How Many Responses Do You Need? A Sample Size Guide for Surveys and Studies

Last updated October 30, 2026·4 min read

Every sample size answer is a sentence about a decision. A survey that will guide a budget call needs tighter precision than one tracking a trend nobody acts on. A study comparing two programs needs to detect the smallest difference that would change the funding choice, not the largest difference anyone can imagine. The formulas in this guide are short. The work is naming the numbers that go into them: the precision or effect worth having, the confidence level, and the realities of the design, clustering, weighting, and response rates included.

Case one: a survey proportion

The commonest planning question asks how many completed responses produce a stated margin of error. At 95 percent confidence, planning at the worst-case proportion of one half:

• For plus or minus five points: n = 1.96² × 0.5 × 0.5 ÷ 0.05² = 384.16, so 385 completes.
• For plus or minus three points: the same formula gives 1,067.11, so 1,068 completes.
• At 400 completes, the margin is plus or minus 4.9 points. At 1,000, it is plus or minus 3.1 points.

Two lessons hide in those numbers. Precision improves with the square root of the sample, so halving the margin of error costs four times the completes, which is why the jump from five points to three points nearly triples the fieldwork. And the worst-case proportion is a planning choice: if the true proportion sits near 0.2 or 0.8, the required n falls by roughly a third, but planning on a friendlier proportion than the data deliver is how surveys end up underprecise.

Case two: comparing two groups

For a two-group comparison of means, the planning input is the standardized effect, Cohen d: the difference divided by the standard deviation. At alpha of 0.05, two-sided, and power of 0.80:

• A medium effect, d of 0.5: exact solution 63.77 per group, so 64 per group and 128 total.
• A small effect, d of 0.2: the requirement rises to roughly 394 per group. Small effects are expensive, and the honest response to that price is a decision about whether the small effect would change anything, not a quieter alpha.

The design takes its share

Real surveys rarely draw simple random samples, and the design charges for its departures:

• Design effect. Clustering and weighting inflate variance. The required sample multiplies by the design effect: 385 completes at a design effect of 1.5 becomes 577. For cluster samples the effect is 1 + (m − 1) × ICC, where m is the cluster size and ICC the intraclass correlation, so large clusters with even modest correlation get expensive quickly.
• Finite population correction. Sampling a large share of a small population reduces the requirement. A town of 2,000 households does not need 385 completes for five-point precision; the correction brings it to about 323. (385 adjusted: 385 ÷ (1 + 384 ÷ 2,000) = 323.0.)
• Response rate. Completes are the denominator of precision; invitations are the denominator of cost. At a 25 percent response rate, 577 completes require 2,308 invitations. Planning that skips this step budgets a survey that cannot finish.

Subgroups are separate studies

A margin of error describes the estimate it is attached to. A survey with 1,000 completes and plus or minus 3.1 points overall holds perhaps 120 completes in a subgroup of interest, where the margin is near nine points. If a subgroup number will be reported, plan its sample on its own, and oversample where the population share is small. Weighting will rebalance the shares afterward; it cannot manufacture subgroup cases that were never collected.

A short planning checklist

• Name the primary estimate or comparison, one only, and plan for it.
• State the precision or smallest effect worth detecting, in subject units.
• Choose alpha and power before seeing any data, and write them in the plan.
• Multiply by the design effect, correct for a finite population where it applies, and divide by the expected response rate.
• Check every subgroup that will appear in the report, and plan the smallest one.
• Round up at every step. Fractions of respondents do not exist, and rounding down always rounds the precision down with them.

Run the numbers

The Sample Size and Power Planner in Analysis Studio solves each case in this guide and shows the sensitivity of the answer to every input:

Module 7: Sample Size and Power PlannerOpen in Analysis Studio →

Sources and method notes

Formulas follow the standard planning literature for t tests (noncentral t), proportions (normal approximation with the worst-case variance), and complex designs (Kish design effect; design-based finite population correction). All arithmetic in this guide was recomputed at drafting and is re-verified at build.