The number of survey responses you need depends on what you want to estimate and how precise the estimate must be. For a two-choice proportion, a simple random sample, a large population, a 95% confidence level, a ±5 percentage-point margin of error, and an assumed proportion of 50%, the standard approximation is about 385 valid completed responses. If one in four invitees responds, you would plan to invite about 1,540 people. These are estimates under stated assumptions, not a universal rule or a guarantee of representativeness.
The response target and the invitation count are different numbers. Define the people you want to learn about, the measure you plan to report, and how you will invite them before calculating either number. This guide explains the inputs, a worked example, a finite-population adjustment, and the limits of sample-size formulas. The examples assume a simple binary outcome unless noted.
Start by defining what the survey will estimate
Write down, in one sentence, whose experience you want to describe and what you want to measure. For example: “Estimate the share of customers who were satisfied with delivery among customers whose orders shipped from July through September.” That sentence identifies a target population, a period, and a measure. “Collect customer feedback” does not define any of those clearly enough to choose a sample size.
The target population might be every customer, customers who purchased in a particular period, users of a specific store, or people who attended an event. The number of eligible people may be known from a list, roughly estimated, or completely unknown. Record which is true. If you do not know the population size, you can begin with the large-population approximation and label the result as such rather than inventing a precise denominator.
The measure matters too. Estimating the proportion who would answer “yes” is not the same task as estimating an average rating, detecting a difference between two campaigns, or comparing several regions. The quick calculation in this guide applies to a two-choice outcome when you are estimating a population proportion. A study of means, treatment effects, several primary outcomes, or a complex sample needs a design suited to that purpose.
The University of Florida IFAS Extension guide explains that proportion sample-size calculations assume a two-choice outcome and simple random sampling. More complex designs, such as stratified or cluster samples, must also account for variation within the relevant groups. Check whether the calculation’s assumptions match your survey.
Four inputs shape a simple sample-size estimate
For an introductory estimate of a binary proportion, decide these inputs:
| Input | What it means | What changes the target |
|---|---|---|
| Confidence level | A property of the method across repeated samples | A higher level generally requires more responses |
| Margin of error | The half-width you are willing to allow around an estimated proportion | A narrower margin requires substantially more responses |
| Expected proportion | The prior estimate of the share answering one way | If unknown, 50% is a conservative planning choice for this calculation |
| Population size | The number of eligible people in the population | A finite population can reduce the adjusted target when the sampling assumptions hold |
A 95% confidence level does not mean that 95% of individual answers are true. It describes the long-run behavior of an interval procedure: over repeated samples using the same method, the intervals would contain the true population proportion at the stated rate. A ±5 percentage-point margin is also not a guarantee that every kind of survey error is at most five points.
The expected proportion can come from a comparable earlier study. If there is no reasonable prior estimate, setting it to 50% produces the largest target under this simple formula because the product p × (1 − p) is largest at one-half. If you already know that a rare outcome is expected, a 50% assumption may be much more conservative than needed. Record the assumption so another person can reproduce the planning estimate.
The acceptable margin should follow the decision. A directional internal check may not need the same precision as a public estimate used to allocate a large budget. Making the margin smaller can increase the required count quickly: halving the margin roughly quadruples the unadjusted sample size. Do not choose a tight margin just because a calculator offers one; decide whether the resulting time, budget, and eligible population make it practical.
A worked example: 95% confidence and ±5 points
For a large population and a two-choice proportion, the basic approximation is:
n₀ = Z² × p × (1 − p) ÷ e²
Here, Z is the value associated with the chosen confidence level, p is the expected proportion, and e is the margin of error expressed as a decimal. For a 95% level, Z is approximately 1.96. With p = 0.50 and e = 0.05:
n₀ = 1.96² × 0.50 × (1 − 0.50) ÷ 0.05²
= 384.16
Round up to approximately 385 valid responses. The result is not a target of 385 invitations. It is an illustrative target for completed observations that meet your analysis rule. If some responses are incomplete, ineligible, or duplicates, your analysis may use fewer than the number of submissions.
Change the confidence level, margin, or prior proportion and the result changes. A 99% confidence level uses a larger Z value and therefore needs more observations. A ±3-point margin requires more responses than a ±5-point margin. A known prior proportion far from 50% can reduce the estimated number under the same simple model. State each input in the study plan; reporting only “n = 385” hides the conditions behind that number.
The University of Florida IFAS Extension guide to determining sample size gives the large-population proportion formula in Equation 1 and the finite-population correction in Equation 3. This compact formula is for learning and initial planning. A formal study, a decision about statistical power, a stratified design, or several primary comparisons may need other methods and professional review.
Adjust when the population is finite and relatively small
When you sample without replacement from a known, finite population, a finite-population correction can reduce the target compared with assuming an indefinitely large population. One common adjustment to the initial estimate is:
n = n₀ ÷ (1 + (n₀ − 1) ÷ N)
N is the number of eligible people. Using n₀ = 384.16 and a fictional population of 1,000 gives an adjusted estimate of about 278. A population of 500 gives about 218; a population of 10,000 gives about 370. The adjustment becomes smaller as the population grows.
This example inherits the simple random sample and binary-proportion assumptions. It is not a shortcut for convenience sampling, a public link with unknown visitors, clustered sampling, or a census in which everyone is invited. If the number of eligible people is unknown, do not guess just to use the adjustment. Report the large-population estimate and its limitation. If there are only a few eligible people and you can contact all of them, inviting everyone and reporting who did or did not respond may be more useful than applying a sample formula.
Convert completed responses into an invitation count
The target is the number of valid completes you need for the analysis, not the number of emails you send. Estimate a response rate using a comparable previous survey, then divide the response target by that expected rate:
planned invitations = required valid responses ÷ expected response rate
If the target is 385 valid responses and you expect 25% of invitees to respond, the plan is about 1,540 invitations. If only 1,000 eligible people are available, the same response-rate assumption yields about 250 completes. That is a planning estimate, not a promise; the actual number can be higher or lower.
Use a response-rate definition whose denominator matches your process. The survey response-rate guide explains how to define the invited population and report the result consistently. A public survey URL may have no list of people who received an invitation. Page views are not automatically an invitation denominator, so do not divide responses by page views and call the result a standard invitation-based response rate.
Once you have a response target and audience, draft the form and check its preview. After publishing, submit a test response yourself and inspect the saved record before inviting people. FORMLOVA Free supports creating and publishing a form, collecting responses, searching responses, and basic analytics. Sample-size assumptions and respondent representativeness remain part of survey design.
Draft a survey around your response target
Draft your questions, check the wording and completion screen in preview, then publish and check a saved test response before sharing the form with your audience.
Sign up free to draft a survey
After registering, verify your email and connect a supported AI chat. Enter your questions and choices to create a private draft and review its preview. Preview test submissions are not saved. After publishing, submit a test response yourself and inspect the saved record before inviting people. FORMLOVA does not calculate sample size, select a random sample, or recruit respondents. The AI chat client has its own terms and costs.
For question wording, see the survey design guide. The form creation hub explains the wider process, and the survey form and analysis guide covers analysis after publication.
Plan subgroup comparisons separately
If you want to compare regions, age bands, customer types, or event groups, a total response target does not tell you how precise each subgroup’s estimate will be. A survey with 400 responses overall may still have only 18 responses in one region. That smaller cell has much greater uncertainty than the overall result.
Define which comparisons are important before invitations go out. Consider how many valid responses are needed in each group, whether groups should be sampled at different rates, and whether you will combine small groups for reporting. The more categories you cross, such as region by age by customer type, the fewer responses are likely to fall in each cell.
Ask only for attributes needed for a stated decision. If age bands are enough, do not collect full birth dates without a reason. Offer an appropriate way not to answer sensitive or optional questions. The survey age-question guide covers choosing age bands and writing that question. These design choices affect who responds and how safe the information is to collect; a sample-size formula does not decide them for you.
For a subgroup comparison, the right calculation may depend on the smallest difference you want to detect, expected proportions in each group, allocation, and analysis method. Estimating a single overall proportion and testing a difference are different statistical tasks. If the comparison will support a consequential decision, get a design review rather than applying the 385 example separately and assuming the study is powered.
A large response count does not remove bias
Reaching the calculated number does not automatically make results representative. An open social-media link may attract people who use that platform or feel strongly about the topic. Inviting only customers whose details are current may exclude people who stopped using the product. A very large convenience sample can estimate the opinions of its respondents precisely while still missing the views of the broader population.
Record how you built the eligible list, how invitations were sent, the collection dates, duplicate rules, exclusions, and any people who could not be contacted. Compare respondents with the target population on attributes you actually know. If a group is underrepresented, describe that limitation. Weighting can sometimes address known differences under additional assumptions; it does not erase every type of selection bias or nonresponse bias.
Statistics Canada’s archived glossary of survey terms distinguishes sampling error, which comes from studying a sample rather than the whole population, from other sources of error. Question misunderstanding, inaccurate recall, data-entry errors, ineligible responses, and nonresponse need their own checks. A sample-size calculation estimates one part of uncertainty under its model; it does not audit the full survey.
Keep a short sample-size planning record
A one-page note makes the estimate easier to review after data collection:
| Field | Fictional example |
|---|---|
| Objective | Improve delivery experience after a purchase |
| Population | Customers whose orders shipped in a defined quarter |
| Measure | Share who report being satisfied with delivery |
| Sampling and invitation method | Email eligible customers from a documented list |
| Calculation assumptions | Binary outcome, 95%, ±5 points, p = 0.50, population about 1,000 |
| Required valid responses | Approximately 278 after finite-population adjustment |
| Invitation target | Calculate separately from a comparable response-rate estimate |
| Exclusions and duplicates | Define before analysis and report actual counts |
The 278 figure belongs to this fictional example only. Substitute your actual population, response-rate evidence, question, and sampling design. Keep planned assumptions separate from measured results. After data collection, report the actual invitations, valid completes, exclusions, and response rate using the same definitions you set in advance.
What to remember
There is no single response count that is sufficient for every survey. Define the population and measure, then choose the precision and confidence assumptions that match the decision. Under a simple random sample of a large population, a binary proportion with 95% confidence, ±5 percentage points, and p = 0.50 needs approximately 385 valid responses. A smaller finite population may allow an adjustment; expected response rate helps turn completes into invitations.
The number is only as useful as the sampling method behind it. A larger sample cannot fix a biased invitation list or unclear questions. Record who was eligible, who was invited, who responded, what you excluded, and which assumptions you used. When the plan is ready, draft and preview the form, then publish it. Submit a test response yourself and check that it was saved before sharing the form with your audience.
Frequently asked questions
How many responses are enough for a survey?
There is no universal count. It depends on the decision, the population, the measure, the sampling method, the desired precision, and whether you need subgroup estimates. A quick estimate for a binary proportion is not a general answer for averages, experiments, or complex designs.
Does a 95% confidence level mean 95% of responses are correct?
No. It describes the long-run coverage of an interval method over repeated samples under its assumptions. It does not certify individual answers or guarantee that the sample is free of selection, measurement, or nonresponse bias.
Will 385 responses always give a ±5-point margin of error?
No. The approximate count assumes a large population, a simple random sample, a binary proportion, a 95% confidence level, and an assumed p of 0.50. Subgroups, biased selection, complex designs, and other measures change the calculation or its interpretation.
Is the response target the same as the number of people I invite?
No. The response target counts usable completed answers. Divide it by a realistic response-rate estimate to plan invitations. If you need 385 valid responses and expect a 25% response rate, the planning estimate is about 1,540 invitations. Actual results can differ.
Can a free form guarantee a statistically valid sample?
A form collects answers; it does not determine how respondents were selected or whether they represent the target population. FORMLOVA Free can be used to create and publish a form and collect responses. Sample design, representativeness, and the validity of the statistical assumptions remain separate decisions.
Sources
- University of Florida IFAS Extension: Determining Sample Size
- Statistics Canada: Appendix D, Glossary of survey terms (archived)
Disclosure and Verification
I reviewed these statistical references on October 18, 2026. The numerical examples are fictional and use a simplified binary-proportion model; they are not survey results or a guarantee of precision for a particular study.
