What is Cluster Sampling?

Cluster sampling is a probability sampling method in which the population is divided into naturally occurring groups (clusters), a random sample of clusters is selected, and then either every unit within the chosen clusters is studied (single-stage) or a random sample of units within them is drawn (two-stage or multistage). The clusters are a means of access, not the object of interest: the researcher wants to know about individuals but can only practically list, reach, or afford to visit groups. Contrast stratified sampling, which divides the population into groups and samples from every group to guarantee coverage; cluster sampling samples some groups to save resources, and the two designs point in opposite directions on the precision-cost axis.

The Design Effect

The cost of clustering is statistical. People inside a cluster resemble each other (colleagues share workflows, households share habits, an account's users share the same configuration), so each additional person from the same cluster adds less new information than a fresh independent draw would. The measure of this resemblance is the intraclass correlation, and its consequence is the design effect: the factor by which variance is inflated relative to a simple random sample of the same size. A survey of 1,000 users drawn from 20 accounts may carry the precision of a few hundred independent respondents. The practical rules follow: more clusters with fewer units each beats fewer clusters with many, and any analysis must account for the clustering (cluster-robust standard errors, multilevel models, or survey-weighted procedures), because treating clustered observations as independent understates uncertainty, sometimes drastically, and manufactures significance out of correlated noise.

Cluster Sampling in Product Research

The design is more common than its name. B2B research routinely samples accounts, then studies several users per account: a cluster design whether or not anyone called it one, with the design effect waiting in the analysis. Experiments assigned at the team or organisation level (to prevent contamination between users who share a workspace) are cluster-randomised trials, and their statistics must run at the cluster level. Field research visiting a handful of sites and observing many people at each is clustered. In every case the honest n is closer to the number of clusters than the number of people, and the readout should say so.

Doing It Well

1. Randomise the clusters.
The design is probability sampling only if clusters are drawn randomly from a complete list of clusters; hand-picked accounts are a judgment sample with extra steps.

2. Prefer many small clusters.
Precision comes from cluster count; when the budget allows either 10 accounts with 20 users or 40 accounts with 5, choose the second.

3. Sample within clusters randomly too.
Letting the account's champion nominate participants re-introduces selection; draw within-cluster units by lottery or with clear rules.

4. Analyse at the right level.
Use methods that respect the hierarchy; report the design effect or effective sample size alongside the raw n.

5. Recruit the access, verify the individuals.
Clusters get you in the door; screening and behavioural evidence for the individuals inside keep the sample honest, the same screener discipline that governs any recruitment.

The Takeaway

Cluster sampling buys access and economy by sampling groups instead of people, and pays in precision through the design effect. Randomise the clusters, favour many small ones, sample within them fairly, and analyse with the hierarchy intact. The number of people in your study is not your sample size when they arrived in groups; the number of groups is closer to the truth, and readouts that know the difference earn their confidence intervals.