How many schools should you survey?
For a typical state or district learning survey at 95 percent confidence and a ±5 point margin, the honest answer is usually 90 to 120 schools. The calculation: take the effective sample a simple random draw would need (about 385 children), multiply by the design effect of surveying through schools (often 3 to 5), then divide by the children you test per school. And never go below roughly 25 to 30 schools per group you want to report on, no matter what the child count says.
This is the question every education evaluation budget hangs on, and the one generic sample size calculators cannot answer, because they stop at "385 children" and never mention that reaching children through schools changes the arithmetic. Here is the full chain, with numbers you can check.
The three-step calculation
Worked, for a state learning survey with 20 children per school and ICC 0.2:
- Effective sample: 385 children (95 percent confidence, ±5 points).
- Design effect: 1 + 19 × 0.2 = 4.8, so field sample = 385 × 4.8 ≈ 1,848 children.
- Schools: 1,848 / 20 ≈ 93 schools.
The lever most teams miss: fewer children per school
Because DEFF depends on the per-school take, shrinking it helps twice:
| Children per school | DEFF (ICC 0.2) | Total children | Schools |
|---|---|---|---|
| 30 | 6.8 | 2,618 | 88 |
| 20 | 4.8 | 1,848 | 93 |
| 10 | 2.8 | 1,078 | 108 |
Reading down the table: visiting 15 more schools saves 770 child assessments for identical precision. Whether that trade wins depends on your travel cost per school versus your cost per assessment; enumerator teams that test a whole classroom anyway often stay at 20, while assessments that take 30 minutes per child push hard toward 10. The point is that this is a choice to make deliberately, not a default to inherit.
What about districts, blocks and strata?
Real designs sample in stages: districts (often chosen purposively or as strata), then schools within districts, then children within schools. Stratifying by district protects representation and costs nothing in sample; it is the clustering by school that inflates it. If every district must be reported separately, each district needs its own 90-plus schools, which is why "state estimate" and "every district estimated" differ in budget by an order of magnitude. See stratified sampling and cluster sampling for the two halves of this design.
Frequently asked questions
How many schools for a state-level learning survey?
Usually 90 to 120: about 385 effective sample, times a DEFF of 3 to 5 from school clustering, divided by 10 to 20 children per school. Sampling Shala's Build mode computes it for your own population, margin and ICC, and flags when the answer exceeds the schools you actually have.
More schools or more children per school?
More schools. Extra children from the same school add little new information because they are alike. Fewer children per school lowers the design effect, cutting total assessments for the same precision, at the price of more travel.
What ICC should I plan with?
For learning outcomes, 0.15 to 0.3 is the common range; 0.2 to 0.25 is a defensible default for Indian school data. Report the value you assumed; it is the single most consequential assumption in the design.
Get your school count, computed honestly
Build mode takes your population, margin, ICC and per-school take, and returns the full field plan with the justification written out.
Open Build mode