Stratified sampling is a probability sampling method in which a researcher first divides a population into distinct, non-overlapping subgroups — called strata — based on a shared characteristic (institution type, disease severity, career stage, geographic region, income band), and then draws an independent random sample from within each stratum. Because every stratum is guaranteed at least some representation, stratified sampling protects against the chance under-representation of a small but important subgroup that can occur with simple random sampling, and it typically produces more precise population estimates when the stratifying variable is related to the outcome being measured.
This guide covers how stratified sampling actually works, the difference between proportional and disproportionate (or “optimal”) allocation, a worked numeric example, when the method is the right choice versus its alternatives, and the practical requirements and limitations a Methods section needs to address.
How Stratified Sampling Works
Stratified sampling proceeds in four steps:
- Identify the stratifying variable. Choose a variable that is known for every member of the population in advance (from a sampling frame, registry, or administrative list) and that is plausibly related to the study’s outcome — for example, hospital size in a study of clinical protocol adherence, or academic rank in a study of research funding behavior.
- Divide the population into strata. Every population member is assigned to exactly one stratum; strata must be mutually exclusive and collectively exhaustive (every unit falls into one, and only one, group).
- Draw an independent sample from each stratum. Within each stratum, units are selected using simple random sampling or systematic sampling. Selection in one stratum has no bearing on selection in another.
- Combine the stratum samples into the full sample. The result is a single sample assembled from the sub-samples, weighted appropriately at the analysis stage if allocation was disproportionate (see below).
The prerequisite that makes or breaks this method is a complete and accurate sampling frame with the stratifying variable recorded for every population member. If that information isn’t available before sampling, stratification isn’t possible and a different method — commonly simple random or cluster sampling — has to be used instead.
Proportional vs. Disproportionate (Optimal) Allocation
Once strata are defined, the researcher decides how many units to draw from each one. There are two standard approaches:
- Proportional allocation — each stratum contributes to the sample in the same proportion it holds in the population. If a stratum is 20% of the population, it supplies 20% of the sample. This is the more common default and keeps the sample’s composition mirroring the population’s.
- Disproportionate (optimal) allocation — strata are sampled at different rates than their population share, usually to oversample a small-but-important stratum (a rare disease subgroup, a minority institution type) so it has enough cases for meaningful within-group analysis, or to allocate more of the sample to strata with higher internal variability (Neyman allocation), where more observations are needed to achieve the same precision.
Disproportionate allocation is powerful but has a direct cost: because the sample no longer mirrors the population’s true composition, any population-level estimate (an overall mean or proportion across the whole sample) must be corrected using sampling weights — typically the inverse of each unit’s selection probability — before it can be reported as representative. Software packages built for complex survey designs (e.g. R’s survey package, Stata’s svy commands, SPSS Complex Samples) handle this weighting directly; treating a disproportionately stratified sample as if it were a simple random sample and skipping the weighting step is a common and serious analytic error.
Worked Example
Suppose a researcher is surveying research-administration staff across a university system with a known population of 1,000 staff, distributed across three strata by institution type: 600 at large research universities, 300 at mid-size institutions, and 100 at small colleges. The researcher wants an overall sample of 200 (a 20% sampling fraction).
Proportional allocation keeps that 20% fraction constant within each stratum:
- Large research universities: 20% of 600 = 120
- Mid-size institutions: 20% of 300 = 60
- Small colleges: 20% of 100 = 20
The resulting 200-person sample (120 + 60 + 20) mirrors the population’s institution-type composition exactly, and — because each unit had the same overall selection probability — no reweighting is needed for population-level estimates.
If the researcher instead wanted to ensure the small-college stratum (only 100 in the population) had enough respondents to support its own subgroup analysis, they might apply disproportionate allocation — for instance, sampling 50 from small colleges instead of 20, and reducing the other two strata to keep the total near 200. Any overall estimate computed afterward would then need to be weighted back down to reflect that small colleges are actually only 10% of the true population, not the roughly 25% they now represent in the raw sample.
When to Use Stratified Sampling
Stratified sampling is the right choice when:
- The stratifying variable is known in advance for the full population (from an administrative list, registry, or sampling frame) — this is a hard requirement, not a nice-to-have.
- The stratifying variable is meaningfully related to the outcome, so isolating it improves precision rather than just adding sampling complexity for no statistical benefit.
- A subgroup is small relative to the whole population but important to analyze in its own right (a rare condition, a minority demographic, a small institutional category) — disproportionate stratification lets the researcher guarantee an analyzable number of cases from that subgroup.
- The researcher needs to report subgroup-level as well as population-level estimates and wants each subgroup’s precision to be controlled directly, rather than left to chance.
It is a poor fit when no reliable frame with the stratifying variable exists, when the population is effectively homogeneous with respect to any variable worth stratifying on (in which case simple random sampling achieves similar precision with less design complexity), or when the practical cost of building and maintaining separate stratum-level sampling operations outweighs the precision gained.
Stratified Sampling vs. Other Sampling Methods
Stratified sampling is one of several probability and non-probability sampling designs; see CASRAI’s overview of sampling methods for the full taxonomy. Two comparisons come up most often:
- Stratified vs. simple random sampling — simple random sampling treats the whole population as one pool and draws randomly from it with no subgroup guarantee; stratified sampling adds a division step first, which costs more setup but reduces the risk of a subgroup being under- or over-represented purely by chance, and generally shrinks the standard error of an estimate when the stratifying variable correlates with the outcome.
- Stratified vs. cluster sampling — the two are easy to confuse because both involve dividing the population into groups first, but the mechanics are opposite. In stratified sampling, the researcher samples within every stratum (all groups are represented, and individuals are sampled from each). In cluster sampling, whole clusters are randomly selected and then either every member of a selected cluster is included, or a sample is drawn only from the selected clusters — meaning most clusters, and everyone in them, are excluded entirely. Cluster sampling is typically chosen for logistical convenience (e.g., sampling entire clinics or classrooms rather than scattered individuals) at some cost to precision; stratified sampling is chosen specifically to improve precision at some cost to logistical simplicity.
Stratified sampling is also distinct from quota sampling, its non-probability cousin: quota sampling fills the same kind of subgroup targets but does so through non-random, convenience-based selection within each group, which means it does not carry the same statistical guarantees or allow formal margin-of-error calculations.
Advantages and Limitations
| Advantages | Limitations |
|---|---|
| Guarantees representation of every defined subgroup, removing the risk that a small stratum is missed entirely by chance. | Requires a complete, accurate sampling frame with the stratifying variable known in advance — often the binding constraint in practice. |
| Generally increases precision (reduces standard error) relative to simple random sampling of the same total size, when the stratifying variable is related to the outcome. | Adds design and administrative complexity — separate sampling operations, and weighting requirements if allocation is disproportionate. |
| Enables planned, adequately powered subgroup analyses via disproportionate allocation. | Choosing a poor or irrelevant stratifying variable adds cost without a meaningful precision benefit. |
| Supports direct comparison between strata as a designed feature of the study, not an after-the-fact subgroup analysis. | Misclassification of population members into the wrong stratum introduces bias that is hard to detect after the fact. |
Reporting Stratified Sampling in a Methods Section
A Methods section using stratified sampling should report: the stratifying variable(s) and why they were chosen; the source of the sampling frame; whether allocation was proportional or disproportionate (and, if disproportionate, the actual allocation and the weighting approach used at analysis); and the final achieved sample size per stratum. Reviewers use this information to judge whether the sample’s precision and generalizability claims are justified — the same standard covered in CASRAI’s comparison of population vs. sample.
Frequently Asked Questions
Is stratified sampling a type of probability sampling?
Yes. Stratified sampling is a probability sampling method: within each stratum, units are selected using a random mechanism with a known, non-zero selection probability, which is what allows standard inferential statistics to be applied to the results.
What’s the difference between stratified sampling and random sampling?
Simple random sampling draws directly from the whole population with no grouping step. Stratified sampling first divides the population into subgroups (strata) and then applies random sampling independently within each one, guaranteeing every stratum is represented rather than leaving that to chance.
Can strata overlap?
No. A valid stratification scheme requires strata to be mutually exclusive and collectively exhaustive — every population member belongs to exactly one stratum. Overlapping or incomplete strata invalidate the method’s statistical properties.
How many strata should a study use?
There’s no fixed rule; the right number balances the precision gained from finer subdivision against the practical cost of very small per-stratum sample sizes and increased data-collection complexity. Most designs use somewhere between two and a handful of strata based on one or two substantively meaningful variables, rather than stratifying on many variables at once, which quickly produces strata too small to sample reliably.
Do I need to weight my data after stratified sampling?
Only if allocation was disproportionate — that is, if strata weren’t sampled at the same rate as their share of the population. Proportionally allocated samples don’t need reweighting for population-level estimates, since every unit already had the same overall selection probability.
