Last updated: 17 September 2026
Quick Answer: Cluster sampling's main advantage is cost and its main limitation is precision. You build a sampling frame only for the clusters you select, and fieldwork concentrates geographically. At an intraclass correlation of 0.05 with 20 interviews per cluster, 400 interviews carry the precision of about 205, a 49% loss.
Cluster sampling divides a population into naturally occurring groups, usually geographic ones, draws a random sample of those groups, and collects data only from inside them. A clustered random sample is the same design under another name.
The National Health and Nutrition Examination Survey draws its sample in four stages, counties, then census blocks, then households, then individuals, and each two-year cycle rests on just 30 counties. It examined 18,248 people in 2015 to 2018, and warns that so few primary sampling units can destabilize variance estimates.
The cost saving comes from two sources, and one of them vanishes when fieldwork stops involving travel.
What Are the Advantages of Cluster Sampling?
The advantages of cluster sampling are operational rather than statistical: where no register of individuals exists, a census area frame is the only route to known, non-zero selection probabilities.
- The frame saving. Most countries hold no list of their adults to sample from. The DHS Program sampling manual is blunt: the best frame is the census enumeration area list.
- The travel saving. Interviewers work one block at a time instead of criss-crossing a region, and this saving exists only where fieldwork has a geography.
What the Cost Ratio Does to Cluster Size
The United Nations handbook on household sample surveys puts the optimum interviews per cluster at the square root of the cost ratio times one minus rho over rho. Across eight surveys the DHS manual tabulates cost ratios of 10 to 52 and optimal takes of 13 to 47.
The cost ratio compares one more cluster against one more interview inside one. On a call or a messaging thread, reaching someone in a new area costs about what one more respondent in the same area costs, so the cost ratio falls toward 1 and the optimum take drops to about four. The arithmetic stops recommending clustering.
What Are the Disadvantages of Cluster Sampling?
The disadvantage of cluster sampling is variance: people in the same cluster resemble each other, so each extra interview inside one adds less new information than one in a fresh cluster.
Rho, the intraclass correlation, is the share of total variance sitting between clusters rather than within them. A methods paper in the Annals of Family Medicine puts typical human-study values at 0.01 to 0.02, and the design effect for equal cluster sizes at 1 plus m minus 1, times rho, m being the number per cluster.
How Many Extra Interviews Does the Design Effect Cost You?
Say you want the precision of 400 simple random interviews, from 20 interviews in each of 20 clusters, on a variable with a rho of 0.05.
- The design effect is 1 plus 19 times 0.05, which is 1.95.
- Multiply the target by it: you need 780 interviews, not 400.
- At 20 per cluster that is 39 clusters, so 380 extra interviews and 19 extra field locations to stand still.
Spreading wider beats digging deeper, and a comparison of emergency field survey designs in Emerging Themes in Epidemiology quantifies it: for an outcome whose design effect is 2.0 under the classic 30 by 30 design, 30 clusters of 7 give 1.21 and 67 of 3 give 1.07.
What Are the Types of Cluster Sampling?
Which type you pick is set by the frame you can obtain and the design effect you can afford. Rows are ordered alphabetically by design name, so the table makes no argument about which is best.
| Design | Frame it needs | Design-effect exposure |
|---|---|---|
| Multistage | A frame at every stage | Compounds at every stage |
| One-stage | Cluster list, then all members | Highest, the take equals cluster size |
| PPS two-stage | Cluster list with a size measure | Controlled, the take stays constant |
| Stratified within cluster | Cluster frame plus an internal classifier | Reduced where the stratifier predicts |
| Two-stage | Cluster list, then a member list per cluster | Governed by the take per cluster |
What Does a Cluster Sample Look Like in Practice?
Two public programs show the design in practice.
A school-based assessment. The National Assessment of Educational Progress selects 50 to 100 geographic primary sampling units of one or more counties in national-only years, then schools, then about 50 students each. The detail worth copying is what it refuses to do: students come from a roster of names, not a list of classrooms.
A multi-country household survey. DHS surveys are two-stage: enumeration areas drawn with probability proportional to size, then households from a listing made fresh in each area. It clusters because in most of the countries it covers, no household list exists.
Which Respondents Does a Cluster Design Systematically Under-Reach?
A cluster design inherits every access constraint of the places it selects, and the people it loses are not missing at random.
- People without a listable dwelling. Anyone not attached to a structure the lister recognizes sits outside the frame before sampling starts: people in institutions, in informal settlements, or living at their workplace.
- Anyone the channel cannot reach once the cluster is fixed. The Current Population Survey assigns about 74,000 housing units a month, finds about 62,000 eligible and completes about 54,000 interviews.
Remote modes lift both constraints: a call or a messaging thread needs neither the respondent at home nor the cluster physically reachable. The frame problem moves instead, since you now need a phone number or a handle, and the people with neither are a different non-random group, the problem behind reaching respondents without smartphones. Who a remote sample still misses works through where that phone frame falls short.
Alchemic runs interviews natively inside WhatsApp with no link and no app and by outbound AI phone call as part of end-to-end consumer research at scale. It publishes 57+ languages including Spanish, Arabic and Mandarin, and offers managed fieldwork or bring your own recruitment across 14 markets including the USA and the UK.
When Is Cluster Sampling the Wrong Choice?
Cluster sampling is wrong in three situations, and in each you pay the variance penalty without the saving that justifies it.
- You already have a usable frame. Customer lists, subscriber files, employee registers or panel databases: draw names directly and clustering buys nothing.
- The fieldwork carries no travel cost. Online, phone and messaging studies push the cost ratio toward 1, where the formula stops recommending clusters.
- The outcome is strongly clustered. Anything set by shared local conditions, water supply, network coverage, one retailer's assortment, carries a high rho.
On whether stratified or cluster sampling is better, neither is: stratification is a precision tool that divides the population and samples every stratum, while clustering is a cost tool that samples some groups and ignores the rest. If you can afford to reach every group, stratify; if you cannot, cluster and budget for the design effect.
None of this makes remote modes universally better. Where a study needs a physical act, a shelf observed or a body measured, geographic clustering is the only design that works. That choice sits with types of market research, and teams running several emerging markets at once often answer differently by country.
The AAPOR Code of Professional Ethics and Practices, revised in June 2026, requires immediate disclosure of the sampling method, the number of completed interviews and the weighting procedure, and for probability samples it asks you to state whether the reported sampling error has been adjusted for the design effect due to clustering. A cluster sample reported as simple random claims a precision the design does not support.