Standard Error Calculator

Two numbers that get printed the same way

This is the single most common error in reported results, and it is usually not deliberate.

The standard deviation is about the data. It says how far individual observations sit from the mean. If adult heights have an SD of 7 cm, that is a fact about people: most are within 7 cm of average. Measure ten thousand more and the SD stays around 7 — you have described the population better, not made people more alike.

The standard error is about the estimate. It says how far your sample mean is likely to sit from the true mean. Collect more data and it shrinks, without limit, because you are pinning down one number more precisely.

The ratio between them is exactly √n. At n = 4 the SE is half the SD. At n = 100 it is a tenth. At n = 10,000 it is a hundredth — and the SD has not moved at all.

So “mean ± 2.1” is genuinely ambiguous, and the two readings can be an order of magnitude apart. A paper reporting “42.3 ± 1.2” from 100 subjects is saying either that most subjects fall between 41.1 and 43.5 (SD), or that the average is pinned to that range while individuals vary between roughly 30 and 54 (SE). Those are different findings.

The rule: say which one you used, every time. A caption reading “mean ± SD” or “mean ± SEM, n = 12” costs four words and removes the ambiguity entirely.

The error bar problem

Error bars are where the confusion does real damage, because the choice changes the picture without changing the data.

Bars at ±1 SEM are √n times shorter than bars at ±1 SD on the same measurements. At n = 25 that is a factor of five. A chart with SEM bars looks dramatically tighter and more convincing than the identical chart with SD bars, and nothing about the underlying variability has changed.

Three conventions are in common use and they are not interchangeable.

±1 SD shows how variable the observations are. Use it when the spread of the data is the point — when a reader wants to know what a typical individual looks like.

±1 SEM shows how precisely the mean is known. It is the most common choice in biology and psychology, and also the most criticised, because ±1 SEM corresponds to only about a 68% interval — noticeably narrower than the 95% most readers assume they are looking at.

A 95% confidence interval is the least ambiguous, because it comes with a stated interpretation. It is roughly twice the SEM for a reasonable sample size, and the tool prints it alongside so the difference is visible.

One thing bars cannot do: overlapping SEM bars are not a significance test in either direction. Two means with visibly overlapping SEM bars can still differ significantly. That comparison is worked through on the Confidence Interval Calculator.

Why dividing by the square root of n is right

The formula looks arbitrary until you see where it comes from, and it comes from one property of variance.

Variances of independent quantities add. Add up n independent observations, each with variance σ², and the sum has variance nσ².

The mean is that sum divided by n. Dividing a quantity by n divides its variance by n², so the mean has variance nσ²/n² = σ²/n. Take the square root and you have σ/√n.

That is the whole derivation, and the square root appears because the arithmetic happens in variances while the answer is reported in standard deviations. It is the same square root behind the margin of error, behind “four times the data to halve the precision”, and behind why independent risks diversify.

The central limit theorem adds the second half of the story: for a large enough n the sampling distribution of the mean is approximately normal whatever the shape of the original data. That is what lets you attach a confidence level to a standard error at all. Skewed, bimodal, discrete — the mean still tends to normality, though strongly skewed data needs a larger n before the approximation is good.

Note the assumption doing the work: independence. Correlated observations — repeated measures on the same subjects, pupils within one class — do not have variances that simply add, and s/√n understates the true standard error, sometimes badly.

Every statistic has its own standard error

s/√n is the standard error of a mean. It is the one everyone memorises, and applying it to anything else is wrong.

A proportion has SE = √(p(1−p)/n). There is no separate standard deviation to confuse it with, because for a Bernoulli variable the spread is determined by the proportion itself.

A difference between two means has SE = √(SE₁² + SE₂²) when the samples are independent — the variances add, so the standard errors combine in quadrature rather than by addition.

A regression slope has a standard error involving the residual variance and the spread of the predictor: the more spread out your x values, the more precisely the slope is determined.

A median has an approximate standard error of about 1.253 × s/√n for normal data — about 25% larger than the mean’s. That is the efficiency cost of using the median, and the price paid for its robustness.

A correlation coefficient has a standard error that depends on r itself, which is why intervals for correlations are usually built through Fisher’s z transformation rather than directly.

When there is no closed form — a ratio, a median difference, an odds ratio in a small sample — the standard modern answer is the bootstrap: resample the data with replacement many times, recompute the statistic each time, and take the standard deviation of those values. It works for anything you can compute.

What the standard error is actually for

It is rarely the final answer. It is the input to almost everything else in inference.

Confidence intervals. estimate ± multiplier × SE. The SE supplies the scale; the multiplier supplies the confidence level.

Test statistics. Every t and z statistic is a difference divided by its standard error. That is literally all a test statistic is: how many standard errors away from the null value the estimate sits.

Margins of error. A margin of error is a multiplier times a standard error, so the square-root law is inherited directly.

Meta-analysis. Studies are combined by weighting each by the inverse of its squared standard error, so more precise studies count for more. This is why a meta-analysis needs the SE from each study and not just the effect.

A useful shortcut that falls straight out of this: if a reported estimate is more than about two standard errors from zero, it is significant at roughly the 5% level. That is the same statement as “zero is outside the 95% interval”, and it lets you read significance off a table of estimates and standard errors without a p-value column.

What a shrinking standard error does not fix

Because the SE falls without limit as n grows, it is tempting to treat a small one as proof of a good estimate. It is not.

The standard error only measures random sampling variation. If your sample is systematically unrepresentative, more data makes the standard error smaller and the estimate no more correct. You converge, with increasing confidence, on the wrong number.

This is the same warning that applies to the margin of error, and it matters more here because the SE shrinks so visibly. A survey of a million self-selected respondents has a tiny standard error and can be far less accurate than a properly drawn sample of a thousand. The 1936 Literary Digest poll had over two million responses and called the presidential election wrongly, because its sampling frame was biased.

It also assumes independence. With clustered or repeated-measures data, s/√n can understate the true standard error by a large factor, and the correct analysis has to account for the structure.

And a small SE says nothing about whether the effect matters. A precisely-estimated trivial difference is still trivial — which is the same point the p-value Calculator makes about significance, arriving from the other direction.

Reading a result that does not say which it used

Often you have to work it out. Three checks, in order of reliability.

Check the size against n. If the ± figure is roughly the same magnitude as the spread you would expect between individuals, it is an SD. If it is several times smaller, it is an SE. With n and the mean, you can often tell immediately.

Check whether it shrank between studies. If a larger study reports a much smaller ± on the same quantity, that is an SE. A standard deviation should be similar across studies of the same population.

Check the field convention. Clinical medicine and epidemiology usually report SD for describing a sample and confidence intervals for effects. Laboratory biology and psychology have historically used SEM for error bars. Physics and engineering tend to report a combined uncertainty with an explicit coverage factor.

If it genuinely cannot be determined, that is worth saying rather than assuming, because the two readings support different conclusions. Journals increasingly require the choice to be stated for exactly this reason.

Sources and methodology

Reporting guidance and the derivation behind the formula.

• Altman & Bland, “Standard deviations and standard errors”, BMJ 331, 2005 — the short paper that states the distinction and why it is misreported
• Cumming, Fidler & Vaux, “Error bars in experimental biology”, J Cell Biol 177, 2007 — on what each error-bar convention actually shows
• NIST/SEMATECH e-Handbook — the standard error of the mean and its assumptions
• Confidence Interval Calculator — where the standard error becomes an interval
• Standard Deviation Calculator — the other quantity on this page, in full
• How we calculate — methodology

Method. The mean’s standard error uses the sample standard deviation with the n − 1 divisor, matching Excel and R, and the proportion’s uses the observed p̂. Both are printed beside the standard deviation rather than alone, because the ratio — which is always exactly √n — is the thing worth seeing. The 95% error bar shown for comparison uses Student-t with n − 1 degrees of freedom, not a flat 1.96, so it is honest at small n. The suite asserts the two closed forms against hand values and checks the √n relationship holds across sample sizes. That engine is verified on every change against 155 assertions. The count and the per-case breakdown are published on the formula verification page.

Related calculators

Where this goes next:

More in Math, or browse all calculators.