The standard error calculator above answers a question that sits underneath almost every statistic you will ever report: how much would this number move if you had drawn a different sample? Feed in a standard deviation and a sample size, a column of raw values, or a count of successes out of a total, and it returns the standard error along with the sample spread, the matching margin of error, and the relative standard error as a percentage.
Arb Digital's analysts use this number as a reality check before any figure goes into a client report. A conversion rate of 4.2% from 500 sessions and a conversion rate of 4.2% from 50,000 sessions are the same number on the page and completely different pieces of evidence. The standard error is what tells them apart, and it takes about ten seconds to work out.
What This Standard Error Calculator Does
The standard error is the standard deviation of a sampling distribution. That sentence is compact enough to be unhelpful, so here is the longer version. Imagine repeating your study hundreds of times, each time drawing a fresh sample of the same size from the same population and recording the mean. Those means would not all be identical — they would form their own distribution, clustered around the true population mean. The standard deviation of that imaginary distribution of means is the standard error of the mean, usually shortened to SEM.
You never actually run the study hundreds of times, which is the elegant part. A single sample is enough to estimate the standard error, because the maths connects it directly to the sample's own standard deviation and its size. This calculator handles the two cases that cover the vast majority of real work: a mean measured on a continuous scale, and a proportion measured as a count of yes/no outcomes. It also applies a finite population correction if you tell it the population size, which matters when your sample is a meaningful fraction of everything there is.
One boundary worth stating plainly, because the two get confused constantly: our margin of error calculator multiplies the standard error by a critical value to produce the half-width of a confidence interval, while this page returns the standard error itself — the raw ingredient that every interval, t-statistic and z-statistic is built from.
How to Use It
- Choose what you are measuring. A mean of something continuous — height, revenue per order, time on page — uses the SD-based formula. A proportion — click-through rate, defect rate, share who answered yes — uses the proportion formula.
- Enter your numbers. If you already have the standard deviation from software output, type it in with the sample size. If you have the raw values, switch to raw-data mode and paste the column; the tool computes the sample SD with the n − 1 denominator for you.
- Set the population size only when it is genuinely finite and small. Surveying 400 of a 1,000-member association is a case where the correction matters. Surveying 400 website visitors out of an unbounded stream is not.
- Pick a confidence level. This affects only the margin-of-error figure in the grid, not the standard error itself, which carries no confidence level at all.
- Read the bars underneath. They show what the standard error would be at half, one and four times your sample size, which is usually a more persuasive argument for a bigger sample than any amount of explanation.
The Formula / How It's Calculated
For a mean, SE = s ÷ √n, where s is the sample standard deviation and n is the number of observations. Work through an example. A sample of 25 students has a mean test score with a standard deviation of 15 points. The standard error is 15 ÷ √25 = 15 ÷ 5 = 3.0. At 95% confidence the margin of error is 1.96 × 3.0 = 5.88 points, so the interval runs roughly 5.9 points either side of the sample mean. This decomposition is set out formally in the NIST/SEMATECH e-Handbook section on confidence limits for the mean, where the s/√N term appears inside the interval formula.
For a proportion, SE = √[p(1 − p) ÷ n], where p is the sample proportion x ÷ n. Suppose 200 of 500 visitors converted. Then p = 0.40, and SE = √(0.40 × 0.60 ÷ 500) = √0.00048 = 0.02191, or about 2.19 percentage points. Notice there is no separate standard deviation to supply — for a binary outcome, the spread is completely determined by the proportion itself, which is why a proportion needs only two numbers.
The finite population correction multiplies the result by √[(N − n) ÷ (N − 1)]. Sample 400 people from a population of 1,000 and the factor is √(600 ÷ 999) = 0.775, shrinking the standard error by nearly a quarter. Sample 400 from a population of a million and the factor is 0.9998, which is why the correction is normally ignored. The rule of thumb is to apply it when your sample exceeds about 5% of the population. The choice of s in the first place is worth thinking about too — our standard deviation calculator shows the difference between the sample and population versions, and the sample version is the one that belongs here.
Standard Error Is Not Standard Deviation
This is the single most common confusion in applied statistics, and it survives into published papers. The standard deviation describes the spread of individual observations. The standard error describes the precision of a summary statistic computed from those observations. They answer different questions, they have different magnitudes, and they behave differently as you collect more data.
Collect more data and the standard deviation does not systematically shrink. If adult heights vary with an SD of 7 cm, measuring ten thousand people instead of a hundred gives you a better estimate of that 7 cm — it does not make people more similar in height. The standard error, by contrast, shrinks towards zero as n grows, because you are becoming more and more certain about where the average sits.
The consequence shows up in charts. Standard deviation bars describe how varied your subjects are; standard error bars describe how confident you are about the mean, and they are always narrower. A figure that quietly switches between them looks far more convincing than the data warrants, so state which you used in the caption. To describe the spread of the data rather than the precision of its average, use the variance calculator and the descriptive statistics calculator.
Why the Square Root Makes Data Expensive
The n sits under a square root, and that single detail governs the economics of nearly every study ever run. Halving your standard error requires four times the data. Cutting it to a tenth requires a hundred times the data. There is no way around this through cleverness in the analysis; it is a property of random sampling itself.
Run the numbers on a marketing example. A landing-page test measuring a 4% conversion rate has a standard error of about 0.62 percentage points at n = 1,000. Push to n = 4,000 and it falls to roughly 0.31 points. Push to n = 16,000 and it reaches about 0.155 points. Each step costs four times the traffic for the same absolute improvement in precision. This is exactly why the returns on collecting more data flatten out, and why the honest answer to "can we get more precision?" is often "yes, and it will cost sixteen times as much."
The corollary is more useful than the warning. If your standard error is close to acceptable, a modest increase in n may finish the job. If it is three times too large, you need nine times the data and should redesign the study — better measurement, tighter inclusion criteria, or a paired design — rather than run for longer. Our sample size calculator works the same relationship in reverse.
When z Is the Wrong Multiplier
The margin of error shown in the grid uses a normal-distribution critical value: 1.645 for 90%, 1.96 for 95%, 2.576 for 99%. That is correct for large samples and for proportions with a reasonable number of successes, but it is optimistic for small samples of a mean.
When you estimate the standard deviation from the same small sample you are analysing, that estimate carries its own uncertainty, and the t-distribution with n − 1 degrees of freedom accounts for it with fatter tails. At n = 25 the 95% t-value is about 2.064 rather than 1.96, so the true margin of error in the worked example above is nearer 6.19 than 5.88 — a difference of about 5%. At n = 10 the gap widens to roughly 15%. At n = 100 it is under 2% and stops mattering. The standard error itself is unchanged in every case; only the multiplier moves. Use the t-test calculator or the confidence interval calculator when n is small enough for the distinction to bite.
Where the Proportion Formula Breaks Down
The p(1 − p) formula rests on a normal approximation to the binomial distribution, and that approximation fails at the extremes. The usual guideline is that both np and n(1 − p) should be at least 5, and preferably 10. With 3 conversions out of 400, np = 3 and the approximation is unreliable — the resulting interval can even extend below zero, which is impossible for a proportion.
Two things follow. First, if your observed proportion is 0 or 1, this formula returns a standard error of exactly zero, which is obviously wrong: zero conversions from 50 visitors is not proof of a 0% rate. Second, for rare events you should use an exact or adjusted method rather than the plain formula. The calculator flags the low-count case in the sub-line rather than silently returning a number you should not use.
There is a related quirk worth knowing: p(1 − p) is largest at p = 0.5, where it equals 0.25. That is why sample-size planning for surveys assumes 50% unless there is good reason not to — it is the most pessimistic, and therefore safest, assumption. A proportion near 5% or 95% has roughly half the standard error of one near 50% at the same sample size.
Reading the Relative Standard Error
The grid reports the relative standard error, which is the standard error expressed as a percentage of the estimate itself. National statistics agencies use it as a publication threshold: estimates with a relative standard error above roughly 25% or 30% are commonly flagged as unreliable or suppressed entirely, because the noise has swamped the signal.
It is a useful sanity check because it is unitless. A standard error of 3.0 means nothing alone — 3.0 on a mean of 2,000 is precision, 3.0 on a mean of 8 is noise. Two cautions: it is meaningless when the estimate is near zero, and it is not usefully defined for quantities that can go negative.
Arb Digital builds free, no-signup calculators for statistics, finance, health and construction. Browse the full library, or get in touch if there is a calculation your team keeps doing by hand.
Browse All Free Tools Contact Arb DigitalCommon Mistakes to Avoid
- Dividing by n instead of √n — the square root is the whole point, and dropping it understates the error dramatically at any realistic sample size.
- Using the population standard deviation formula on sample data — the n − 1 denominator corrects a known downward bias, and this tool applies it automatically in raw-data mode.
- Labelling standard error bars as standard deviation, or leaving the caption blank — readers cannot tell them apart visually, and the difference changes what the chart claims.
- Applying the finite population correction to an open-ended population — website visitors, future customers and ongoing production runs have no fixed N, so leave the field at zero.
- Trusting the proportion formula at extreme rates — with fewer than about ten successes or ten failures, the normal approximation behind it is not sound.
Related Free Tools From Arb Digital
Turn the standard error into an interval with the confidence interval calculator, plan a study around a target precision with the sample size calculator, check whether a difference could be chance with the p-value calculator, and standardise a single observation with the z-score calculator. If you are sizing a study around detectability rather than precision, the statistical power calculator is the companion piece. The full free online tools hub holds the rest of the statistics set. NIST also publishes a clear reference on the underlying measures of scale that feed into the standard deviation this calculation depends on.
Frequently Asked Questions
Standard deviation describes how spread out the individual observations are. Standard error describes how precisely you have estimated a summary statistic such as the mean. Standard error equals the standard deviation divided by the square root of the sample size, so it is always the smaller of the two for any sample beyond one observation.
Divide the sample standard deviation by the square root of the sample size. For a sample of 25 values with a standard deviation of 15, the standard error is 15 divided by 5, which equals 3.0.
There is no universal threshold, because the number carries the units of your measurement. Compare it against the size of the estimate itself using the relative standard error. Statistical agencies commonly treat a relative standard error above 25% to 30% as too unreliable to publish.
Yes, but only in proportion to the square root of the sample size. Quadrupling your sample halves the standard error. Cutting the standard error to one tenth requires one hundred times the data.
Take the square root of p times one minus p, divided by n, where p is the number of successes divided by the sample size. With 200 successes from 500 trials, p is 0.40 and the standard error is about 0.0219, or 2.19 percentage points.
Apply it when your sample is more than about 5% of a genuinely fixed population, such as 400 members of a 1,000-person association. For open-ended populations like website visitors there is no fixed total, so the correction does not apply.
Mathematically yes, if every value in the sample is identical or if a proportion is exactly 0 or 1. In practice a zero standard error usually signals too little data rather than perfect precision, and for extreme proportions you should use an exact method instead.
Use z for large samples and for proportions with plenty of successes and failures. Use t with n minus 1 degrees of freedom when estimating a mean from a small sample, because the standard deviation itself is uncertain. The difference is around 5% at n equals 25 and negligible beyond n equals 100.
This calculator is provided for education and planning only. Statistical output should be read alongside a plot of your data and, for consequential decisions, reviewed by someone qualified in your field.
