Cluster sampling is a probability sampling method in which a population is divided into non-overlapping groups, called clusters, and a random sample of those clusters is selected. Researchers then study every unit in the chosen clusters or randomly sample units within them. It is especially useful for large, geographically dispersed populations.
Cluster sampling allows researchers to collect data from a large population without creating a complete list of every individual or travelling to every location. Instead, the researcher works with naturally occurring or administratively defined groups, such as schools, hospitals, neighbourhoods, villages, workplaces or census areas.
This article explains how cluster sampling works, its main types, how it differs from stratified sampling, how clustering affects sample size and precision, and how cluster-sampled data should be analysed and reported.
Key Takeaways
- Cluster sampling selects groups first rather than selecting individuals directly from the entire population.
- In one-stage sampling, every unit in a selected cluster is studied; in two-stage sampling, units are randomly sampled within selected clusters.
- Cluster sampling usually reduces travel, listing and administrative costs.
- Units in the same cluster may be similar, increasing sampling variance and reducing the effective sample size.
- Unequal cluster sizes should be addressed through an appropriate selection method, allocation and sampling weights.
- Analyses should account for cluster identifiers, strata and survey weights when these are part of the design.
What Is Cluster Sampling?
Cluster sampling is a probability-sampling technique in which the target population is divided into distinct groups called clusters. A random sample of clusters is selected, after which researchers collect data from all or some of the units inside the selected clusters.
For example, suppose a researcher wants to study digital-learning access among secondary-school students throughout a country. Obtaining a national list of every student may be impractical. The researcher could instead use schools as clusters, randomly select a sample of schools and then survey either:
- every eligible student in each selected school; or
- a random sample of students within each selected school.
The first design is one-stage cluster sampling. The second is two-stage cluster sampling.
Cluster sampling is a probability method only when selection is based on a genuine probability mechanism. Choosing the nearest schools, the most cooperative hospitals or the easiest villages to visit is convenience or purposive selection, not probability cluster sampling.
Key Terms in Cluster Sampling
Population
The population is the full group about which the researcher wants to draw conclusions.
Examples include:
- all public-school teachers in a state;
- all households in a district;
- all registered patients attending primary-care facilities;
- all employees of a multinational organisation.
Cluster
A cluster is a non-overlapping group of population units used as a sampling unit.
Clusters may be:
- geographic, such as census blocks, villages or districts;
- organisational, such as schools, clinics, factories or branches;
- administrative, such as classrooms, departments or electoral areas;
- time-based, such as work shifts or scheduled service sessions.
Primary sampling unit
The primary sampling unit, commonly abbreviated as PSU, is the unit selected at the first sampling stage.
If schools are selected first and students second, schools are the PSUs.
Secondary sampling unit
The secondary sampling unit, or SSU, is selected within a sampled PSU.
If schools are PSUs and classrooms are sampled within schools, classrooms are SSUs. Students may then form a third-stage sampling unit.
Sampling frame
A sampling frame is the operational list or map from which units are selected.
A cluster-sampling design may require:
- a complete list of first-stage clusters;
- a measure of each cluster’s size;
- an updated list of units within each selected cluster.
A major practical advantage is that a researcher may need a complete individual-level list only for the selected clusters rather than for the entire population.
Inclusion probability
An inclusion probability is the probability that a particular unit will be included in the sample.
In a multistage design, an individual’s final inclusion probability is generally based on the probability of selection at every relevant stage.
For example:
[
\pi_{ij} = \pi_i \times \pi_{j|i}
]
where:
- (\pi_i) is the probability that cluster (i) is selected; and
- (\pi_{j|i}) is the probability that individual (j) is selected, given that cluster (i) was selected.
The basic design weight is often the inverse of this probability:
[
w_{ij} = \frac{1}{\pi_{ij}}
]
Additional adjustments may be made for nonresponse, calibration or post-stratification.
How Does Cluster Sampling Work?
Cluster sampling works by selecting progressively smaller units from a structured population.
The basic process is:
- Define the population.
- Identify non-overlapping clusters.
- create or obtain a cluster-level sampling frame.
- Randomly select clusters.
- Survey everyone in those clusters or randomly select units within them.
- Calculate selection probabilities and weights where required.
- Analyse the data using methods that recognise the sampling design.
The essential feature is that only a subset of clusters is observed. If researchers select individuals from every group, the design is more likely to be stratified sampling than cluster sampling.
Types of Cluster Sampling
One-Stage Cluster Sampling
In one-stage cluster sampling, researchers randomly select clusters and collect data from every eligible unit within each selected cluster.
Suppose a university has 120 tutorial groups and wants to survey student satisfaction. It randomly selects 15 tutorial groups and invites every student in those groups to participate.
The tutorial group is the PSU, and all students within selected groups are included.
One-stage sampling is practical when:
- clusters contain a manageable number of units;
- complete enumeration within a cluster is inexpensive;
- a reliable list of individual units is unavailable before cluster selection;
- the cost of visiting a cluster is high relative to the cost of interviewing additional units within it.
Its main statistical disadvantage is that collecting many observations from a small number of highly similar clusters may provide less information than collecting fewer observations from more clusters.
Two-Stage Cluster Sampling
In two-stage cluster sampling, researchers randomly select clusters first and then randomly select units within each chosen cluster.
For example:
- Randomly select 30 schools from all schools in a region.
- Randomly select 25 students within each selected school.
The first-stage units are schools, and the second-stage units are students.
Two-stage sampling is often preferred when selected clusters are too large to survey completely. It gives the researcher greater control over total sample size and fieldwork.
Multistage Cluster Sampling
Multistage cluster sampling uses three or more probability-selection stages to move from large clusters to the final units of observation.
A national household survey might:
- stratify the country by region and urban–rural status;
- select enumeration areas within each stratum;
- select neighbourhood segments within enumeration areas;
- select households within segments;
- select one eligible adult within each household.
Multistage sampling is common in national surveys because no single complete list of every person may exist. However, each additional stage introduces design, weighting and documentation requirements.
One-Stage, Two-Stage and Multistage Sampling Compared
| Feature | One-stage | Two-stage | Multistage |
|---|---|---|---|
| First step | Select clusters | Select clusters | Select large clusters |
| Later selection | None | Select units within clusters | Select progressively smaller units |
| Units surveyed | Every eligible unit in chosen clusters | A random subset within each chosen cluster | A final sample after three or more stages |
| Main advantage | Simple field procedure | Controls workload and sample size | Practical for very large populations |
| Main limitation | May collect many correlated observations | Requires a within-cluster frame | More complex weighting and analysis |
| Example | Survey all students in selected schools | Survey sampled students in selected schools | Select regions, schools, classes and students |
How to Conduct Cluster Sampling: Seven Steps
Step 1: Define the Target Population
Specify exactly who or what the findings should represent.
A complete definition may include:
- geographic boundaries;
- age or eligibility rules;
- institutional categories;
- study period;
- exclusions.
Example:
All students aged 12–18 enrolled in government secondary schools in the state during the 2026 academic year.
Step 2: Identify Suitable Clusters
Choose clusters that:
- cover the target population;
- do not overlap;
- can be clearly identified;
- have known or estimable sizes;
- can be selected using a probability method.
Natural administrative groups are often preferable because lists and contact arrangements already exist.
Step 3: Build the Cluster Sampling Frame
The first-stage frame should list all eligible clusters.
Useful fields include:
| Field | Example |
|---|---|
| Cluster ID | SCH-001 |
| Cluster name | North Valley Secondary School |
| Region | Western District |
| Urban/rural status | Rural |
| Estimated number of eligible units | 840 students |
| Contact or access information | District education office |
| Selection stratum | West–Rural |
Outdated frames can create undercoverage if new clusters are omitted or closed clusters remain listed.
Step 4: Decide the Number of Stages
Choose between:
- one-stage enumeration;
- two-stage sampling;
- a multistage design.
The decision should reflect cost, desired precision, cluster size, available frames and the feasibility of listing units within selected clusters.
Step 5: Select Clusters Randomly
Clusters may be selected through:
- simple random sampling;
- systematic random sampling;
- stratified random sampling;
- probability proportional to size.
The randomisation process should be recorded, including software, seed, ordering method and replacement rules.
Step 6: Select Units Within Clusters
For a two-stage design, create or update the within-cluster list and use a probability method such as:
- simple random sampling;
- systematic sampling;
- PPS sampling at an additional stage.
Do not permit fieldworkers to choose the easiest respondents. Convenience selection within randomly chosen clusters can still produce biased results.
Step 7: Document and Analyse the Design
Record:
- cluster and unit identifiers;
- strata;
- probabilities at each stage;
- sampling weights;
- cluster-size measures;
- nonresponse outcomes;
- substitutions or deviations;
- finite population information, when relevant.
These variables should be retained in the analytical dataset.
Detailed Cluster-Sampling Example
A researcher wants to estimate the proportion of secondary-school teachers in a state who use generative AI tools for lesson planning.
Population
All 18,600 eligible secondary-school teachers in the state.
Available frame
The education authority has a list of 310 schools and the number of eligible teachers at each school. It does not maintain a single updated list of all individual teachers.
Proposed design
The researcher uses schools as clusters.
- Divide schools into urban and rural strata.
- Select schools within each stratum with probability proportional to the number of teachers.
- Obtain an updated teacher list from every selected school.
- Randomly select 20 teachers from each school.
- Calculate the probability that each teacher was selected.
- Apply survey weights and account for schools as PSUs during analysis.
This is a stratified two-stage cluster sample.
It combines:
- stratification, to ensure urban and rural coverage;
- clustering, to reduce listing and travel costs;
- PPS selection, to handle different school sizes;
- within-school random sampling, to limit workload.
Cluster Sampling and Unequal Cluster Sizes
Clusters in real populations are rarely equal in size. Schools may have 30 or 200 teachers, villages may contain 100 or 2,000 households, and hospitals may serve very different numbers of patients.
Unequal size does not automatically invalidate cluster sampling. The important questions are:
- How are clusters selected?
- How many units are sampled from each selected cluster?
- What is each final unit’s inclusion probability?
- Are appropriate weights used?
Equal-Probability Cluster Selection
Suppose every school has the same probability of first-stage selection and every teacher in a selected school is surveyed.
Each teacher’s first-stage inclusion probability is the probability that their school is selected. If that probability is equal across schools, all teachers may have the same inclusion probability at this stage even though schools differ in size.
Equal Numbers Sampled from Unequal Clusters
Now suppose every school has the same selection probability, but exactly 20 teachers are sampled from each selected school.
A teacher in a school with 40 teachers has a one-half within-school selection probability. A teacher in a school with 200 teachers has a one-tenth probability.
The final probabilities differ, so weights are required.
Probability Proportional to Size
In PPS sampling, larger clusters have a greater probability of being selected.
If clusters are selected proportional to size and the same number of final units is sampled from each selected cluster, the two stages can sometimes produce an approximately self-weighting design.
For instance:
- large schools are more likely to be selected;
- each selected school contributes the same number of teachers;
- the higher school-selection probability can offset the lower within-school sampling probability in large schools.
The exact probabilities must still be calculated and verified.
Cluster Sampling Formula: Design Effect
Cluster sampling often produces correlated observations because people within the same cluster share environments, institutions, services or social conditions.
Students in one school may experience the same curriculum and leadership. Patients in one hospital may receive similar procedures. Households in one neighbourhood may share infrastructure and economic conditions.
A commonly used approximation for the cluster design effect is:
[
DEFF \approx 1 + (m-1)\rho
]
where:
- (DEFF) is the design effect;
- (m) is the average number of sampled units per cluster;
- (\rho) is the intracluster correlation coefficient, or ICC.
This approximation assumes a relatively simple design and roughly equal cluster sizes. More complex formulas or simulation may be needed when cluster sizes vary substantially or the design contains multiple stages, unequal probabilities or stratification.
Interpreting the Design Effect
- (DEFF = 1): approximately the same variance as simple random sampling.
- (DEFF > 1): less statistical precision than an equal-sized simple random sample.
- (DEFF < 1): greater precision, which can occasionally arise from stratification or other design features.
Worked Design-Effect Example
Assume:
- average completed sample per cluster: (m = 25);
- expected ICC: (\rho = 0.04).
Then:
[
DEFF = 1 + (25-1)(0.04)
]
[
DEFF = 1 + 24(0.04)
]
[
DEFF = 1.96
]
The variance is therefore estimated to be approximately 1.96 times the variance expected under a comparable simple random sample.
Effective Sample Size
The effective sample size estimates how many independent simple-random-sample observations would provide similar precision.
[
n_{\text{effective}} \approx \frac{n}{DEFF}
]
If 800 people are sampled and the design effect is 1.96:
[
n_{\text{effective}} \approx \frac{800}{1.96} \approx 408
]
Although the dataset contains 800 respondents, its precision may resemble a simple random sample of approximately 408 independent respondents.
This does not mean the remaining responses are useless. It means that respondents within the same cluster provide partially overlapping statistical information.
Adjusting Sample Size for Cluster Sampling
A simple planning approximation is:
[
n_{\text{cluster}} \approx n_{\text{SRS}} \times DEFF
]
Suppose a simple-random-sample calculation indicates that 400 completed responses are needed.
With (DEFF = 1.96):
[
n_{\text{cluster}} = 400 \times 1.96 = 784
]
The cluster design would therefore require approximately 784 completed responses to target similar precision.
If 25 completed respondents will be obtained from each cluster:
[
\text{Number of clusters} =
\frac{784}{25}
= 31.36
]
The researcher would round upward and select at least 32 clusters.
If the expected response rate is 80%, the number of eligible units approached may be estimated as:
[
n_{\text{approached}} =
\frac{784}{0.80}
= 980
]
This calculation is only a starting point. Formal sample-size planning should consider:
- the main outcome and estimator;
- expected prevalence or effect size;
- confidence level;
- statistical power;
- cluster-size variation;
- expected ICC;
- finite population corrections;
- stratification;
- nonresponse;
- subgroup estimates;
- the number of available clusters;
- practical cost at each sampling stage.
Why the Number of Clusters Matters
For many cluster designs, sampling more clusters with fewer units per cluster produces better precision than sampling many units from only a few clusters.
Consider two designs with 1,000 participants:
- Design A: 10 clusters × 100 people;
- Design B: 50 clusters × 20 people.
When people within clusters are positively correlated, Design B will often contain more independent information because it captures variation across more clusters.
A large participant count cannot fully compensate for an extremely small number of clusters. This is particularly important when:
- estimating cluster-level variance;
- calculating cluster-robust standard errors;
- comparing cluster characteristics;
- conducting cluster-randomised experiments.
The optimal allocation depends on the cost of adding a cluster, the cost of sampling another person within an existing cluster, cluster correlation and the research objective.
Cluster Sampling vs. Stratified Sampling
Cluster sampling selects some groups and leaves other groups unobserved. Stratified sampling selects units from every stratum.
| Feature | Cluster sampling | Stratified sampling |
|---|---|---|
| Main purpose | Reduce logistical and fieldwork cost | Improve precision or ensure subgroup representation |
| Groups | Clusters | Strata |
| Groups included | A random subset of clusters | Normally every stratum |
| Selection within groups | Everyone or a sample from selected clusters | A sample from each stratum |
| Ideal group pattern | Clusters are similar to one another and internally varied | Members within a stratum are relatively similar |
| Typical examples | Schools, villages, clinics, census blocks | Age groups, regions, degree levels, income categories |
| Main statistical concern | Within-cluster correlation | Correct allocation and weighting across strata |
| Can the methods be combined? | Yes | Yes |
Example of Cluster Sampling
Randomly select 20 schools and survey students only in those schools.
Example of Stratified Sampling
Divide all students into undergraduate, master’s and doctoral strata, then randomly sample students from each stratum.
Important Qualification
Real clusters do not need to be perfect miniature versions of the population for a probability sample to be valid.
Validity depends more fundamentally on:
- adequate frame coverage;
- known selection probabilities;
- genuine random selection;
- appropriate weighting;
- correct variance estimation;
- a sufficiently informative number of clusters.
Having internally diverse clusters may improve efficiency, but it is not a substitute for a valid probability design.
Cluster Sampling Compared with Other Sampling Methods
| Method | What is selected? | Does every group have to be sampled? | Main advantage | Typical limitation |
|---|---|---|---|---|
| Simple random sampling | Individuals directly | Not applicable | Straightforward inference | Requires a complete individual frame |
| Systematic sampling | Every (k)th unit after a random start | Not applicable | Easy to implement on ordered lists | Can be affected by periodic patterns |
| Stratified sampling | Individuals within every stratum | Yes | Ensures subgroup representation and may improve precision | Requires stratum information |
| Cluster sampling | Some groups, then all or some units inside them | No | Reduces travel, listing and administrative costs | Correlation may reduce precision |
| Multistage sampling | Progressively smaller units | Only selected units at each stage | Practical for national and hierarchical populations | Complex probabilities, weights and analysis |
When Should Cluster Sampling Be Used?
Cluster sampling is appropriate when:
- the population is geographically dispersed;
- a complete list of individuals is unavailable or expensive to create;
- a reliable list of clusters exists;
- data collection requires travel or institutional access;
- population units naturally belong to schools, hospitals, households, villages or organisations;
- the cost of opening a new cluster is high;
- the researcher can calculate selection probabilities;
- enough clusters can be selected to support the intended analysis.
Cluster sampling may be inappropriate when:
- only a few highly unusual clusters are available;
- cluster selection cannot be randomised;
- important population groups are concentrated in clusters with little chance of selection;
- the research requires precise estimates for every cluster;
- there are too few clusters for valid inference;
- a complete individual-level frame already exists and simple or stratified sampling is affordable.
Advantages of Cluster Sampling
Lower Data-Collection Cost
Researchers can concentrate fieldwork in selected locations rather than travelling throughout the entire study area.
Reduced Frame-Construction Burden
Only a list of clusters may be required initially. Detailed household or individual lists can be prepared after clusters are selected.
Administrative Convenience
Schools, clinics and organisations can coordinate access for groups of respondents.
Suitability for Large-Scale Surveys
Cluster and multistage designs make regional, national and international surveys operationally possible.
Flexible Combination with Other Methods
Clusters can be stratified before selection. PPS, systematic sampling and within-cluster random sampling can also be incorporated.
Practical Use of Digital Mapping
Geographic information systems can support the definition, measurement, selection and mapping of area-based clusters.
Limitations of Cluster Sampling
Reduced Precision
People or units within the same cluster are often similar. Positive intracluster correlation increases variance compared with a simple random sample of the same nominal size.
More Complex Sample-Size Planning
The researcher needs an anticipated design effect or ICC, not only a conventional simple-random-sample calculation.
More Complex Analysis
Naive analysis may produce standard errors and confidence intervals that are too small.
Unequal Selection Probabilities
Different cluster sizes and sampling fractions can create unequal probabilities that must be reflected in weights.
Dependence on the Sampling Frame
Missing clusters, outdated size measures or boundary errors can cause undercoverage or incorrect PPS probabilities.
Nonresponse at Multiple Levels
A selected institution may refuse participation, and individuals within participating institutions may also fail to respond.
Risk from Too Few Clusters
A small number of clusters limits the amount of independent cluster-level information, regardless of the total number of respondents.
Sampling Error, Bias and Representativeness
Sampling error and bias are not the same.
Sampling Error
Sampling error is the random difference between an estimate from a sample and the value that would be obtained from the full population under the same measurement conditions.
Cluster sampling often increases sampling error because units within clusters may be correlated.
Bias
Bias is a systematic tendency for the method to produce estimates that differ from the target population value.
Bias can arise from:
- incomplete frames;
- nonrandom cluster selection;
- convenience selection within clusters;
- unrecorded substitutions;
- nonresponse;
- incorrect weights;
- measurement error;
- excluding inaccessible areas without redefining the target population.
Cluster sampling is not inherently biased. A properly selected and analysed cluster sample can produce unbiased or design-consistent estimates even when clusters differ substantially.
Representativeness
Representativeness should not be judged only by whether a selected cluster appears visually similar to the entire population.
A stronger assessment considers whether:
- the target population is adequately covered;
- the selection mechanism is probabilistic;
- all relevant units have known, non-zero selection probabilities;
- differential probabilities and response patterns are adjusted;
- the sample contains enough clusters;
- estimates and uncertainty are calculated for the actual design.
Real-World Uses of Cluster Sampling
Public-Health Surveys
Public-health agencies use cluster designs for household needs assessments, vaccination-coverage studies and population-health surveys.
For example, an emergency assessment may select geographic clusters first and households within clusters second.
Demographic and Household Surveys
Large demographic surveys often use stratified two-stage designs:
- enumeration areas are selected;
- households are selected within each area.
Educational Research
Researchers may sample:
- school districts;
- schools within districts;
- classes within schools;
- students within classes.
Health-Service Research
Hospitals or clinics may be sampled first, followed by patients, clinicians or records.
Organisational Research
A company may randomly select branches, factories or departments and then survey employees within those locations.
Agricultural and Environmental Research
Researchers may select geographic areas, farms, fields and plots at successive stages.
Market Research
Stores, service locations or sales territories can act as clusters when a population is widely dispersed.
How to Analyse Cluster-Sampled Data
Correct analysis begins with the sample design, not merely the number of completed responses.
Identify the Design Variables
The analytical dataset should contain, where applicable:
- PSU identifier;
- lower-stage cluster identifiers;
- stratum identifier;
- final survey weight;
- replicate weights;
- finite population information;
- response and eligibility indicators.
Use Survey-Analysis Procedures
Common options include:
- R’s
surveypackage; - Stata’s
svysetandsvy:commands; - SAS
SURVEYMEANS,SURVEYFREQ,SURVEYREG,SURVEYLOGISTICand related procedures; - IBM SPSS Complex Samples.
These tools can incorporate clustering, weights and stratification when estimating means, proportions, regression coefficients, standard errors and confidence intervals.
Do Not Apply Weights Without Adjusting Variance
Applying a weight in a standard statistical procedure may change point estimates but still leave standard errors incorrect.
The analysis should account for both:
- unequal probabilities through weights; and
- dependence through cluster-aware variance estimation.
Cluster-Robust Standard Errors Are Not Always Sufficient
Cluster-robust standard errors can address some within-cluster dependence, but they do not automatically reproduce a complete complex-survey analysis involving:
- stratification;
- multiple sampling stages;
- probability weights;
- finite population corrections;
- replicate-weight systems;
- calibrated survey estimators.
The analytical method should match the inferential framework and supplied survey documentation.
Distinguish Population Inference from Multilevel Modelling
Complex-survey analysis and multilevel modelling answer related but different questions.
- Survey methods focus on inference under the sampling design.
- Multilevel models estimate relationships and variation at several hierarchical levels.
- Some studies require both survey weights and multilevel structure.
Researchers should not assume that adding a random intercept automatically resolves unequal selection probabilities.
Digital Tools, GIS and Artificial Intelligence
Geographic Information Systems
GIS software can help researchers:
- define non-overlapping area clusters;
- attach population-size measures;
- identify inaccessible areas;
- select clusters;
- produce field maps;
- monitor geographic coverage.
The rules for boundaries and exclusions should be established before selection.
Random-Selection Software
Researchers can use R, Python, SAS, Stata or audited survey-sampling software to:
- generate random numbers;
- select clusters;
- perform PPS sampling;
- document random seeds;
- calculate probabilities and weights;
- reproduce the selected sample.
Survey Platforms
Online survey platforms can administer questionnaires within selected clusters, but the platform does not itself create a probability sample. Probability depends on how clusters and participants are selected.
Appropriate Uses of AI
Artificial intelligence can assist with:
- drafting sampling documentation;
- checking code syntax;
- creating fieldworker instructions;
- translating questionnaires;
- generating validation rules;
- identifying missing design variables;
- explaining output.
AI should not be trusted to:
- invent a sampling frame;
- guess cluster sizes;
- replace inaccessible clusters;
- assign undocumented weights;
- choose a design without population and cost information;
- interpret specialised survey output without verification.
Random-selection code produced by AI should be reviewed, tested and saved with the seed and software version.
Privacy Considerations
Cluster identifiers can reveal schools, clinics or small geographic areas. Public-use datasets may therefore mask, aggregate or perturb cluster information.
Researchers should avoid publishing information that makes sensitive institutions or small communities identifiable.
Common Cluster-Sampling Mistakes
Mistake 1: Selecting Clusters Conveniently
Choosing the nearest five schools is convenience sampling, even when the schools are called clusters.
Mistake 2: Sampling from Every Group and Calling It Cluster Sampling
If researchers sample students from every academic department, departments are functioning as strata rather than selected clusters.
Mistake 3: Ignoring Cluster Size
Sampling the same number of people from differently sized clusters can produce unequal inclusion probabilities.
Mistake 4: Replacing a Difficult Cluster
Replacing a selected rural village with an easier nearby village changes the probability design unless a valid replacement procedure was specified beforehand.
Mistake 5: Treating the Nominal Sample Size as Fully Independent
A sample of 2,000 students from four schools does not contain the same amount of independent school-level information as 2,000 students from 100 schools.
Mistake 6: Using a Standard Sample-Size Calculator Without a Design Effect
Conventional calculations usually assume independent observations. A cluster design normally requires adjustment.
Mistake 7: Analysing the Data as a Simple Random Sample
Ignoring PSUs, weights and strata may produce misleading confidence intervals and significance tests.
Mistake 8: Assuming Every Cluster Must Be a Perfect Mini-Population
Internal diversity can improve efficiency, but probability selection and correct estimation are more fundamental than visual resemblance.
Mistake 9: Confusing Cluster Sampling with Cluster Analysis
Cluster analysis is a data-analysis method that groups similar cases or variables. Cluster sampling is a method of selecting a sample.
Mistake 10: Confusing Sampling with Cluster Randomisation
A cluster-randomised trial assigns interventions to groups. Cluster sampling selects groups into a study. A study may use either or both procedures.
Cluster Sampling vs. Cluster-Randomised Trials
| Feature | Cluster sampling | Cluster-randomised trial |
|---|---|---|
| Main purpose | Select units from a population | Assign an intervention |
| Randomised object | Cluster selection | Treatment allocation |
| Example | Randomly select 30 clinics | Randomly assign clinics to intervention or control |
| Main inferential issue | Selection probabilities and survey variance | Treatment-effect estimation with clustered outcomes |
| Can both occur together? | Yes | Yes |
A trial may recruit a convenience set of hospitals and randomise hospitals to treatment. That is cluster randomisation, but not probability cluster sampling from all hospitals.
Conversely, a descriptive survey may randomly sample schools without assigning any intervention. That is cluster sampling, not a cluster-randomised trial.
How to Report Cluster Sampling in a Methodology
A clear methods section should state:
- the target population;
- the sampling frame;
- what constituted a cluster;
- any stratification;
- the number of clusters in the frame;
- the number selected;
- the first-stage selection method;
- the size measure used for PPS, if applicable;
- the within-cluster selection method;
- planned and achieved sample sizes;
- inclusion probabilities and weights;
- nonresponse procedures;
- prohibited or permitted replacement rules;
- analysis software and variance-estimation method.
Cluster-Sampling Methodology Template
The target population comprised [population] in [location and period]. The first-stage sampling frame contained [number] clusters, defined as [definition]. Clusters were stratified by [variables, if applicable] and [number] clusters were selected using [simple random/systematic/PPS] sampling. Within each selected cluster, [all eligible units were included/a random sample of units was selected using method]. Final inclusion probabilities combined selection probabilities across stages. Sampling weights were calculated as [brief explanation] and adjusted for [nonresponse/calibration, if applicable]. Analyses accounted for clustering, stratification and weights using [software and procedure].
Example Methodology Paragraph
The study population included all teachers employed in government secondary schools in the state during the 2026 academic year. Schools served as primary sampling units. The school frame was stratified by region and urban–rural classification, after which 40 schools were selected with probability proportional to the number of eligible teachers. An updated teacher list was obtained from each selected school, and 20 teachers were selected by simple random sampling. Final weights were based on the inverse of the combined school- and teacher-selection probabilities and were adjusted for teacher nonresponse. Estimates and standard errors were calculated using survey procedures that incorporated the sampling weights, strata and school identifiers.
Conclusion
Cluster sampling is a probability method designed for populations organised into geographic, institutional or administrative groups. Its main benefit is practical efficiency: researchers can reduce travel and frame-construction costs by selecting clusters before individuals.
A defensible design requires more than randomly choosing a few locations. Researchers must define the population and clusters, maintain a suitable frame, calculate inclusion probabilities, address unequal cluster sizes, plan for the design effect and analyse the data using methods that recognise clustering, weights and stratification.
