When Is Cluster Sampling Appropriate?

Should You Randomly Select Groups Instead of Individual Participants?

Imagine conducting a nationwide study of teachers' digital teaching practices. You want to draw conclusions about teachers across the country, but there are thousands of schools and no readily available national list containing every eligible teacher.

Even if such a list existed, randomly selecting individual teachers could require visiting schools scattered across distant geographical areas. Travel expenses, institutional permissions, and coordination might make the study impractical.

One alternative is to randomly select schools and collect data from teachers within those selected schools. This is the central idea behind cluster sampling.

The challenge is that teachers within the same school may share institutional conditions and therefore provide more similar responses than teachers selected independently across schools. Cluster sampling may reduce data collection costs while changing the statistical information provided by the sample.

Determining whether the design is appropriate requires evaluating both its operational advantages and its inferential consequences.

Use Cluster Sampling When Sampling Groups Makes Population Research Feasible

What Makes Cluster Sampling Appropriate?

What Exactly Is Cluster Sampling?

Cluster sampling is a probability sampling design in which population units are organized into groups called clusters, and a probability sample of those groups is selected.

Clusters may be schools, classrooms, villages, census enumeration areas, hospitals, workplaces, or households, depending on the population and research objectives.

For example, if a researcher wants to study individual teachers, schools may serve as clusters. If the target population consists of households, geographical enumeration areas may serve as clusters.

The cluster is the unit selected at the relevant sampling stage. The ultimate unit of analysis may be different. A study may select schools as clusters while measuring individual teachers' experiences.

This distinction matters because the number of selected clusters and the number of observed individuals contribute differently to statistical precision.

Why Would Researchers Sample Clusters Instead of Individuals?

The main advantage is often operational rather than statistical.

Consider a population of 20,000 teachers distributed across 1,000 schools. A simple random sample of individual teachers could require contacting hundreds of separate schools.

Alternatively, selecting a smaller number of schools and surveying teachers within them may substantially reduce travel, coordination, and administrative costs.

Cluster sampling may therefore be appropriate when:

• Population members are naturally grouped into institutions, communities, or geographical areas.
• Contacting scattered individuals is considerably more expensive than contacting groups.
• A usable list of clusters exists, even when an individual-level population frame is unavailable.
• Institutional or geographical access procedures favor group-based recruitment.
• The research budget and timeline make unrestricted individual sampling impractical.

These advantages do not mean that cluster sampling always produces equally precise estimates at a lower cost. The design often trades some statistical efficiency for operational feasibility.

What Is the Difference Between One-Stage and Two-Stage Cluster Sampling?

Cluster sampling can involve different procedures after clusters are selected.

Two-stage cluster sampling is a form of multistage sampling. More elaborate designs may select regions, then schools, then classrooms, and finally students.

When several selection stages are required, the researcher should explicitly justify the use of multistage sampling and calculate selection probabilities across the stages.

What Conditions Must the Clusters Satisfy?

Clusters should be defined in a way that supports valid probability selection and appropriate population coverage.

For conventional cluster sampling, the researcher should establish whether clusters collectively cover the intended population and whether individual units belong to one or more clusters.

For example, schools may form a suitable cluster frame for studying teachers if every eligible teacher can be assigned consistently to a school and the school registry adequately covers the intended population.

Complications arise when teachers work at multiple schools, institutions have recently closed, or certain school categories are excluded from the registry.

These problems may require explicit assignment rules, adjustments for multiple frame memberships, or reconstruction of the frame.

A cluster frame should also contain reliable information about cluster size when the selection method or weighting procedure depends on it.

Why Does Similarity Within Clusters Matter?

Individuals belonging to the same cluster often share contextual influences.

Teachers in one school may experience the same leadership policies, infrastructure, professional development programs, and institutional expectations. Students in the same classroom may share instructional conditions and peer environments.

As a result, their responses may be correlated.

This similarity is commonly described using the intracluster correlation coefficient, or ICC. In a simple random-intercept model, the ICC represents the proportion of total outcome variance attributable to differences between clusters.

For many educational and organizational outcomes, positive intracluster correlation means that observing additional individuals within the same cluster provides less new information than observing an equivalent number of independently sampled individuals.

Cluster sampling can therefore produce less precise estimates than an equally sized simple random sample, particularly when clusters are large and within-cluster similarity is substantial.

However, the magnitude of the effect varies by outcome, population, cluster structure, weighting, and sampling design. It should not be assumed to be identical across all variables.

What Is the Design Effect in Cluster Sampling?

The design effect compares the sampling variance under a particular design with the variance under a specified simple random sampling benchmark.

For an elementary equal-sized cluster model with approximately equal weights, a common approximation is:

The formula is an approximation, not a universal expression for every cluster design. Unequal cluster sizes, unequal selection probabilities, stratification, finite population corrections, and other design features may change the actual design effect.

The implications for required sample size are examined more fully through the effect of clustering on sample size and the use of design effects in sample size calculations.

Does Selecting More People Within Each Cluster Improve Precision?

Usually, adding observations provides some information. However, when intracluster correlation is positive, the gain may be smaller than adding observations from additional independent clusters.

Suppose a study can recruit 600 teachers. One option is to select 10 schools and survey 60 teachers per school. Another is to select 30 schools and survey 20 teachers per school.

Both designs produce 600 observations, but they do not necessarily have the same precision.

Under a simplified equal-cluster-size model with an ICC of 0.05, the first design has an approximate clustering design effect of 3.95, while the second has an approximate design effect of 1.95.

The second design would generally provide more statistical information under those assumptions, although it may involve higher travel and administrative costs.

Researchers should therefore consider both the number of clusters and the number of individuals within clusters when determining sample size.

The optimal balance depends on the ICC, cluster sizes, fixed costs of accessing each cluster, costs of collecting data within clusters, and the intended analysis.

How Do Unequal Cluster Sizes Affect Selection Probabilities?

Clusters frequently differ in size. A school may have 15 teachers while another has 150. A village may contain hundreds of households while another contains thousands.

If clusters are selected with equal probability and every eligible individual in each selected cluster is included, individuals have equal inclusion probabilities when each belongs to exactly one cluster.

However, if a fixed number of individuals is sampled within each selected cluster, individuals in larger clusters may have lower overall inclusion probabilities than those in smaller clusters.

One alternative is probability proportional to size sampling, commonly called PPS sampling. Under PPS, clusters with larger measures of size receive higher selection probabilities.

When PPS cluster selection is combined with an appropriate fixed-size within-cluster sampling procedure, the overall design may produce approximately equal individual inclusion probabilities, depending on the exact implementation and accuracy of the size measures.

The design must be specified carefully. PPS does not automatically produce a self-weighting sample under every arrangement.

Where inclusion probabilities differ, researchers need to account for unequal selection probabilities and sampling weights when estimating population characteristics.

How Does Cluster Sampling Differ from Stratified Sampling?

Cluster sampling and stratified sampling both divide the population into groups, but the purpose and selection mechanism differ.

In stratified sampling, all defined strata generally contribute to the sample. In cluster sampling, only selected clusters contribute observations.

Stratification often seeks to reduce sampling variance by grouping relatively similar units into strata and sampling within each one. Cluster sampling commonly reduces fieldwork costs by selecting groups, but may increase sampling variance when units within those groups are similar.

The two approaches are not mutually exclusive. A national education survey may stratify schools by region and sector, select schools within each stratum, and then sample teachers within selected schools.

When Is Cluster Sampling Less Appropriate?

Cluster sampling may offer limited advantages when a complete individual-level frame is available and population units can be contacted at low cost.

It may also be problematic when very few clusters can be selected, cluster sizes vary substantially without an appropriate design response, or the outcome exhibits strong intracluster correlation.

For example, a study selecting only three schools to estimate national teacher attitudes would face serious limitations even if it surveyed hundreds of teachers within those schools.

The large number of teachers would not compensate fully for the limited number of independently selected schools.

Cluster sampling also requires an analysis plan that recognizes the sampling structure. Treating all observations as independent can substantially understate standard errors.

Where the study is intended to estimate characteristics of individual teachers rather than schools, researchers must also ensure that the selection and weighting procedures support inference to the teacher population rather than merely describing the selected institutions.

Sampling Schools for a Study of Teachers' Digital Competence

This is a two-stage cluster sampling design. Schools are selected at the first stage, and teachers are selected at the second stage.

The researcher then considers the expected statistical precision. Suppose an ICC of 0.05 is used as a planning assumption based on relevant prior evidence or a justified sensitivity scenario.

With 20 teachers per school, the approximate clustering design effect is:

The researcher recognizes that selecting 30 schools rather than 10 schools may improve precision by reducing the number of teachers concentrated within each cluster. However, the larger number of schools may increase coordination costs.

The final design is therefore evaluated against the available budget, expected ICC, anticipated response rates, and desired precision. The analysis will account for school-level clustering rather than treating the 600 teachers as independently sampled observations.

The example illustrates why cluster sampling is not simply a convenient recruitment arrangement. It is a probability design whose statistical properties must be planned and documented.

Common Misunderstandings About Cluster Sampling

How to Decide Whether Cluster Sampling Is Worth the Trade-Off

Cluster sampling is often a practical solution to a real fieldwork problem. The key is to establish that the operational savings justify the design's additional statistical and analytical requirements.

When documenting the methodology, identify the ultimate population units, the clusters, the sampling stages, the frame used at each stage, and the probability selection procedures.

Report the number of clusters selected and the intended number of observations within clusters. Where relevant, explain how unequal cluster sizes, nonresponse, sampling weights, and intracluster correlation were considered.

For analysis, use methods appropriate to the actual design. If the survey involves stratification, clustering, and unequal weights, these features should be represented in the survey-design specification wherever required for valid estimation.

The distinction between selected clusters and participating clusters also matters. Institutional refusal or inaccessibility can introduce nonresponse problems even when the original cluster selection was random.

Before Using Cluster Sampling, Verify These Design Requirements

Frequently Asked Questions About Cluster Sampling

Use Cluster Sampling When Group-Based Selection Makes Research Feasible

How to Cite This Guide

This guide is intended to be read, shared, and used in research, teaching, and academic work. If you draw on its ideas, explanations, or other content, please acknowledge the source by citing the guide. Doing so gives appropriate credit and helps your readers locate the original resource.