Back
Survey Margin of Error: Formula, Table and Examples

Matthieu Saussaye

"52% prefer concept A" means nothing until you know whether the real figure could just as easily be 47%.
The margin of error tells you how far a survey result can drift from the true value in the population, purely because you asked a sample instead of everyone. It decides how many respondents you need, which differences you are allowed to report, and which subgroups you should not slice. This guide gives you the formula, a ready-to-use table, worked examples with real numbers, and the traps that catch most teams. If you want to run the study itself quickly, SmartInterview drafts the questionnaire, recruits the sample and returns analysed results, usually in about 48 hours.
The Short Answer
For a proportion, the margin of error at 95% confidence is MOE = 1.96 × √(p(1−p)/n), where p is the observed share and n the number of respondents.
The worst case is p = 50%, which gives about ±9.8 points at n=100, ±4.9 at n=400 and ±3.1 at n=1,000. A quick rule: MOE ≈ 1/√n.
To halve the margin of error you need four times the sample, so going from ±5 to ±2.5 points costs 4x the respondents.
Every subgroup has its own, larger margin of error. 180 people under 30 in a sample of 1,000 carry about ±7.3 points, not ±3.1.
The margin of error only covers random sampling error. It says nothing about bad questions, non-response or a biased panel, and it is only strictly valid for probability samples.
Margin of Error by Sample Size and Confidence Level
Values below are for a proportion at p = 50% (the most conservative case) in a large population, in percentage points.
Sample size (n) | 90% (z = 1.645) | 95% (z = 1.96) | 99% (z = 2.576) | Typical use |
|---|---|---|---|---|
50 | ±11.6 | ±13.9 | ±18.2 | Directional only; qualitative signal |
100 | ±8.2 | ±9.8 | ±12.9 | Early pulse, large differences only |
200 | ±5.8 | ±6.9 | ±9.1 | Small concept test, one market |
400 | ±4.1 | ±4.9 | ±6.4 | The classic "±5 points" study |
600 | ±3.4 | ±4.0 | ±5.3 | Two or three subgroups to compare |
1,000 | ±2.6 | ±3.1 | ±4.1 | National poll, brand tracking |
2,000 | ±1.8 | ±2.2 | ±2.9 | Many subgroups, small shifts over time |
Two things stand out. First, the gains flatten fast: going from 100 to 400 respondents removes almost 5 points of error, going from 1,000 to 2,000 removes less than 1. Second, the confidence level matters as much as the sample: the same 400 respondents give ±4.1 at 90% and ±6.4 at 99%. Most market research reports 95%. How you pick the people matters too; see sampling methods.
How the Margin of Error Actually Works
Definition
If you could repeat the same survey many times with fresh random samples, each result would land slightly differently. The margin of error is the half-width of the range that contains the true population value in, for example, 95% of those repetitions. A result of 40% ±4 points at 95% confidence means the confidence interval runs from 36% to 44%, and the method that produced it captures the true value 95 times out of 100.
The formula, piece by piece
MOE = z × √(p(1−p)/n). The square root part is the standard error of a proportion. z is the critical value for your confidence level: 1.645 for 90%, 1.96 for 95%, 2.576 for 99%. p(1−p) peaks at 0.25 when p = 0.5, which is why pollsters quote the margin at 50%: it is the largest it can be. n sits under a square root, which is why precision gets expensive.
Worked example 1: one result, n = 400
You survey 400 people and 30% say they would switch brands. Standard error = √(0.30 × 0.70 / 400) = √0.000525 = 0.0229. Margin of error = 1.96 × 0.0229 = 0.045, so ±4.5 points. The 95% confidence interval is 25.5% to 34.5%. Note that it is narrower than the ±4.9 in the table, because 30% is further from 50%. At 5% (a rare answer) it would be only ±2.1 points.
Worked example 2: is the difference real?
Two groups of 500 rate a new pack: 42% of group A and 35% of group B say "very appealing". Each figure has a margin of about ±4.3 points, so the intervals overlap, and many people stop there. That is a mistake. The margin of error of a difference is 1.96 × √(0.42×0.58/500 + 0.35×0.65/500) = ±6.0 points. The gap is 7 points, larger than 6.0, so it is statistically significant at 95%. Overlapping intervals do not prove "no difference". This is exactly the comparison you make in monadic concept tests, where each cell sees one concept.
Worked example 3: means and NPS
For an average on a 5-point Likert scale, use MOE = z × s/√n. A mean of 3.8 with a standard deviation of 1.1 and n = 400 gives 1.96 × 1.1/20 = ±0.11. For Net Promoter Score, the variance is (promoters + detractors) − NPS², all as proportions. With 50% promoters and 20% detractors (NPS = 30) among 400 respondents: √((0.70 − 0.09)/400) = 0.039, × 1.96 = ±7.7 NPS points. Your quarterly NPS moving from 30 to 34 on 400 answers is noise.
Finite population correction
The standard formula assumes a very large population. When your sample is more than about 5% of the population, for example when you survey your own customer base, multiply the margin by √((N−n)/(N−1)). With 2,000 customers and 400 answers: √(1,600/1,999) = 0.895, so the ±4.9 becomes ±4.4. Working the other way, the sample needed for ±5 points falls from 385 to n = n₀ / (1 + (n₀−1)/N) = 384.2 / (1 + 383.2/2,000) = 323 respondents.
For the broader picture of which statistics fit which question, see our survey questions guide, the NPS calculator and how to read relationships with correlation in survey data.
How to Plan and Report the Margin of Error, Step by Step
1. Start from the decision, not the sample size
Write down the smallest difference that would change your decision. If you will launch concept A only if it beats B by 10 points, you do not need ±2 precision. If you track a brand metric that moves 2 to 3 points a year, you do.
2. Compute the sample for the margin you need
Rearrange the formula: n = z² × p(1−p) / E². For ±5 points at 95%: 1.96² × 0.25 / 0.05² = 384.2, round up to 385. For ±3 points: 1,068. For ±10 points: 97. Use p = 0.5 unless you have a solid prior estimate.
3. Size for the smallest subgroup you will report
If you will compare three age bands, each band needs its own sample. A total of 1,000 split 18% / 52% / 30% leaves 180 in the youngest band, with ±7.3 points. To report each band at ±5, plan about 385 per band and set quotas. Treat any subgroup under 100 as directional and say so in the report.
4. Adjust for weighting
Weighting a sample to match the population inflates the variance. Leslie Kish's design effect expresses this: effective sample = n / deff. A sample of 1,000 with a design effect of 1.4 behaves like 714 respondents, and its margin grows from ±3.1 to ±3.7. Pew Research Center publishes its margins including the design effect for this reason.
5. Add the correction for small populations
Surveying a known list (customers, employees, members) where the sample is more than 5% of the total? Apply the finite population correction from the section above. It is the one adjustment that lowers your margin.
6. Report it the right way
Write the sample size, the confidence level and the margin together: "n = 612, margin of error ±4.0 points at 95% confidence for the total sample; larger for subgroups." For an opt-in online panel, the American Association for Public Opinion Research (AAPOR) advises against quoting a classic margin of sampling error, since there is no random selection; describe it as a modelled or "credibility" interval, or report the precision as an approximation.
7. Doing it in SmartInterview
In SmartInterview, you describe the objective ("compare purchase intent for two pack designs among Swiss supplement buyers") and the AI drafts the full questionnaire with a quality score. You set the sample and quotas per subgroup, and they are enforced live, so your youngest band does not end at 60 people. Respondents come from your own customer list through a reusable link, tagged by source, or from an audience recruited in Switzerland and Europe. Open answers get an AI follow-up when they are vague, then are coded into themes with sentiment, and you read cross-tabs per subgroup or export to Excel or SPSS to compute the intervals. The first 50 respondents are free; plans start at CHF 45 per month for 150 respondents, see pricing.
Where I draw the line
I do not chase small margins on early product questions. When Pharmalp, a Swiss natural health brand, tested four new supplement concepts, the decision rested on which claim made people hesitate and what price felt fair, not on a 2-point gap. A few hundred well-targeted respondents with probed open answers told them more than 2,000 tick-boxes would have. Read the Pharmalp case study.
Where I am strict is on subgroups and on trend lines. I have seen teams celebrate a "+6 points among under-30s" that rested on 70 people, and quarterly dashboards that coloured every wobble green or red. If a difference is smaller than its margin of error, it does not go in the headline. And a tiny margin of error on a biased sample is still a biased number.
Matthieu Saussaye, Founder of SmartInterview.
Margin of Error Checklist
Define the smallest difference that matters before choosing n.
Pick the confidence level (95% by default) and write it down.
Compute n with n = z² × p(1−p) / E², using p = 0.5 if unsure.
Multiply by the number of subgroups you must report on their own, and set quotas.
Apply the finite population correction if you sample more than 5% of a known list.
Inflate n for weighting using the expected design effect.
Recompute the margin on the achieved sample, not the planned one.
Test differences with the formula for differences, not by eyeballing overlapping intervals.
Print n and the margin on every chart, including subgroup charts.
Common mistakes
Quoting the total margin for every subgroup. ±3.1 for 1,000 people becomes ±9.8 for the 100 who live in Ticino.
Treating the margin as a measure of quality. It ignores leading wording, order effects and non-response. See survey bias.
Concluding "no difference" because intervals overlap. Two results can overlap and still differ significantly.
Using the margin of a proportion for a mean or an NPS. Each statistic has its own standard error; NPS margins are roughly double those of a single proportion.
Reading a trend from noise. A 2-point move on n = 400 is inside the margin of error of each wave.
Ignoring low response rates. A big n from a small, self-selected slice of your list can be precise and wrong. More on falling response rates.
Presenting a classic margin of error for an opt-in panel as if it were a random sample.
Panels and their limits are covered in what panel research is; for analysis software see survey analysis tools.
Frequently Asked Questions
What is a good margin of error for a survey?
For most market research, ±5 points at 95% confidence (about 385 respondents) is the usual target. National polls aim for ±3 points (about 1,000). Early product tests can live with ±8 to ±10 points if you only need to spot large differences.
How do you calculate the margin of error?
Multiply the critical value for your confidence level (1.96 for 95%) by the standard error √(p(1−p)/n). For 400 respondents and a 50% result: 1.96 × √(0.25/400) = ±4.9 points.
What sample size gives a 5% margin of error?
About 385 respondents at 95% confidence in a large population, using p = 0.5. If your whole population is small, for example 2,000 customers, the finite population correction brings it down to about 323.
What is the margin of error for 1,000 respondents?
About ±3.1 points at 95% confidence for a result near 50%, ±2.6 at 90% and ±4.1 at 99%. For a result near 10% or 90% it drops to about ±1.9 points.
Does population size affect the margin of error?
Barely, once the population is large. A sample of 1,000 gives roughly the same precision for a city of 200,000 as for a country of 80 million. It only matters when the sample is a meaningful share of the population, more than about 5%.
What is the difference between margin of error and confidence level?
The confidence level is how often the method captures the true value (for example 95%). The margin of error is how wide the range is. Raising the confidence level with the same sample widens the margin.
Can you calculate a margin of error for an online panel?
You can compute the number, but strictly it assumes random sampling, which opt-in panels do not have. AAPOR recommends describing such figures as modelled or credibility intervals, and the bigger risk with panels is usually bias, not sampling error.
Want to size and run your next study without the spreadsheet? SmartInterview is free for your first 50 respondents.


