Does sample size change the standard deviation?

What the standard deviation is estimating

A sample standard deviation is an estimate of a fixed property of the population: how spread out the individual values are. Adult heights, the weights of cereal boxes off a production line, the response times of a web server each have some true spread, σ, that exists whether you measure 5 cases or 5,000. Measuring more cases does not make the boxes more alike. It only tells you more precisely how alike they already were.

So as n grows, the sample SD should settle toward σ, from above or below, rather than drift toward zero.

A simulation

To show this, we drew random samples from a normal population with mean 50 and σ = 10, using a fixed random seed so the numbers are reproducible (NumPy's default_rng(4), drawing samples of 5, 20, 100 and 1,000 in that order). The values are illustrative: another seed gives different numbers, with the same pattern.

The SD column wanders around 10 and ends up close to it. It happens to fall from 12.60 to 10.02 over the first three rows, which is exactly the pattern that makes people believe the myth, and then it goes back up to 10.26 at n = 1,000. The standard error column behaves completely differently: it falls at every step, because √n in its denominator keeps growing while s stays near 10.

One sample per size is only an anecdote. To see the typical behaviour, we repeated the experiment 10,000 times at each sample size (seed 2026) and summarised the results:

Three things are visible. The average sample SD stays close to 10 at every size. The range of sample SDs narrows sharply: with 5 observations, one sample in ten gives an SD outside 4.18 to 15.36, while with 1,000 the middle 90% is within 4% of the truth. And the spread of the sample means matches σ/√n almost exactly, which is the central limit theorem at work.

Try it: standard deviation of the sample mean calculator

The calculator below opens with its defaults; enter σ = 10 and n = 5, 100 or 1,000 to reproduce the standard errors above, or use your own SD and sample size.

Open the full standard deviation of the sample mean calculator to work from raw data or apply the finite population correction.

Why the average s is slightly below 10 for small n

The small-sample row shows an average s of 9.39, not 10. That is not the myth sneaking back in. The n − 1 divisor makes the sample variance unbiased, but taking a square root pulls the average down a little. For normal data the average of s is σ times a constant called c₄, which is 0.940 for n = 5, 0.987 for n = 20 and 0.997 for n = 100. The simulated averages, 9.39, 9.86 and 9.98, match those almost exactly. The bias is toward smaller values at small n, the opposite of what the myth predicts, and it disappears as the sample grows.

The standard error is what shrinks

The quantity that does depend on n is the standard error of the mean:

SE = s / √n

It measures how much the sample mean would vary from sample to sample. Quadruple the sample and the SE halves; multiply it by 100 and the SE falls to a tenth. With s = 10, that is 4.47 at n = 5, 1.00 at n = 100 and 0.32 at n = 1,000. The standard error calculator computes it from raw data, and the sample size calculator runs the relationship backwards to find how many observations a target precision needs.

The confusion usually comes from a chart or a paper reporting "mean ± error" without saying which error. Error bars that shrink as a study grows are standard errors or confidence intervals. If they were standard deviations, they would stay about the same width.

The n − 1 correction fades as n grows

There is one real way the sample size enters the formula: the divisor. Dividing by n − 1 instead of n makes the sample SD larger than the population-formula SD by a factor of √(n / (n − 1)). The reasons for the correction are on the sample vs population page; here is how much it matters:

For a sample of 5 the choice of formula changes the answer by almost 12%; for 1,000 it is lost in the rounding. In the simulation, the n = 1,000 sample gave 10.264 with n − 1 and 10.259 with n. So a large sample does make the two formulas agree, but it does that by making the sample SD stop depending on the divisor, not by making it smaller.

What to take from this

If you are planning a study, do not expect more data to reduce the variability you see in individual measurements; expect it to reduce your uncertainty about the mean. If a small pilot study gave an SD of 12.6 and you size the main study on that, the full study may well find something closer to 10, or to 14. Small-sample SDs are rough, and the second table shows how rough. When comparing SDs across studies of different sizes, compare them directly: they estimate the same kind of thing. When comparing standard errors, remember that n is baked into them.