Ask a field program “what share of children here are fully vaccinated?” and the honest answer usually comes from a survey, not a registry. Health information systems in low-resource settings routinely undercount or overcount doses given, so operational researchers reach for a sampling method that can estimate coverage directly from households. For decades that method has been the WHO Expanded Programme on Immunization’s 30-cluster, or 30×7, survey — a design built to turn a sample of roughly 210 children into a defensible national or district estimate. Understanding how it works, and why WHO revised it in 2018, is a useful case study in how field epidemiology trades precision for feasibility.
The two-stage design behind the 30×7 survey
The 30×7 approach is a two-stage cluster sample, developed for EPI in the late 1970s. In the first stage, the target area is divided into non-overlapping clusters, usually villages, wards, or enumeration areas, and 30 of them are selected with probability proportional to population size, so larger clusters have a proportionally larger chance of being drawn. In the second stage, field teams pick a starting point inside each selected cluster, often a household chosen at random, and then work outward — historically to the “nearest” household — until seven eligible children (commonly those aged 12–23 months) have been enrolled. Thirty clusters times seven children gives the classic total sample of 210.
The appeal of this structure is practical rather than statistically elegant: it does not require a complete list of every household in the survey area, which most immunization programs never had. Teams only need a map or population estimate at the cluster level and a simple rule for moving through households once they arrive. That is what made it exportable to dozens of countries with thin administrative data.
Why 30 clusters and seven children, specifically
The 30×7 parameters were not arbitrary. Analysis in the early 1980s that compared dozens of real EPI coverage surveys against thousands of simulated ones found that a 30-cluster, seven-per-cluster design would typically place a coverage estimate within about 10 percentage points of the true value at 95 percent confidence — precise enough to flag a program that was clearly underperforming, cheap enough that a national program could actually afford to run it every year or two. Seven respondents per cluster was, in effect, the smallest workload that kept the overall sample near 210 once the target precision and an assumed level of within-cluster similarity were factored in.
The design effect: why 210 clustered children are not 210 independent ones
The number that makes cluster surveys harder to interpret than a simple random sample is the design effect. Children in the same village tend to resemble each other on vaccination status more than children picked at random from the whole country would — they share the same health post, the same outreach schedule, the same community attitudes toward vaccination. That within-cluster similarity is captured by the intracluster correlation coefficient (ICC); the higher the ICC, the more redundant information each additional child in the same cluster contributes.
The design effect translates that redundancy into a concrete penalty on precision: it is the ratio between the variance you actually get from a clustered sample and the variance you would have gotten from a simple random sample of the same total size. A design effect of 2, for instance, means the clustered survey’s uncertainty is twice as large as an unclustered sample of the same 210 respondents would produce — equivalent to having only about 105 truly independent observations.
| Design effect | What it implies | Practical consequence |
|---|---|---|
| ~1.0 | Clusters are no more alike than random | 210 children behave like a simple random sample |
| ~1.5–2.0 | Moderate within-cluster similarity | Effective sample size drops to roughly 105–140 |
| >2.5 | Strong within-cluster clustering (common where outreach is patchy) | Confidence intervals widen well past the intended ±10 points |
The original 30×7 design assumed ICCs for vaccination coverage would rarely exceed roughly 0.17, which kept the expected design effect manageable. Field experience showed that assumption did not always hold, particularly in settings where coverage varies sharply from one community to the next — exactly the settings where an accurate estimate matters most.
Where the original design fell short
Three limitations pushed WHO to revisit the 30×7 formula. First, precision built around a ±10-point target is coarse once national coverage climbs into the 80s and 90s, where the policy-relevant differences are much smaller than 10 points. Second, the classic “nearest household” rule for the second stage is not a true probability method: it systematically favors households clustered near the starting point over those on the periphery, introducing a selection bias that is hard to quantify after the fact. Third, because the underlying ICC assumption often understated real-world clustering, published coverage figures sometimes carried more uncertainty than their confidence intervals implied.
The 2018 revision and what changed
WHO’s updated guidance moved cluster-survey design away from a fixed 30×7 formula and toward calculating sample size explicitly from the target precision, an empirically estimated design effect, and a genuine probability method for selecting households within each cluster, rather than the nearest-household shortcut. The number of clusters and respondents per cluster is now derived for each survey’s specific coverage target and expected ICC instead of assumed to be universally 30 and 7. The underlying two-stage logic — population-weighted cluster selection, then a bounded number of respondents per cluster — survives; what changed is that programs are now expected to justify their sample size with real numbers rather than inherit a 40-year-old rule of thumb.
Conclusion
The 30×7 cluster survey earned its place as a field standard because it solved a real constraint: how to measure something as consequential as vaccination coverage without a population register, on a budget a district health office could actually carry. Its successor keeps the same two-stage architecture but replaces a fixed formula with a design effect calculated for the survey at hand. For anyone reading a coverage estimate coming out of a field program, the number worth asking about is not just the percentage covered, but how many effectively independent observations actually sit behind it.