How to determine the appropriate sample size for a survey

Sample size is not a software default. It follows the design, the outcome, and the precision the study needs.

The decision comes first

The first question is not how many people you can reach. The first question is what decision or claim the study must support. A national prevalence estimate, a comparison between two groups, and a model with several predictors do not share one formula.

Write down the primary outcome. Decide whether you need a proportion, a mean, a difference, or an association. Then choose the precision you can defend: a margin of error, a minimum detectable difference, or an interval width that still lets a reader judge the result.

Inputs have to be honest

Power calculations assume an effect, a variance, a design, and a loss rate. If those inputs are guesses, the number they produce is a guess. Use published studies or a small pilot to set realistic values. If you cannot, report a range of sample sizes under honest assumptions rather than a single figure that looks exact.

Design effects matter. Cluster samples, stratified samples, and unequal selection probabilities change the effective sample size. A simple random sample formula applied to a clustered field survey will understate how many interviews you actually need.

Loss in the field is part of the plan

People who cannot be reached, who refuse, or who are ineligible are not a surprise. Inflate the target by a realistic completion rate for the setting, not a round number copied from another country.

A larger sample does not repair a biased design. If the frame misses the people who matter, or if interviewers skip households that are hard to reach, extra interviews will not save the estimate. Sample size answers a precision question. Coverage and measurement answer a validity question.

Sample size follows the design. It is not a number the software owes you.

What this note is for

Software can compute the arithmetic once the design is specified. It cannot tell you which outcome is primary, whether a three percentage point margin is tight enough, or whether the clusters in the field match the clusters in the formula.

This note is a way to think, not a substitute for a statistician on a funded protocol. If the study will be used to allocate funds, change a programme, or support a licence, get the sample size written into the protocol and reviewed before the field starts.