How to Calculate Survey Sample Size

"How many people do I need to survey?" is one of the most common research questions - and one of the most misunderstood. The right sample size is not a fixed percentage of your audience; it depends on how precise and how confident you need your results to be. This guide walks through the underlying concepts, the actual formula, and worked examples so you can calculate a defensible sample size for any survey.

Why Sample Size Matters

You survey a sample because surveying everyone is usually impractical. But a sample only approximates the truth, and the size of that approximation error depends on how many people you ask. Too few responses and your results are too noisy to act on; too many and you waste time and money for precision you do not need. Calculating sample size deliberately gives you exactly the confidence your decision requires - no more, no less.

Crucially, what matters statistically is the absolute number of responses, not the fraction of your population. Surveying 384 people gives roughly the same precision whether your population is 20,000 or 20 million. This counterintuitive fact is why national polls of a few thousand people can describe entire countries, and why a rule like "survey 10% of my users" is statistically meaningless - it would demand absurdly large samples for big populations and dangerously small ones for tiny populations.

It helps to separate two ideas that beginners often conflate. Sample size controls how precise your estimate is - the random scatter around the true value. It does not control bias, which is a systematic skew from a flawed frame or selective non-response. A sample of 50,000 drawn from a biased frame is just a very precise measurement of the wrong thing. Calculating sample size correctly is necessary, but it only buys you precision; representativeness comes from good sampling and high response rates. Keep that distinction in mind as you work through the formula below.

The Three Inputs: Confidence, Margin, Population

Confidence level is how sure you want to be that the true population value falls within your stated range. Common choices are 90%, 95%, and 99%. A 95% confidence level means that if you repeated the survey many times, about 95% of the resulting confidence intervals would contain the true value. Each confidence level maps to a z-score: 1.645 for 90%, 1.96 for 95%, and 2.576 for 99%.

Margin of error (also called the confidence interval) is the precision of your estimate, expressed as plus or minus a percentage. If 60% of respondents pick an option with a 5% margin of error, the true population value is likely between 55% and 65%. Smaller margins require larger samples.

Population size is the total number of people in the group you are studying. It only meaningfully affects the calculation for small, finite populations; for large populations its influence is negligible, which is why many calculators ignore it once the population exceeds about 20,000.

A fourth, hidden input is the population proportion (p) - the expected split of responses. When you do not know it, use 0.5 (a 50/50 split), because that produces the largest required sample and therefore the most conservative, safe estimate.

The Sample Size Formula

For a large or unknown population, the required sample size is:

n = (z² × p × (1 - p)) / e²

where z is the z-score for your confidence level, p is the expected proportion, and e is the margin of error as a decimal. For a finite population of size N, apply the finite population correction:

n_adjusted = n / (1 + ((n - 1) / N))

The correction always reduces the required sample, and the reduction is large only when the population is small relative to n.

Worked Examples

Example 1 - large population. You want 95% confidence and a 5% margin of error, with an unknown split so p = 0.5. Then z = 1.96, so n = (1.96² × 0.5 × 0.5) / 0.05² = (3.8416 × 0.25) / 0.0025 = 0.9604 / 0.0025 = 384.16. You need about 385 completed responses. This is the famous "384" figure behind many national surveys.

Example 2 - tighter precision. Keep 95% confidence but demand a 3% margin of error. Now n = (3.8416 × 0.25) / 0.03² = 0.9604 / 0.0009 = 1067.1, so you need about 1,068 responses. Halving the margin roughly quadruples the sample - precision is expensive.

Example 3 - finite population. Suppose you have only 2,000 customers and want the Example 1 precision (n = 385 before correction). Apply the correction: n_adjusted = 385 / (1 + (384 / 2000)) = 385 / 1.192 = 323. You need about 323 responses from your 2,000 customers. This is common for a focused NPS survey sent to an existing customer base.

Adjusting for Response Rate

The numbers above are completed responses, not invitations. If you expect a 20% response rate, divide the target by 0.20 to find how many people to invite. For 385 completes at a 20% rate, you must invite 385 / 0.20 = 1,925 people. Realistic response rates vary widely - email surveys to engaged customers may reach 20-40%, while cold outreach often falls below 5% - so always confirm your historical rate before committing to a list size. Sending reminders and keeping the survey short are the most reliable ways to lift completion.

Common Mistakes to Avoid

First, do not confuse sample size with response count needed for subgroups: if you plan to analyze segments separately, each segment needs its own adequate sample, which can multiply your total. Second, do not assume a bigger sample fixes bias - sample size addresses random sampling error only, not systematic bias from a flawed frame or non-response. Third, avoid "percentage of population" rules of thumb like "survey 10%"; they over-sample large populations and under-sample small ones. Finally, remember that an enormous sample can produce statistically significant differences that are too small to matter in practice; always judge effect sizes, not just significance.

When you research a defined market segment - for example, a market research survey aimed at SaaS startups - decide your subgroup cuts before you field so you can size each one correctly. A useful planning habit is to write down, before launch, every comparison you intend to make in the final report: "free versus paid," "new versus tenured," "region A versus region B." Each of those comparisons implies two subgroups that each need an adequate sample. If you discover at planning time that a small but important segment will only yield 40 responses, you can oversample it deliberately rather than ending up unable to say anything about it.

It also pays to budget for data cleaning. Not every submitted response is usable: some respondents straight-line through a grid, some fail attention checks, some abandon halfway. If you expect to discard, say, 10% of completes as low quality, inflate your target by that amount so your clean sample still meets the precision you calculated. Combining the response-rate adjustment with a cleaning buffer gives you a realistic invitation list rather than an optimistic one that leaves you short when the data arrives.

Frequently Asked Questions

How many survey responses do I need to be statistically valid? For most studies at 95% confidence and a 5% margin of error, about 385 completed responses are enough when the population is large. Tighter margins or subgroup analysis require more.

Does population size really not matter? For large populations it barely matters - 385 responses give roughly the same precision whether the population is 50,000 or 5 million. Population size only changes the result meaningfully for small, finite groups, where the finite population correction reduces the required sample.

What confidence level and margin of error should I use? A 95% confidence level with a 5% margin of error is the standard default for business surveys. Use 99% confidence or a 3% margin only when a high-stakes decision justifies the much larger sample.

Why use p = 0.5 when I do not know the split? Because p × (1 - p) is maximized at p = 0.5, using it produces the largest, safest sample size. If you genuinely expect a lopsided split, a value like 0.2 or 0.8 lowers the required sample.