Sample size calculator for surveys and program evaluation
How many people do you need to survey? For estimating a proportion at 95 percent confidence with a 5 percent margin of error, the answer is about 385 respondents from a large population. But for a real program evaluation, that is only the starting point: clustering, statistical power, attrition and multiple stakeholder groups all change the number. Sampling Shala computes all of it, free, in your browser.
Most sample size calculators answer one narrow question: how many people for one survey estimate. Program evaluations are rarely that simple. You survey students through schools, so clustering inflates the sample. You interview teachers, head teachers and parents too, each needing their own method and size. You measure change over time, so power and attrition matter. Sampling Shala's Build mode is a five step wizard (Question, Scale, Who, Precision, Strategy) that handles all of this and writes a defensible justification you can paste into a proposal.
Quick reference: survey sample sizes at 95 percent confidence
For a proportion near 50 percent (the most conservative case), in a simple random sample from a large population:
| Margin of error | Sample size needed |
|---|---|
| ±3 percentage points | about 1,067 |
| ±5 percentage points | about 385 |
| ±7 percentage points | about 196 |
| ±10 percentage points | about 97 |
Note what this table hides: halving the margin of error roughly quadruples the sample, because the margin is squared in the formula. Precision gets expensive fast. Our margin of error explainer walks through why.
Power: sample size for detecting impact
If your question is "did the program work?" rather than "what is the rate?", you need a power calculation, not a margin of error. For a two group comparison (treatment versus control):
At 80 percent power and a 0.3 standard deviation effect, that gives roughly 175 per arm, about 350 children in total, before clustering and attrition. Sampling Shala computes this for RCTs, quasi-experimental designs with matching inflation, baseline-endline designs where panel correlation reduces the sample, and longitudinal cohorts with compound attrition across waves.
What this calculator does that generic ones do not
- Multi-stakeholder designs. Students, teachers, head teachers, parents, and officials at cluster, block, district and state level, each set to statistical, quota, census or purposive sampling with its own size logic.
- Cluster reality. Design effect (DEFF = 1 + (m - 1) × ICC) with geography-aware ICC defaults, so school-based surveys are sized honestly. See how many schools to survey.
- Eight evaluation designs. Descriptive survey, baseline-endline, longitudinal cohort, RCT, quasi-experimental, LQAS, mixed methods, and qualitative saturation-based designs.
- A plan, not just a number. The output is a full strategy: stakeholder table, precision parameters, field cascade, structural warnings, and a paste-ready defensible justification.
- Honest maths, in the open. Every formula is published in the methodology. A deterministic engine does all arithmetic; nothing is estimated by AI.
Frequently asked questions
How many people do I need for a survey at a 5 percent margin of error?
About 385 respondents, for a proportion at 95 percent confidence in a simple random sample from a large population. If your population is small, the finite population correction reduces this. If you reach people through clusters such as schools or villages, the design effect increases it.
What sample size do I need to detect an impact of 0.3 standard deviations?
Roughly 175 participants per arm at 80 percent power and a two-sided 5 percent significance level, so about 350 in total. Add more for clustering (multiply by DEFF) and expected attrition (divide by the retention rate).
Why does surveying through schools increase my sample size?
Children in the same school tend to resemble each other, so each additional child from the same school adds less new information than a child from a new school. The design effect quantifies the penalty: with 20 children per school and an intra-cluster correlation of 0.2, you need almost five times the simple random sample size. Surveying more schools with fewer children per school reduces the penalty.
Is this calculator really free?
Yes. The core tools are free, need no account, and run entirely in your browser; nothing you enter is sent to any server. Sampling Shala is a planning and learning aid, and its outputs should be confirmed with a qualified statistician before use in a real evaluation.
Build your sampling plan in five steps
Question, scale, who, precision, strategy. A complete plan with a defensible justification, free.
Start building