Surveys
How many people do you need to survey?
It depends on the margin of error you can live with, not on the size of the population. About 385 responses gives roughly ±5 percentage points, 97 gives ±10, and 1,068 gives ±3 — whether you are asking a town or a country. Small groups need fewer, and there is a correction for that.
The answer people do not expect
The usual instinct is that a bigger population needs a bigger sample — that you should survey some percentage of the people involved, ten per cent perhaps. That is not how it works.
Once a population is reasonably large, the number you need depends almost entirely on how precise you need the answer to be, and barely at all on how many people there are. Around 385 responses gives you roughly ±5 percentage points whether you are asking a town of ten thousand or a country of fifty million. That is why national polls report samples of one or two thousand and are not embarrassed about it.
So the question to settle first is not “how many people are there?” but “how wrong am I willing to be?”
How many for the margin you want
These are the standard figures for a 95% confidence level, calculated at the worst case for precision — a result near 50/50, where a proportion is hardest to pin down.
| Margin of error | Responses needed | What that buys you |
|---|---|---|
| ±20 points | 25 | Almost nothing. A 60% result spans 40–80% |
| ±15 points | 43 | A rough impression only |
| ±10 points | 97 | Enough to see a large difference |
| ±7 points | 196 | Enough for most internal decisions |
| ±5 points | 385 | The usual target for a published figure |
| ±3 points | 1,068 | Polling standard |
Read the first row before the others. Twenty-five responses is a perfectly normal size for an informal survey, and it tells you almost nothing: a 60% result is consistent with anything from 40% to 80%, which includes “most people disagree”. If you take one number from this page, make it that one.
Small groups need fewer
The figures above assume the population is large. When it is not — a club of fifty, a team of thirty, a school year of two hundred — you need proportionally fewer, because each response covers more of the whole. This is the finite population correction, and it is the one case where population size genuinely matters.
| Population | Responses needed | Share of the group |
|---|---|---|
| 30 | 28 | 93% |
| 50 | 45 | 90% |
| 100 | 80 | 80% |
| 200 | 132 | 66% |
| 500 | 218 | 44% |
| 1,000 | 278 | 28% |
| 5,000 | 357 | 7% |
| 100,000+ | 385 | under 1% |
The practical reading: for a small group, precision means asking nearly everybody. A committee of thirty needs twenty-eight responses for a ±5 point answer, at which point you may as well ask all thirty and stop estimating. For groups under about fifty, a survey is usually the wrong instrument — a census is easier and exact.
Diminishing returns
The margin shrinks with the square root of the sample, which is why the table climbs so steeply at the precise end.
- 97 → 385 responses, four times the work, halves the margin from ±10 to ±5.
- 385 → 1,068, nearly three times again, only takes it from ±5 to ±3.
So the honest planning question is where on that curve the answer stops changing your decision. Going from 25 responses to 100 transforms what you know. Going from 400 to 1,000 usually does not change what you would do, and the effort is better spent on who you asked.
The bigger threat is who you ask
Every number above assumes a random sample: everyone in the group you want to describe had an equal chance of being counted. Almost no informal survey manages this, and when it is missed, no sample size repairs it.
A survey of whoever walks past a table is a convenience sample. It describes people who were in that place, at that time, and willing to stop. That is a real group, but it is not the group you meant, and the gap between them does not shrink as you collect more responses — it is bias, not noise, and bias does not average out. Two thousand responses collected badly are worse than two hundred collected well, because the bad ones come with false confidence.
Three that catch people out:
- Self-selection. People with strong opinions answer; the indifferent walk on. An open online form is the extreme case.
- Timing and place. Whoever is in that corridor at 3pm is not whoever is there at 9am.
- The question itself. A leading question moves the answer further than any sample size will, and asking it slightly differently once you are tired of asking it introduces a drift nothing can correct for.
When to stop counting
Decide the number before you start, not while you are collecting. Choosing to stop once the result looks convincing is how a survey becomes a way of confirming what you already believed.
- Decide what margin would change your decision. If you would act the same way at 55% and at 65%, you do not need ±5 — ±10 will do, and that is a quarter of the work.
- Look up the number, and write it down.
- Decide how you will choose people, and be honest about whether it is random.
- Collect to the number, then stop.
- Report the count, the method and the margin together. A percentage travelling on its own loses all three.
Frequently asked questions
How many people do I need to survey?
About 385 for a margin of error of five percentage points, 97 for ten points and 1,068 for three, at 95% confidence. These hold for any large population, so the size of the group barely matters once it is more than a few thousand.
Does sample size depend on population size?
Hardly at all, once the population is reasonably large. A sample of 385 gives about five points either way whether you are surveying a town or a country. Population size only matters for small groups, where the finite population correction reduces what you need.
How many people should I survey in a group of 100?
About 80 for a margin of five points. Small groups need a large share of themselves: 45 out of 50, or 28 out of 30. Below about fifty people it is usually easier to ask everyone and have an exact answer than to sample.
Is 100 responses enough for a survey?
It gives roughly ten percentage points either way, which is enough to see a large difference and not enough to resolve a close one. A 55% result at that sample size is consistent with 45%, so it does not establish a majority.
Is 10% of the population a good rule?
No, and it misleads in both directions. Ten per cent of a country is absurdly more than you need, and ten per cent of a club of thirty is three people, which tells you nothing. The margin you need is what sets the number.
Why does doubling my sample not halve the error?
Because the margin falls with the square root of the sample size. Quadrupling the responses halves the margin: 97 to 385 takes you from ten points to five, and it takes 1,068 to reach three.
What if my sample is not random?
Then the margin of error does not describe the population you meant, and collecting more responses does not fix it. A convenience sample carries bias rather than noise, and bias does not average out. Report the count and how you collected it instead.
When should I stop collecting responses?
At the number you decided before starting. Stopping when the result looks convincing turns a survey into a way of confirming what you already thought, and it is the easiest way to get a confident wrong answer.