The standard error is the standard deviation of the sample mean across repeated studies — the typical distance between your x̄ and the μ you are chasing. It equals σ/√n, so raising the sample size narrows the sampling distribution. But it narrows by √n, not by n, and that one square root sets the price of every precise study ever run. Draw samples, move n, and watch the pile tighten.
Watch it narrow
simulated sample meansnormal model at your nmodel at 1, 4, 16, 64±1 SE±2 SEμ
The exchange rate: ×4 the data, ÷2 the error
Bell-shaped populationn = 1 · press “Draw 200 samples” to begin
Population you are sampling from
Sample size — the quadrupling ladder
1
Population σ15.0
Sample size n1
Divide by √n1.00
Standard error15.00
Where the sample means actually landed
Top row: measured from the x̄ values you piled up. Bottom row: what σ/√n predicted before any data existed.
SD of the x̄'s—σ/√n = 15.00
Within ±1 SE—68.3% if normal
Within ±2 SE—95.4% if normal
Biggest miss——
How many would you need?
Turn the question around. Name the precision you want and the √n law hands you the sample size: solve σ/√n = target for n, giving n = (σ/target)².
0.25σ
Four times the data, half the error. That is the whole section. The standard error is not a mistake anyone made — it is the built-in wobble of a mean computed from n individuals instead of everyone, and it shrinks only as fast as √n. Doubling your sample does not halve your uncertainty; it divides it by √2, about 1.41, a 29% improvement. To halve it you need four times the data. To cut it to a tenth you need a hundred times the data. Sample size is bought on a square-root sale, and the sale gets worse the more you buy.
Three things √n cannot do.(1) It does not shrink your data. σ is a property of the population and it never moves; a bigger sample gives a sharper mean, not tidier observations. (2) It does not fix bias. If your sampling method systematically misses people — volunteers only, one neighborhood, whoever answers the phone — then x̄ is centered on the wrong number, and raising n shrinks the standard error around that wrong number. A large biased study is a confidently wrong study. (3) It does not depend on the size of the population. There is no N in σ/√n: a sample of 1,000 is equally precise for a town or a nation. (Sampling without replacement from a population only a few times larger than the sample is the one exception, and it makes the error smaller, not larger.) One more honest caveat: σ/√n needs σ, which you almost never know. In real work you report the estimated standard error s/√n, using the sample's own SD — the same √n law, with one more thing being guessed.