The difference in one sentence
Standard deviation describes how spread out individual observations are. Standard error describes how much a sample mean would vary if you repeated the study.
They measure different things about different populations of numbers. One is about people; the other is about averages of people.
Why standard error is always smaller
The formula makes it obvious:
SE = σ ÷ √n
Since √n is greater than 1 for any sample bigger than one person, the standard error is always smaller than the standard deviation — and shrinks as your sample grows.
The intuition matters more than the algebra. Individual people vary a lot. Averages of many people vary much less, because unusually high and unusually low values cancel out. Take a sample of 100 and one outlier barely moves the mean.
A concrete example
Suppose adult heights have μ = 170 cm and σ = 10 cm.
- Standard deviation = 10 cm. Roughly 68% of individuals are between 160 and 180 cm.
- Standard error for n = 100 = 10 ÷ 10 = 1 cm. Roughly 68% of sample means from samples of 100 fall between 169 and 171 cm.
Same population, tenfold difference. Ask “how tall is a random person?” and you need the standard deviation. Ask “how close is my sample mean to the truth?” and you need the standard error.
The mistake this causes in practice
Consider a population with μ = 70 and σ = 12, and a sample of 36. What is the probability of a value above 73?
It depends entirely on what “a value” means:
- For one individual: z = (73 − 70) ÷ 12 = 0.25, giving a probability of about 0.40.
- For a sample mean of 36: SE = 12 ÷ 6 = 2, so z = 3 ÷ 2 = 1.5, giving about 0.067.
Same numbers, and the answers differ six-fold. Using σ where SE belongs is the commonest error in the whole of inference, and it is almost always fatal to the answer.
How to tell which one a question wants
Read the question and ask: is this about one thing, or about an average of many things?
| The question says… | Use |
|---|---|
| “a randomly chosen person…” | Standard deviation σ |
| “the mean of a sample of n…” | Standard error σ/√n |
| “a confidence interval for μ…” | Standard error |
| “how variable is the data?” | Standard deviation |
The square-root law, and why it is expensive
Because the divisor is √n rather than n, precision improves slowly:
- To halve the standard error, you must quadruple the sample.
- To cut it to a third, you need nine times as much data.
This is why surveys of 1,000 people are so common. Getting the margin of error from ±3 points to ±1.5 would mean surveying 4,000 — four times the cost for a modest gain. Diminishing returns set in fast.
Standard error is not only for means
Every statistic has a sampling distribution, so every statistic has a standard error. For a proportion:
SE = √[ p(1 − p) ÷ n ]
Regression slopes have one too, which is exactly what the “Std. Error” column in any regression output reports. Dividing an estimate by its standard error gives the t-statistic — the same move, over and over, throughout inference.
Which should a report show?
Error bars on a chart can be either, and the caption should say which. Standard error bars look reassuringly small; standard deviation bars are wider and honestly show individual variation. Showing SE bars and letting readers assume they represent spread is one of the quieter ways graphs mislead.
The safest option is usually a confidence interval, which is built from the standard error but comes with an explicit interpretation attached.
The takeaway
Standard deviation answers “how different are individuals?” Standard error answers “how precisely do I know the average?” Both are useful. Using one where the other belongs will get you a wrong answer that looks entirely reasonable — which is the worst kind.