Cluster sampling is a probability sampling method in which a researcher divides a population into naturally occurring groups — called clusters — such as schools, clinics, villages, or households, then randomly selects a subset of whole clusters rather than sampling individuals directly from the full population. Everyone in a selected cluster may be included (one-stage cluster sampling), or a sub-sample may be drawn from within each selected cluster (two-stage cluster sampling). Most clusters that are not selected, and everyone in them, are excluded entirely.
This guide covers how cluster sampling works, why it is chosen, how it differs from stratified sampling (the comparison readers confuse most often), the design effect it introduces, one-stage/two-stage/multistage designs, how to analyze clustered data correctly, and when cluster sampling is the wrong choice.
What Is Cluster Sampling?
Cluster sampling starts from an existing grouping structure in the population — a list of schools, clinics, villages, city blocks, or households — rather than a list of individuals. The researcher:
- Divides the population into clusters. Clusters are naturally occurring or administratively convenient groups, not artificial categories created for the study.
- Randomly selects a subset of clusters. Selection is typically simple random sampling or probability-proportional-to-size (PPS) sampling from a list of all clusters — the sampling frame in cluster sampling is a frame of clusters, not individuals.
- Samples within each selected cluster. Either every member of a selected cluster is included (one-stage), or a further random sub-sample is drawn from within each selected cluster (two-stage, or multistage with additional levels).
The defining feature is that clusters, not individuals, are the unit initially selected at random. An individual’s chance of appearing in the final sample depends on their cluster being selected first.
Why Researchers Use Cluster Sampling
The honest reason cluster sampling is used is cost and feasibility, not statistical merit — unlike stratification, which is chosen specifically to improve precision, clustering is a deliberate trade of precision for practicality. Two structural advantages drive that trade-off:
- No population-wide sampling frame is needed. Building and maintaining a complete list of every individual in a population is often impossible or prohibitively expensive. A list of clusters — every clinic in a health system, every school in a district, every village in a region — is usually available even when a list of every individual is not.
- Fieldwork is geographically and logistically concentrated. Interviewing 500 people scattered randomly across a country requires traveling to 500 different locations. Interviewing 500 people drawn from 20 randomly selected villages requires traveling to 20 locations. For in-person data collection, this difference determines whether a study is affordable at all.
Large national household surveys — including USAID’s Demographic and Health Surveys (DHS) and UNICEF’s Multiple Indicator Cluster Surveys (MICS) — use cluster sampling for exactly this reason: a full national list of individuals or households doesn’t exist in a form that could be sampled directly, but a list of census enumeration areas does.
Cluster Sampling vs. Stratified Sampling
These two methods are the most commonly confused in the sampling methods taxonomy, because both start by dividing the population into groups. The mechanics — and the statistical consequences — are opposite. See CASRAI’s stratified sampling guide for the full mechanics of that method.
| Cluster Sampling | Stratified Sampling | |
|---|---|---|
| What is sampled | A random subset of whole clusters (groups are the sampling unit) | A random subset of individuals from every stratum (individuals are the sampling unit) |
| How many groups are represented | Only the selected clusters — most clusters are excluded entirely | Every stratum — all groups are represented |
| Ideal internal composition of each group | Heterogeneous — each cluster should look like a mini-population | Homogeneous — each stratum should be internally similar |
| Effect on variance / precision | Increases variance relative to simple random sampling of the same size | Reduces variance relative to simple random sampling of the same size |
| Primary reason to choose it | Cost and logistical feasibility — no full population frame needed | Statistical precision and guaranteed subgroup representation |
| Sampling frame required | A list of clusters only | A list of every individual, with the stratifying variable known for each |
The core distinction to hold onto: stratified sampling groups people so that each group is internally similar and then samples from every group, which improves precision. Cluster sampling groups people so that, ideally, each group is internally diverse (a good miniature of the whole population) and then samples only some groups, which trades precision for cost. In practice, naturally occurring clusters (a clinic’s patients, a school’s students) tend to be more similar to each other than the population at large — patients at one clinic often share demographics, referral patterns, or care quality in ways that make the cluster less heterogeneous than the ideal. That similarity is exactly what produces the design effect covered below.
One-Stage, Two-Stage, and Multistage Cluster Sampling
- One-stage cluster sampling — clusters are randomly selected, and every member of each selected cluster is included in the sample. Simple to execute, but sample size is determined entirely by cluster sizes, which the researcher doesn’t control precisely.
- Two-stage cluster sampling — clusters (the primary sampling units, or PSUs) are randomly selected first, then a random sub-sample is drawn from within each selected cluster (the secondary sampling units). This gives the researcher direct control over final sample size and reduces fieldwork within any single cluster.
- Multistage cluster sampling — more than two levels of random selection, each stage narrowing the frame further. This is the design most large national surveys actually use.
Worked example (illustrative): a national survey of primary-care patient experience could use a three-stage design: Stage 1 — randomly select a sample of districts (or census enumeration areas) from a national list, often with probability proportional to population size (PPS), so larger districts are more likely to be selected. Stage 2 — within each selected district, randomly select a sample of clinics. Stage 3 — within each selected clinic, randomly select a sub-sample of patients from recent visit records. Each stage narrows a frame that would have been unmanageable to build directly at the individual level, while still preserving a known, calculable selection probability for every patient in the final sample — the property that keeps the design a valid probability sample rather than a convenience sample.
The Statistical Cost: The Design Effect
Cluster sampling isn’t free. Because people within a naturally occurring cluster tend to resemble each other — patients at the same clinic, students in the same classroom — observations within a cluster carry less independent information than the same number of observations drawn via simple random sampling. The design effect (DEFF) quantifies exactly how much less:
DEFF = 1 + (m − 1) × ICC
where m is the average cluster (sub-)sample size and ICC is the intraclass correlation coefficient — a measure of how much of the total variance in the outcome is attributable to differences between clusters rather than within them. When ICC = 0 (clusters are no more similar internally than the population at large), DEFF = 1 and clustering costs nothing statistically. As ICC rises, DEFF rises, and a clustered sample of size n becomes worth progressively fewer than n independent observations.
Worked example (illustrative): suppose the primary-care survey above samples m = 25 patients per selected clinic, and prior research suggests patient-satisfaction scores have an ICC of 0.10 across clinics (a plausible order of magnitude for outcomes that cluster moderately by site — ICC values in health-services and education research are commonly reported in roughly the 0.01–0.20 range, though this varies by outcome and setting and should be checked against comparable prior studies, not assumed).
DEFF = 1 + (25 − 1) × 0.10 = 1 + 2.4 = 3.4
If a simple random sample of 400 patients would have been sufficient to detect a target effect at the planned power and significance level, the clustered design needs 400 × 3.4 = 1,360 patients recruited through the cluster structure to achieve the same precision — more than three times as many respondents, purely because of clustering. Required sample size under cluster sampling is calculated by multiplying the simple-random-sample requirement by DEFF, then working backward to how many clusters and how many patients per cluster are needed to reach that total.
Critically, adding more individuals within already-selected clusters helps far less than adding more clusters: because within-cluster observations are correlated, a 1,360-patient sample spread across 5 clusters delivers much worse precision than the same 1,360 patients spread across 54 clusters, even though the total sample size is identical. The number of clusters, not the number of individuals, is usually the binding constraint on precision. For the full mechanics of ICC, its interpretation benchmarks, and the design effect specifically in the context of cluster-randomized trials, see CASRAI’s dedicated guides on the intraclass correlation coefficient (ICC) and cluster randomized trials and the design effect — this page introduces the concept; those guides go deeper on estimation and CONSORT reporting requirements for randomized designs specifically.
Analysis Must Account for Clustering
Treating clustered data as if every observation were independent — running an ordinary regression or t-test as though the sample were a simple random sample — understates standard errors and inflates the Type I error rate, because the effective sample size is smaller than the raw observation count once clustering is accounted for. Three standard approaches correct for this:
- Mixed-effects (multilevel/hierarchical) models — explicitly model a random intercept (and optionally random slopes) for cluster, partitioning variance into within-cluster and between-cluster components directly.
- Generalized estimating equations (GEE) with cluster-robust standard errors — model the population-average effect while using a robust (“sandwich”) variance estimator clustered on the grouping variable, which corrects standard errors without requiring a fully specified random-effects structure.
- Survey-weighted estimators — software built for complex survey designs (R’s
survey/srvyrpackages, Stata’ssvycommand suite, SASPROC SURVEYMEANS/SURVEYREG, SPSS Complex Samples) take the cluster (and stratum, and selection-probability weight) structure as inputs and produce correctly adjusted point estimates and standard errors.
Which approach is appropriate depends on the research question: mixed-effects models are typically preferred when the between-cluster variance itself is of substantive interest; GEE and survey-weighted estimators are typically preferred when the goal is a population-average estimate with correctly sized standard errors. In every case, the clustering variable itself — not just the outcome and predictors — must be supplied to the software; omitting it silently produces the standard-error understatement described above.
When Cluster Sampling Is the Wrong Choice
- A small number of clusters. Precision in a cluster design depends heavily on the number of clusters sampled, not just total observations. With too few clusters (commonly cited rules of thumb start around 20–30, though the right number depends on the ICC and effect size), between-cluster variance is estimated too imprecisely for cluster-robust standard errors or mixed-effects models to be reliable.
- High ICC. When clusters are strongly internally homogeneous, each additional individual sampled within an already-selected cluster adds very little new information, since they closely resemble others already in the sample. High ICC drives DEFF up sharply, and the cost savings from clustering can be outweighed by the much larger total sample size needed to compensate.
- When precision matters more than cost. If a study has a hard requirement on the precision of its estimate (a regulatory threshold, a pre-registered minimum detectable effect) and the budget genuinely allows it, simple random sampling or stratified sampling will reach that precision with a smaller total sample — cluster sampling should not be the default just because it is cheaper to execute.
See CASRAI’s comparisons of random sampling vs. convenience sampling and purposive sampling vs. convenience sampling, and the simple random sampling and snowball sampling guides, for the other methods in the same decision space. For determining how many total observations any of these designs need before accounting for clustering, see CASRAI’s guide to power analysis and sample size calculation.
Reporting Cluster Sampling in a Methods Section
A Methods section using cluster sampling should state, at minimum:
- What the clusters are (clinics, schools, villages, households) and how the cluster-level sampling frame was constructed.
- How many stages the design used (one-stage, two-stage, or multistage) and what was selected at each stage.
- Selection probabilities at each stage — equal-probability or probability-proportional-to-size (PPS), and how any resulting sampling weights were calculated and applied at analysis.
- The number of clusters sampled and the average (and range of) cluster sample sizes — not just the total individual sample size.
- The observed (or assumed, if planned in advance) ICC and the resulting design effect, and how the analysis (mixed-effects model, GEE with robust standard errors, or survey-weighted estimator) accounted for clustering.
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. For cluster-randomized designs specifically, CASRAI’s cluster randomized trials guide covers the additional CONSORT cluster-extension reporting requirements.
Frequently Asked Questions
What is cluster sampling?
Cluster sampling is a probability sampling method that divides a population into naturally occurring groups (clusters), randomly selects a subset of those clusters, and then samples everyone within selected clusters (one-stage) or a sub-sample within them (two-stage/multistage), rather than sampling individuals directly from the whole population.
What’s the difference between cluster sampling and stratified sampling?
Stratified sampling divides the population into homogeneous strata and samples from every stratum, which improves precision. Cluster sampling divides the population into (ideally heterogeneous) clusters and samples only some of them, which increases variance but avoids needing a full population-level sampling frame.
What is the design effect in cluster sampling?
The design effect (DEFF) measures how much a clustered sample’s precision is reduced compared to a simple random sample of the same size, calculated as DEFF = 1 + (m − 1) × ICC, where m is the average cluster sample size and ICC is the intraclass correlation coefficient.
How many clusters do I need for cluster sampling?
There’s no universal minimum, but precision depends more on the number of clusters than the number of individuals; commonly cited practical thresholds start around 20–30 clusters, with the exact requirement driven by the ICC, the desired precision, and the planned analysis method.
What’s the difference between one-stage and two-stage cluster sampling?
One-stage cluster sampling includes every member of each randomly selected cluster. Two-stage cluster sampling randomly selects clusters first and then draws a further random sub-sample from within each selected cluster, giving the researcher direct control over final sample size.
How do you analyze cluster-sampled data correctly?
Clustered data should be analyzed with methods that account for within-cluster correlation — mixed-effects (multilevel) models, generalized estimating equations (GEE) with cluster-robust standard errors, or survey-weighted estimators in software built for complex survey designs. Ignoring clustering understates standard errors and inflates the Type I error rate.
