Standard Error Calculator – How Precise Is Your Sample Mean?

Related tools

How it works, in three steps

Going deeper

The formula, and why the square root is there

Standard error is the standard deviation divided by the square root of the sample size, and the square root is the whole story of the page.

At the defaults, a sample standard deviation of 15 and a sample size of 25 give a standard error of 3. The working the page shows is direct: the square root of 25 is 5, and 15 divided by 5 is 3. In the raw-data mode, the same formula runs on a standard deviation the page computes for you: the data 2, 4, 6, 8, 10 give a sample standard deviation of 3.1623, a sample size of 5, a square root of 5 of 2.2361, and a standard error of 1.4142, with a mean of 6. The page also shows the confidence-interval mode, in which a mean and a standard error you already know are combined with a z value, but the standard error itself always comes from the same division, whether you type the standard deviation or let the page compute it from raw data.

Every figure in that table was produced by this page’s own calculator and is quoted exactly as displayed. The square-root law is visible in the first two rows: quadrupling the sample from 25 to 100, at the same standard deviation of 15, only halves the standard error, from 3 to 1.5, because the square root of 100 is 10, only twice the square root of 25, which is 5. That is why bigger samples help, but with diminishing returns, and why the page’s own note says that quadrupling the sample only halves the standard error. To halve the error again, from 1.5 to 0.75, the sample would have to quadruple again, from 100 to 400, and the cost of each halving grows fourfold while the benefit stays the same. Sample size planning is therefore a trade-off between precision and cost, and the square root is what sets the exchange rate.

The intuition behind the square root is averaging. A sample mean averages n individual values, and averaging cancels out part of their individual variation: some values fall above the true mean and some below, and their errors partly offset. The cancellation is not complete, because the errors do not cancel perfectly, and the square root is the mathematical residue of that partial cancellation. If the errors cancelled completely, the standard error would fall as one over n, and quadrupling the sample would quarter the error. If they did not cancel at all, the standard error would not fall with n. Reality sits between the two, at one over the square root of n, and the page’s formula is that reality, stated as arithmetic.

Standard error is not standard deviation

The two names sound alike, the two numbers behave differently, and confusing them is the most common standard-error mistake.

Standard deviation describes how spread out individual values are. In the raw-data example, the values 2, 4, 6, 8, 10 have a sample standard deviation of 3.1623, which says something about how far a typical single value falls from the mean of 6. Standard error describes how precise the sample mean is as an estimate of the true population mean. For the same data, the standard error is 1.4142, which is smaller, because the mean of five values is more stable than any single value. As the sample grows, the standard deviation stays roughly the same, because the spread of individual values does not shrink when you collect more of them, while the standard error falls, because the mean of more values is more precise. That divergence is the point of the whole page: one number describes the data, the other describes the estimate, and they move in different directions as n grows.

The page’s own FAQ draws the same distinction in plain words, and its mode structure enforces it. In SD-and-n mode you supply a standard deviation and a sample size, and the page divides. In raw-data mode you supply the values, and the page first computes the sample standard deviation, then divides by the square root of the count. The base guide notes that the raw-data mode uses the sample version of the standard deviation, because that is the version the standard error formula needs, and that if you already know the full population standard deviation, you should use the SD-and-n mode and type it in directly. The sample-versus-population distinction matters because the sample version divides by n minus 1 rather than n, which makes it slightly larger for small samples, and the standard error inherits that difference. For the five-value example the sample standard deviation of 3.1623 is the figure the page uses, and the standard error of 1.4142 follows from it.

The confidence-interval mode builds on the same distinction. It does not compute a standard error; it takes a mean and a standard error you already have, and adds and subtracts a margin, which is the z value times the standard error. At 95 percent confidence the z value is 1.96, so a mean of 100 with a standard error of 3 gives a margin of 5.88 and an interval from 94.12 to 105.88. At 99 percent confidence the z value is 2.576, so the margin grows to 7.728 and the interval widens to 92.272 to 107.728. At 90 percent confidence the z value is 1.645, the margin shrinks to 4.935, and the interval narrows to 95.065 to 104.935. All three intervals are centred on the same mean of 100 with the same standard error of 3, and all three were produced by this page’s own calculator and are quoted exactly as displayed. The pattern is the price of confidence: a wider interval buys a higher stated confidence level, and the z value is the exchange rate, just as the square root is the exchange rate for sample size.

Reading a confidence interval correctly

A 95 percent interval is a statement about a procedure, not a probability about one interval, and the distinction changes how it should be quoted.

The interval from 94.12 to 105.88, built from a mean of 100 and a standard error of 3 at z equals 1.96, is often misread as “there is a 95 percent chance the true mean lies between 94.12 and 105.88”. That reading is not what the method means. The correct reading is procedural: if the same sampling procedure were repeated many times, and an interval were built the same way each time, about 95 percent of those intervals would contain the true mean. This particular interval either contains it or it does not; the 95 percent describes the long-run behaviour of the procedure, not a probability attached to these two endpoints. The distinction sounds pedantic, but it matters when intervals are used in decisions: a procedure that captures the truth 95 percent of the time is a reliable procedure, while a single interval with a supposed 95 percent chance attached to it invites overconfidence about one specific result.

The width of the interval carries the same lesson in a different form. A wider interval, such as the 99 percent interval from 92.272 to 107.728, is more likely to capture the truth across repetitions, but it is also less informative about where the truth lies, because it covers more ground. A narrower interval, such as the 90 percent interval from 95.065 to 104.935, is more informative but captures the truth less often across repetitions. There is no interval that is both maximally confident and maximally precise, and the choice of confidence level is a choice about which error to favour. The page offers 90, 95 and 99 percent, with z values 1.645, 1.96 and 2.576, and the three intervals above show the trade-off on identical inputs, so the choice can be made with its price visible.

One more reading habit protects against the most common interval error. An interval is centred on the sample mean, not on the true mean, and its width is set by the standard error, not by the standard deviation. Centring on the true mean is impossible, because the true mean is what the interval is trying to locate, and widening by the standard deviation instead of the standard error would produce an interval about the spread of individual values, not about the precision of the mean. The page’s layout prevents both errors by labelling the mean and standard error inputs separately from the standard deviation inputs of the other modes, but the habit of checking which quantity is in which field is still worth building, because the numbers look alike and the fields sit close together.

Five mistakes with standard error

Each mistake below produces a number that looks like a standard error but is not one, and each is caught by checking the formula, not the result.

1. Dividing by n instead of the square root of n. With a standard deviation of 15 and a sample of 25, dividing by 25 gives 0.6, while dividing by the square root of 25, which is 5, gives the correct 3. The wrong divisor understates the error by a factor of five in this case, and the understatement grows with n, which is exactly backwards: the mistake makes larger samples look more precise than they are, by an ever larger factor.

2. Using the standard deviation as the standard error. Reporting 15, the standard deviation, as if it were the precision of the mean, overstates the uncertainty of the mean by a factor of five at n equals 25, and by a factor of ten at n equals 100. The standard deviation describes individuals; the standard error describes the mean. The page shows both, labelled, in the raw-data mode, so the mistake is a labelling error, and labelling errors are the ones that survive into reports.

3. Using a sample size of 1, or forgetting why n must exceed 1. The page’s own FAQ explains that with only one observation there is no sample variation to measure, and the sample standard deviation itself needs at least two values, which is why the calculator asks for a sample size greater than 1. A standard error from n equals 1 is not merely imprecise; it is undefined in the sample framework, because there is no variation to divide.

4. Mixing the sample and population standard deviations. The raw-data mode uses the sample version, dividing by n minus 1, because that is what the standard error formula needs. Typing a population standard deviation, computed by dividing by n, into the SD-and-n mode is legitimate, as the base guide advises, but only when you know it is the population figure. Mixing the two without noting which was used makes the standard error slightly wrong, and the error is largest exactly where samples are smallest and precision matters most.

5. Reading a confidence interval as a prediction interval. A confidence interval for the mean is narrower than an interval that predicts where a single new observation will fall, because the mean is more stable than any individual value. Using the interval from 94.12 to 105.88 as if it described individual observations would understate their spread badly. The page’s interval mode is explicitly for the mean, with a mean and a standard error as inputs, and individual prediction needs a different calculation that this page does not offer.

Planning a sample with the square root

The square-root law turns sample-size planning into a simple question: how much precision is the next quadrupling worth?

Suppose a standard deviation of 15 is expected, and the current plan uses 25 observations, giving a standard error of 3. Quadrupling to 100 observations halves the error to 1.5, as the page displays. Quadrupling again to 400 observations would halve it again, to 0.75, by the same law. Each quadrupling costs four times the data collection of the previous stage and buys the same halving, so the decision is never “more data is better” in the abstract, but “is this halving worth four times the cost?” For many purposes the answer changes between the first and second quadrupling: the move from 3 to 1.5 may be worth it, while the move from 1.5 to 0.75 may not be, and the page makes that judgement possible by showing the error for any n you type, rather than by recommending one.

The same logic works backwards from a target precision. If a standard error of 1 is required with a standard deviation of 10, the page shows that n equals 100 achieves it exactly, because 10 divided by the square root of 100, which is 10, is 1. If a standard error of 1 is required with a standard deviation of 15, the required n is 225, because the square root of 225 is 15, and 15 divided by 15 is 1. That 225 figure is arithmetic from the page’s own formula, using its own standard deviation values, and it illustrates the planning method: square the ratio of the standard deviation to the target error, and that is the sample size the formula implies. Entering 225 as the sample size with a standard deviation of 15 will display a standard error of 1 on the page itself, which is the check that the planning arithmetic and the calculator agree.

Finally, remember that precision is not the only thing a larger sample buys, and the standard error is not the only thing that limits a study. A larger sample does not fix a biased sampling method, a flawed measurement, or a wrong model, and a small standard error around a biased mean is precisely wrong rather than roughly right. The page computes the precision of the mean given the data, and it does so exactly for its formula. Whether the mean itself deserves that precision is a question about how the data were collected, and no calculator can answer it from the numbers alone.

From a target error to a sample size

The formula runs backwards as easily as forwards, and running it backwards is how sample sizes are planned.

Start from the precision you need, and solve for n. The formula is standard error equals standard deviation divided by the square root of n, so the square root of n equals the standard deviation divided by the target standard error, and n equals that ratio squared. With a standard deviation of 10 and a target standard error of 1, the ratio is 10 divided by 1, which is 10, and squaring gives 100, which is exactly the case the page displays: a standard deviation of 10 with a sample of 100 gives a standard error of 1. With a standard deviation of 15 and a target standard error of 1, the ratio is 15, and squaring gives 225, and the page displays a standard error of 1 for a standard deviation of 15 with a sample of 225, confirming the planning arithmetic against the calculator itself. With a standard deviation of 15 and a target of 0.75, the ratio is 20, and squaring gives 400, and the page displays 0.75 for a standard deviation of 15 with a sample of 400.

The planning formula also shows why target precision has a cost curve, not a price list. Halving the target error from 2 to 1 quadruples the required sample, because the ratio doubles and the sample is the square of the ratio. Halving again, from 1 to 0.5, quadruples it again. A study that needs a standard error of 0.5 with a standard deviation of 15 would need a sample of 900, because 15 divided by 0.5 is 30, and 30 squared is 900, and the page displays 0.5 for a standard deviation of 15 with a sample of 900, confirming the planning arithmetic against the calculator itself. The practical question is always whether the extra precision is worth the extra observations, and the formula makes that question quantitative instead of rhetorical.

One caution applies to backwards planning. The standard deviation used in the planning formula is an estimate, usually from a pilot study or from prior work, and if that estimate is wrong, the planned sample will miss its target error. Underestimating the standard deviation is the dangerous direction: planning for a standard deviation of 10 when the true figure is 15 means the achieved standard error will be fifty percent larger than planned, because the error scales directly with the standard deviation. The remedy is to plan with a conservative, slightly high standard deviation estimate, or to plan in stages, collecting a first batch, re-estimating the standard deviation from it, and then deciding how many more observations the target error actually requires. The page supports staged planning directly, because any standard deviation and sample size can be typed in and the resulting standard error read at once, without any commitment to a single plan made in advance.

What the page cannot tell you

A standard error is a precise number about an imprecise world, and its precision ends where the data collection begins.

The page computes the standard error from the numbers you enter, assuming those numbers are a random sample from the population the mean is meant to describe. If the sample was not random, if it systematically excludes part of the population, or if the measurements themselves are biased, the standard error will be precise about the wrong thing: it will describe how much the sample mean would vary across repeated samples drawn the same biased way, not how far the sample mean sits from the true population mean. No increase in sample size fixes that, because the square-root law only shrinks the variation across repetitions, not the bias shared by all of them. A large sample with a small standard error and a large bias is the most dangerous combination in statistics, because the small error invites confidence that the bias does not deserve.

Measurement quality is a second limit. The standard deviation fed into the formula reflects the variation in the measurements, including measurement noise, and the standard error inherits it. Noisier measurements give a larger standard deviation and a larger standard error, which is correct, but they also mean the mean itself is estimated from noisier data, and the interval built from it, however correctly computed, inherits the noise. Improving measurement, rather than increasing the sample, is sometimes the cheaper route to a smaller standard error, because the error scales directly with the standard deviation: halving the measurement noise halves the error without adding a single observation, while halving the error by sample size alone requires quadrupling the sample.

Finally, the page’s intervals assume the normal approximation behind the z values, which is reasonable for large samples and for roughly symmetric data, but less so for small samples from skewed distributions, where the sampling distribution of the mean is itself skewed and a symmetric interval centred on the sample mean misstates the uncertainty on at least one side. The page does not warn about this, because it is a single-purpose calculator, not a statistics course, and its z values, 1.645, 1.96 and 2.576, are the standard normal values for 90, 95 and 99 percent. For small samples, a t-based interval would be wider and more honest, and the difference matters most exactly where samples are smallest. Knowing that the page’s intervals are z-based, and that z-based intervals are approximations whose quality depends on sample size and shape, is part of using the page well, even though the page itself does not offer the t alternative.

Frequently asked questions

More calculators

This calculator is an educational tool for working out standard error and confidence intervals. It does not give statistical advice, and its results should be checked before being used in any formal report or decision.