• Cluster sampling is a sampling method in which the population is divided into naturally occurring groups, and the researcher then randomly selects some groups and studies all members of those groups.
• You sample clusters rather than individuals, so you only need a list of clusters, not of every person.
• Each cluster should ideally be a miniature version of the population, whereas strata in stratified sampling are population subgroups that have a certain variable in common.
• Cluster sampling is cheaper than other probability sampling methods and less precise.

What is cluster sampling in research?

Cluster sampling is a probability sampling technique in which a population is divided into groups that already exist, such as schools, hospitals, villages, or city blocks. The researcher then selects some of those groups at random and collects data from the people inside them.

The crucial feature is what gets sampled first. In most techniques you draw individuals; in cluster sampling you draw whole groups, and individuals enter the study because their group was chosen.

Why use cluster sampling

Cluster sampling makes conducting research cheaper and faster. Studying 1,200 students spread across 2,000 schools would mean visiting hundreds of sites. Studying 1,200 students drawn from 40 selected schools means 40 visits.

It also doesn’t require a sampling frame of every individual in the population. You need a list of clusters rather than a list of every individual, and lists of schools, clinics, or districts usually already exist.

Important definitions

Term	What it means
Cluster	A naturally occurring group, such as a school or clinic
Primary sampling unit	The unit selected at the first stage, usually the cluster
Secondary sampling unit	The unit selected at the second stage, often the individual
Intracluster correlation	How similar members of the same cluster are to each other
Design effect	How much precision is lost compared with a simple random sample

 

Steps in cluster sampling

• Step 1. Define the population and choose the clustering unit. Decide what counts as a cluster, such as a school, a ward, or a pharmacy.
• Step 2. Build a list of clusters. You need a complete frame of clusters, but not of individuals.
• Step 3. Decide how many stages to use. Choose between studying everyone in a selected cluster or subsampling within it.
• Step 4. Select clusters at random. Use simple random sampling and consider weighting by cluster size.
• Step 5. Collect the data. Study every member of the chosen clusters, or draw a random sample within each.

 

Step 4 is where the probability requirement is met. If clusters are chosen because they are nearby or cooperative, the design stops being a probability sample and becomes convenience sampling with extra steps.

Also keep in mind that a good cluster contains as much variety as the population itself, so that a handful of clusters can stand in for the whole. A cluster full of similar people is a poor cluster, because studying 50 of them tells you little more than studying 5. Clusters whose members closely resemble each other make your sample less representative. For example, a public school that accepts all children living in the district vs a private school with high fees and a competitive entrance test.

 

When should you use cluster sampling?

Cluster sampling is the right choice in these situations.

• The population is spread across a wide geographic area.
• No list of individuals exists, but a list of groups does.
• Travel or site access is the dominant cost in your budget.
• Natural groupings already exist and are easy to define.
• An intervention must be delivered to whole groups rather than individuals.

Don’t use cluster sampling when

• Members of each cluster are very similar to one another (e.g., a Mandarin immersion school in the US will likely have most students with a Chinese background)
• You need maximum precision from a limited sample size.
• Only a few clusters exist, which leaves too little to randomize over.
• A complete list of individuals is already available.

Types of cluster sampling

1. Single-stage cluster sampling

Clusters are selected at random, and every member of each selected cluster is included. It is simple to administer and it works well when clusters are small, but it can produce a large sample very quickly if they are not.

2. Two-stage cluster sampling

Clusters are selected at random, then a random sample of members is drawn within each selected cluster. This keeps the sample size manageable and is the version most surveys use.

3. Multistage cluster sampling

Selection proceeds through 3 or more levels, such as region, then district, then school, then student. National surveys almost always work this way, because no single list covers the whole population.

4. Probability proportional to size

Clusters vary in size, and treating a school of 2,000 like a school of 200 distorts selection probabilities. This refinement gives larger clusters a proportionally greater chance of being selected, then takes an equal number of members from each, so every individual ends up with the same overall probability of selection.

Cluster sampling methods at a glance

Type	How members are chosen	Best used when
Single-stage	Everyone in the selected clusters	Clusters are small
Two-stage	A random subsample within each cluster	Clusters are large
Multistage	Selection through 3 or more levels	The population is national
Proportional to size	Larger clusters more likely to be picked	Clusters differ greatly in size

 

Examples of how to conduct cluster sampling

Example 1: education

A research team wants to measure adolescent wellbeing across a country with 500,000 secondary students in 2,000 schools. No national list of students exists, and no ethics committee would approve building one.

A list of schools does exist, so the team samples schools instead, then samples students within the schools it selects.

Stage	Unit sampled	Number
Stage 1	Schools, drawn at random	40 of 2,000
Stage 2	Students within each school	30 per school
Total	Students in the final sample	1,200

 

Fieldwork now involves 40 visits rather than hundreds. The cost saving is large enough to make the study possible at all, which is the usual reason cluster designs are chosen.

The weakness is that students at the same school resemble each other. They share a catchment area, a teaching culture, and often a social background, so the 30th student from a school adds less new information than a 30th student drawn at random from the country would.

Example 2: management science

A researcher wants to measure employee engagement across a retail chain with 1,200 stores nationwide. Head office holds a list of stores, but staff turnover means no reliable central list of individual employees exists.

The researcher randomly selects 60 stores, which is 5 percent of the chain, then surveys 50 employees at each one, giving 3,000 responses in the final sample.

Fieldwork concentrates on 60 sites instead of chasing employees scattered across the country, and every store had an equal chance of selection. No regional manager can claim their stores were targeted, which matters when results feed into performance discussions.

 

Cluster sampling vs stratified sampling

Stratified sampling samples from every stratum. Cluster sampling samples only some of the clusters and ignores the rest entirely. Strata should be homogeneous within and different from each other. Clusters should be heterogeneous internally and similar to each other externally.

Feature	Cluster sampling	Stratified sampling
Groups formed by	Natural grouping, often location	A shared characteristic
Which groups are sampled	A random selection of them	All of them
Ideal group make-up	Varied inside, like the population	Group members are similar to each other but different from the rest of the population
Main aim	Lower cost	Greater precision
Sampling frame required	A list of clusters	A list of individuals plus subgroup data

 

Both techniques divide the population into groups before sampling, which is why they are so often confused. They differ in how many groups are sampled and in what a good group looks like.

The 2 are often combined rather than chosen between. A national survey might stratify regions by urban and rural status, then sample clusters within each stratum, capturing the cost advantage of one and the coverage guarantee of the other.

Advantages and disadvantages of cluster sampling

Advantages

• Much cheaper: fieldwork concentrates at a limited number of sites.
• No list of individuals needed: a frame of clusters is enough to begin.
• Practical for dispersed populations: it makes national and regional studies feasible.
• Supports inference: clusters are selected at random, so it remains a probability technique.
• Allows larger samples: the money saved on travel can fund more participants.

Disadvantages

• Lower precision: confidence intervals are wider than for a simple random sample of the same size.
• Intracluster correlation: members of a cluster resemble each other, so each adds less new information.
• Larger samples required: the design effect must be built into the sample size calculation.
• More complex analysis: standard formulas assume independent observations and must be adjusted.
• Risk of unrepresentative clusters: selecting few clusters means an unusual one can distort the whole result.

How to reduce the disadvantages of cluster sampling

Problem	How to reduce it
Intracluster correlation	Sample more clusters with fewer members in each
Unrepresentative clusters	Stratify clusters first, then sample within strata
Unequal cluster sizes	Select clusters with probability proportional to size
Understated uncertainty	Use survey analysis methods that account for clustering

 

Frequently asked questions

Is cluster sampling probability or non-probability sampling?

It is a probability technique, provided the clusters are selected at random. Every individual then has a known and calculable chance of selection, even though nobody was chosen individually.

If clusters are picked for convenience, such as the 3 nearest schools, the design is no longer a probability sample and the results do not support population estimates.

What is the difference between cluster sampling and multistage sampling?

Multistage sampling is a form of cluster sampling, not a separate technique. Single-stage cluster sampling stops after selecting clusters and studies everyone inside them.

Multistage designs continue sampling at each level: regions, then districts, then schools, then students. Nearly every large national survey is multistage, because no single list spans the whole population.

What is the design effect in cluster sampling?

The design effect measures how much precision a cluster sample loses compared with a simple random sample of the same size.

Design effect = 1 + ((m – 1) × ICC)

Here m is the number of members sampled per cluster and ICC is the intracluster correlation. With 30 students per school and an ICC of 0.02, the design effect is 1 + (29 × 0.02), which is 1.58.

A sample of 1,200 students therefore carries about as much information as a simple random sample of 760. Sample size calculations must be multiplied by the design effect, or the study will be underpowered.

How many clusters should you select?

More clusters with fewer members in each beats fewer clusters with many members, because precision depends far more on the number of clusters than on the number of people inside them.

A common working guide is at least 30 clusters, and more when the intracluster correlation is high. Sampling 60 schools with 20 students each will outperform 20 schools with 60 students each, even though both give 1,200 participants.

What is a cluster randomized trial?

It is an experiment in which whole groups, such as clinics or pharmacies, are randomly assigned to the intervention or the control condition, rather than randomizing individual participants.

It is used when an intervention cannot be confined to individuals, or when treating some patients differently within the same site would contaminate the comparison. Analysis must account for clustering, and the required sample size is larger than an individually randomized trial would need.