Standard Error

When an election poll reports a result "accurate to within 3 points," that little margin is built from the standard error — the number that says how far a sample's average is likely to sit from the truth. It quietly sits behind every poll, drug trial and factory quality check you will ever read.

You take one sample, compute its mean \bar{x}, and it lands near the true population mean \mu — but not exactly on it. So here is the question that everything hinges on: how much does a sample mean typically miss the truth by? Not the worst case, not the best case — the typical gap.

The standard error answers exactly this. It is the standard deviation of the sampling distribution of the mean — the typical distance between your sample's mean and \mu. And it shrinks as your sample grows, in a very specific and very useful way.

The sampling distribution of the mean has a spread, and that spread has a name: the standard error. It is just the standard deviation of the sample mean \bar{x} — a measure of how much \bar{x} wobbles from sample to sample:

\operatorname{SE} = \frac{\sigma}{\sqrt{n}}.

Don't confuse it with \sigma. The population standard deviation \sigma measures how spread out the individual values are; the standard error measures how spread out the averages are. Averaging always tames the wobble, so \operatorname{SE} < \sigma whenever n > 1.

Precision grows with √n, not n

The \sqrt{n} in the denominator carries a sober warning about diminishing returns. Because it is the square root — not n itself — that drives the shrinkage, precision improves slowly:

\operatorname{SE} = \frac{\sigma}{\sqrt{n}} \;\Longrightarrow\; \text{to halve the SE you must } \textbf{quadruple } n.

Going from n=100 to n=400 — four times the data, four times the cost — only cuts the standard error in half. Each extra digit of precision gets dramatically more expensive.

Three worked examples

Example 1 — compute an SE. A population of test scores has \sigma = 15. You take a sample of n = 25 students. Then

\operatorname{SE} = \frac{\sigma}{\sqrt{n}} = \frac{15}{\sqrt{25}} = \frac{15}{5} = 3.

A single student's score wobbles by 15 points, but the class average of 25 students wobbles by only 3 points around the true mean. The average is five times steadier than an individual.

Example 2 — quadruple the sample, halve the error. Keep \sigma = 15 but grow the sample fourfold to n = 100:

\operatorname{SE} = \frac{15}{\sqrt{100}} = \frac{15}{10} = 1.5.

Four times the students, and the error dropped only from 3 to 1.5 — exactly half. That is the famous \sqrt{n} law of diminishing returns in one line: to double your precision you must quadruple your effort.

Example 3 — how close is the truth likely to be? For a roughly normal sampling distribution, about 95% of sample means land within 2\,\operatorname{SE} of \mu. In Example 2, with \operatorname{SE} = 1.5, that means your sample mean is very likely within 2 \times 1.5 = 3 points of the true mean. The standard error is the raw material from which every "margin of error" and confidence interval is built.

Feel the diminishing returns

The faint curve is the population, with spread \sigma. The bold curve is the distribution of \bar{x}, with spread \operatorname{SE} = \sigma/\sqrt{n}. Raise n and the bold curve narrows — but notice you have to go from n=4 to n=16, a fourfold jump, just to halve its width.

The standard error is not the same thing as the standard deviation, and swapping one for the other is a genuinely common mistake — it appears in published research papers.

With n = 100 the SE is ten times smaller than the SD. Report the SD when you meant the SE and you make your estimate look ten times shakier than it is; report the SE when you meant the SD and you make the population look ten times more uniform than it is. One symbol, two utterly different claims about uncertainty — always ask which spread you mean: the spread of the people, or the spread of the average.

A poll for a country of 300 million and a poll for a town of 30,000 use samples of almost the same size — usually around 1,000 to 2,000 people. That surprises everyone. Shouldn't a bigger country need a bigger sample? No: the standard error depends on n, not on the size of the population you drew from.

And the \sqrt{n} law is why they stop there. Going from 1,000 to 4,000 respondents costs four times as much but only halves the margin of error — from about ±3% to about ±1.5%. Push to ±0.75% and you need 16,000 people. Pollsters stop where the cost/precision trade-off flattens out. The very same \sqrt{n} shows up all over science and beyond: the tiny fixed edge that guarantees a casino profits over millions of bets, and the way physicists average N noisy repeats of a measurement to sharpen it by \sqrt{N}. Averaging beats down noise everywhere — but always at the same stubborn square-root rate.

See it explained