Sample Size Calculator

Confidence level and margin of error

Confidence level (typically 95%) expresses how often the true population value would fall within your reported margin of error if you repeated the survey many times. A 95% confidence level means 95 out of 100 independent samples would produce an interval containing the true value.

Margin of error (often ±5%) is the maximum expected difference between the survey result and the true population value. Cutting the margin of error in half requires roughly four times as many respondents because the relationship is quadratic — halving E quadruples n.

Sample size formula

The standard formula for an infinite (or very large) population uses the critical z-value for the chosen confidence level, the margin of error, and the expected proportion p. Setting p = 0.5 gives the most conservative (largest) required sample, which is why 50% is the default when the true proportion is unknown.

Critical z-values for common confidence levels

Each confidence level maps to a fixed z critical value from the standard normal distribution: 80% → 1.282, 85% → 1.440, 90% → 1.645, 95% → 1.960, 99% → 2.576. The 95% level is by far the most common in published research and business surveys.

Moving from 95% to 99% confidence increases the required sample size by about 73% (for p = 0.5, E = 5%) because the z-value grows from 1.96 to 2.58 — a 31% increase that gets squared in the formula.

Finite population correction

When your population is small enough that your sample would represent a substantial fraction of it, the standard formula overestimates the required sample size. The finite population correction reduces n proportionally — for a population of 1,000 and an infinite-population requirement of 385, the corrected sample size is about 278.

Leave the population size blank if your population is large (over 100,000) — the correction becomes negligible (less than 0.4%) and the infinite-population formula is sufficient.

Completed responses and response rate

The number returned by this calculator is the target number of completed responses, not the number of invitations to send. If you expect a 25% response rate and need 385 completed surveys, you would plan on roughly 1,540 invitations. In general, invitations needed = required sample size / expected response rate.

That distinction matters because searchers often ask how many people they need to survey when they really mean how many replies they need to collect. Rival survey sample size calculators frame the result the same way: the sample target is about completed responses, while response rate is a separate planning step.

Planning with common confidence levels

The confidence-level comparison on this page lets you see how the same survey plan behaves at 80%, 85%, 90%, 95%, and 99% confidence. A looser confidence target can reduce fieldwork a lot, while a stricter target can make the required completes and invitations rise quickly.

For a proportion survey using 50% as the expected proportion, 95% confidence remains the standard benchmark because it balances reliability and practicality. If the project is only directional, 90% confidence or a wider margin of error may be enough. If the decision is high-stakes or the audience is small, the stricter settings are easier to justify.

Why margin of error usually drives the budget more than confidence level

Survey teams often focus first on confidence level because 90%, 95%, and 99% are easy labels to recognise. In practice, margin of error is usually the setting that changes fieldwork costs most dramatically. The formula squares the error term, so narrowing the interval from ±5% to ±3% can roughly triple the completed-response target, and pushing from ±5% to ±2% can multiply it by more than six.

That is why directional customer pulse surveys often tolerate a wider interval while board-level tracking studies, election polling, and regulated research budgets push for tighter precision. The question is not whether smaller error is better in the abstract. The question is whether the extra precision changes a real decision enough to justify the larger sample and outreach cost.

A practical planning habit is to compare at least three scenarios before launching: a lean read such as ±7% or ±10%, a standard read such as ±5%, and a high-precision read such as ±3%. Seeing those scenarios side by side usually makes the cost-accuracy tradeoff much clearer than treating the sample size calculator as a one-number answer.

Proportion surveys vs mean estimates

This calculator uses the proportion formula because that is the standard setup for yes/no questions, choice shares, and prevalence estimates. In those cases the expected proportion p drives the answer, which is why 50% is the conservative default when you do not know p.

If your study is about an average or mean rather than a share, the sample-size problem changes. Continuous outcomes such as blood pressure, spend, or reaction time need a standard-deviation-based formula, so you should not use a proportion sample-size calculator for that job.

A/B tests also need a different framing. Instead of a survey margin of error, they usually ask for a minimum detectable effect, a baseline conversion rate, statistical power, and a significance level. The math is related, but the planning question is different enough that a dedicated A/B test sample size or power calculator is a better fit.

Response rate, completion rate, and completed responses are different planning layers

The formula gives you completed responses, which is the number of usable survey records you want in the final dataset. That is not the same as the number of people who start the survey, and it is not the same as the number of invitations you have to send. Those are operational layers that sit on top of the statistical target.

Response rate describes how many invited people begin or submit the survey, depending on how your team defines it. Completion rate describes how many of the people who start the questionnaire make it through to the end. If the survey is long, has screening questions, or contains sensitive items that create drop-off, the completion-rate layer can materially raise the number of starts you need even when the statistical sample size itself does not change.

For example, if your sample size calculator result says you need 385 completed responses, an 80% completion rate means you need about 482 starts. If response rate is only 25%, you still need about 1,540 invitations to achieve those 385 usable completes. Keeping those layers separate makes the launch plan much more realistic.

Subgroup analysis is where total sample size often explodes

A single sample-size result only guarantees one headline estimate for the full audience. The moment you want stable results for smaller segments such as age bands, regions, customer tiers, or device types, the total fieldwork target usually grows. In the simplest case, if each subgroup needs the same precision, you multiply the completed-response target by the number of important segments.

This is why a survey that looks comfortably sized for one overall result can become underpowered the moment stakeholders ask for cuts by gender, product line, or tenure. A total sample of 400 might be fine for a single topline proportion, but it is not enough if you need four equally sized subgroups with similar precision. In that case you are closer to 1,600 completes before you even account for response rate or completion rate losses.

Real projects are often messier than equal-sized subgroups. The smallest critical subgroup tends to drive the plan, especially when the audience is imbalanced. If one region or customer tier is rare, you may need targeted recruitment or quotas rather than a simple multiply-the-sample rule.

Design effect: why 400 raw responses do not always behave like 400 random responses

The classic sample-size formula assumes simple random sampling. Real surveys often violate that assumption because they use panel weights, clustered recruitment, quota balancing, or multi-stage sampling. Those features can increase variance, which means the raw number of completed interviews overstates how much statistical precision you really bought.

Survey methodologists summarise that loss of efficiency with design effect, often written as DEFF. A design effect of 1.5 means your finished sample behaves more like a simple-random sample that is one-third smaller. In planning terms, that usually means multiplying the baseline completed-response target by the expected design effect before you turn the result into survey starts or invitations.

That is why the calculator now keeps a simple-random-sample baseline visible alongside the design-adjusted completed-response target. It helps answer the operational question stakeholders actually ask: how many extra responses do we need once weighting, clustering, or other survey-design inefficiencies are acknowledged?

Why weighting can shrink the effective sample

Weighting can improve representativeness, but unequal weights also tend to inflate variance. That means a weighted sample of 1,000 respondents can have the precision of a much smaller simple-random sample. Survey teams often describe that smaller benchmark as the effective sample size.

The practical lesson is not that weighting is bad. It is that sample-size planning should happen with the expected analysis design in mind. If a survey will rely heavily on weights or clustered recruitment, the raw completed-response target should usually be larger than the simple-random-sample formula suggests.

Why use 50% as the expected proportion?

The variance term p × (1 − p) in the formula is maximised when p = 0.5, giving the largest and most conservative sample size estimate. If you have prior knowledge suggesting the true proportion is close to 10% or 90%, using that value instead will reduce the required sample size significantly.

For example, at 95% confidence and ±5% margin of error, p = 0.5 requires 385 respondents, but p = 0.1 (or 0.9) requires only 139 — a 64% reduction. Use the conservative 50% when you have no prior estimate.

Worked example: a 1,000-person survey

Start with the classic survey benchmark: 95% confidence, ±5% margin of error, and a 50% expected proportion. For a very large population, the calculator returns 385 completed responses. If your total audience is only 1,000 people, the finite population correction lowers that to about 278 completes.

If the same survey is expected to get a 25% response rate, the fieldwork target becomes about 1,112 invitations for the 1,000-person audience or about 1,540 invitations for a very large audience. That is why the page separates completed responses from invitations to send: the statistical target and the operational target are related, but they are not identical.

When a census is better than a sample

A sample size calculator is most useful when the audience is too large or too expensive to measure in full. If the population is tiny and reachable, a census can be simpler than sampling. Employee surveys for a 60-person team, membership checks for a small association, or classroom polls for one seminar section often fall into this category.

The finite population correction already points in that direction by shrinking the required completes when the total population is small. Once the corrected sample is close to the full audience, the operational convenience of surveying everyone may outweigh the effort of defending a smaller sample plan. In those cases the better planning question is often census versus sample, not which sample size to choose.