined so that each population unit belongs to exactly one stratum, and samples are selected within the strata. This structure can support coverage of important subgroups and more precise estimation when the grouping is useful for the measured characteristic.

What it means

Strata make a population's structure explicit. A business survey might separate small, medium, and large firms; an evaluation might separate requests by language or task type. The definition should be operational enough to assign units consistently.

The groups need not have equal sizes, and the sample need not take the same fraction from every group. A small but important stratum may be sampled heavily or completely. When combining results, the analysis must respect the design rather than treating the resulting sample mix as the population mix. Strata are also used in analysis and experimental allocation, but this entry emphasizes their survey-sampling role.

How it works

Define the population and choose grouping variables available before selection. Create non-overlapping categories that collectively cover the population, including a deliberate approach to missing or uncertain group information. Document boundaries so that two analysts would assign the same unit to the same stratum.

Choose sample sizes and selection methods within the strata. Allocation may reflect population size, variability, cost, or the precision required for subgroup reporting. The grouping and allocation should serve the research question, not merely create visually balanced sample counts.

Estimate stratum-specific quantities and combine them with appropriate weights. Under simple random sampling within each stratum, a basic design weight is the stratum population size divided by its sample size. More complex designs and nonresponse may require additional treatment.

Evaluate whether the strata remain accurate and useful. Outdated classifications, sparse groups, or variables weakly related to the target can reduce the expected benefit. Stratification does not by itself solve nonresponse, measurement error, or poor coverage of the population frame.

Three practical examples

1. A survey separates firms by size and samples more heavily from a small group of very large firms that contribute substantially to total activity.

2. An AI evaluation samples enough requests from each major language group to support meaningful subgroup estimates, then weights appropriately for an overall result.

3. A customer study defines mutually exclusive subscription tiers and reports both tier-specific outcomes and a population-weighted summary.

Think of it this way

Think of checking fruit quality in separate labeled baskets. Sampling from every basket ensures that a small basket is not accidentally overlooked. But if the baskets contain different numbers of fruit, equal sample counts do not mean the baskets should contribute equally to the overall quality estimate.

Worked example

Consider an original population of 1,000 customers. Stratum A contains 800 customers and stratum B contains 200. The team independently selects simple random samples of 80 from A and 100 from B, with complete responses.

Suppose 40 sampled A customers and 80 sampled B customers report using a feature. The sample proportions are 40 / 80 = 0.5 and 80 / 100 = 0.8. Population weights are 800 / 1,000 = 0.8 and 200 / 1,000 = 0.2. The combined estimate is 0.8 × 0.5 + 0.2 × 0.8 = 0.56, or 56%.

Equivalently, each A respondent has design weight 800 / 80 = 10 and each B respondent has weight 200 / 100 = 2. The estimated number of users is 40 × 10 + 80 × 2 = 560.

Ignoring the design gives 120 users out of 180 respondents, about 66.7%. That overrepresents the more heavily sampled B stratum in this example. The weighted estimate still has sampling uncertainty; weighting does not turn a sample into a census or correct every possible data-quality problem.

Illustrative case study

In a fictional company, Aspen Assist evaluates an AI service across two subscription tiers. The smaller tier receives a larger sampling fraction because the team needs a usable subgroup estimate. An early report pools all responses and shows an overly favorable overall result because the smaller tier performs better. The analyst rebuilds the summary using the intended population shares and keeps both tier results visible. Managers can now see the difference between overall performance and the experience of a deliberately oversampled group.

Common misconceptions and mistakes

Strata do not have to be equally sized or equally sampled. Equal allocation can be useful, but it requires appropriate interpretation when population sizes differ.

Stratification is not the same as cluster sampling. A basic stratified design samples within every stratum, whereas cluster designs select groups and may omit unselected groups entirely.

A subgroup created after observing outcomes is not automatically a prespecified sampling stratum. Poststratification and exploratory subgroup analysis have different procedures and assumptions. Finally, similar units within a stratum are often desirable for precision, but perfect homogeneity is not required and cannot simply be assumed.

Business use

Strata help businesses plan data collection around meaningful population differences. They can protect important small groups from being overlooked and make sampling resources more effective.

Before collecting data, identify which subgroup estimates will influence decisions and whether the frame supports the intended classification. Preserve the stratum identifier and selection probabilities with each record. When communicating results, distinguish the sample composition from the target population composition. An AI benchmark with equal numbers per task type may be useful diagnostically but needs justified weighting to represent actual usage.

Check your understanding

Answer: A population is 90% group A and 10% group B, but the sample contains equal numbers from both.

Should an overall population mean automatically average the two group means equally?

Answer: No. Under the appropriate simple stratified design, population shares should determine their contributions.

Related terms

Stratified sampling
Population
Cluster Sampling
stratum
sampling frame
allocation
design weight
poststratification

Learn next

Study proportional and disproportionate allocation, then learn how stratified estimates and their uncertainty are calculated under the actual design.

Sources

NIST Engineering Statistics Handbook glossary, strata: https://www.itl.nist.gov/div898/handbook/glossary.htm . Statistics Canada, Probability sampling, stratified-sampling section: https://www150.statcan.gc.ca/n1/edu/power-pouvoir/ch13/prob/5214899-eng.htm . The weighted customer calculation and fictional AI evaluation are original teaching examples.

Previous: Straight-through processing Next: Strategic form