A Confidence Interval for a Mean

Every "average \pm margin" you have ever seen reported — average household income, average blood pressure in a trial, the average weight of a factory's cereal boxes — is built by one calculation. This is that calculation, assembled from parts you already have.

The move is simple: take a sample, compute its mean, work out its standard error, and wrap the mean in a margin to get a confidence interval for the true population mean \mu. It is the workhorse of applied statistics — learn it once and you will see it everywhere.

The recipe

Centre the interval on the sample mean, then reach out by a multiple of the standard error:

\bar{x} \pm z^{\*}\cdot \mathrm{SE}, \qquad \mathrm{SE} = \frac{\sigma}{\sqrt{n}}.

The half-width z^{\*}\cdot\mathrm{SE} is the margin of error. The multiplier comes from the standard normal — courtesy of the central limit theorem, which makes \bar{x} approximately normal — read off as a z-score. For 95\% confidence, z^{\*} \approx 1.96; for 99\%, z^{\*} \approx 2.58.

One fork in the road. If the population spread \sigma is known (or n is large), use the normal's z^{\*}. If \sigma is unknown and you must estimate it with the sample SD s from a small sample, swap z^{\*} for the slightly larger critical value of Student's t-distribution — the same machinery that powers the t-test.

Worked example: cereal boxes, step by step

A factory samples n = 100 cereal boxes. The sample mean weight is \bar{x} = 500\text{ g} and the sample SD is s = 20\text{ g}. Build a 95\% confidence interval for the true mean weight.

  1. Standard error: \mathrm{SE} = \dfrac{s}{\sqrt{n}} = \dfrac{20}{\sqrt{100}} = \dfrac{20}{10} = 2\text{ g}.
  2. Critical value: 95\% confidence, large sample, so z^{\*} \approx 1.96.
  3. Margin of error: 1.96 \times 2 = 3.92\text{ g}.
  4. The interval: 500 \pm 3.92 = [496.08,\ 503.92]\text{ g}.

In plain English: we are 95% confident that the true mean weight of all boxes lies between about 496 g and 504 g — meaning the procedure we used lands an interval around the real mean 95% of the time. If the label promises 500 g, the factory is comfortably on target.

Worked example: same data, tighter or looser

Keep \bar{x} = 500 and s = 20, and watch two dials move the interval.

Note the \sqrt{n}: quadrupling the sample only halved the margin. Precision is real but expensive.

Reading one in the wild

Once you know the shape, you can decode any reported interval on sight. A clinical trial announces: "the drug lowered blood pressure by 8.0 mmHg on average (95% CI: 5.6 to 10.4)." Unpack it:

That last reading — does the interval contain the boring value? — is the bridge from confidence intervals to hypothesis testing, which asks the same question from the other direction.

Build the interval

The dot is the sample mean \bar{x}; the bracket spans \bar{x} \pm z^{\*}\cdot\mathrm{SE} with \sigma fixed. Slide n up and the bracket tightens; switch to higher confidence and it stretches.

The formula \bar{x} \pm 1.96\cdot s/\sqrt{n} is only right when you know \sigma or your sample is large. With a small sample, you don't know the true spread — you estimated it with s, and that estimate is itself uncertain. Pretending s is exact makes your interval too narrow, so it overstates your precision and quietly inflates your confidence.

The fix is the t-distribution. Its critical value is a bit larger than z — for n = 5 at 95\% it's about 2.78 instead of 1.96 — which widens the interval to honestly account for the extra uncertainty in s. The correction matters most when n is tiny and vanishes as n grows: by n \approx 30 the t- and z-values are nearly identical. Small sample, unknown \sigma? Reach for t.

The t-distribution owes its existence to beer. Around 1900, William Sealy Gosset was a young chemist at the Guinness brewery in Dublin, wrestling with tiny samples of barley and hops — just a handful of measurements, far too few for the big-sample theory of his day. So he worked out the correct maths for small samples himself, and in doing so invented what we now call the t-distribution.

There was one snag: Guinness forbade its employees from publishing, terrified that rivals would learn its trade secrets. So Gosset published his 1908 paper under a modest pseudonym — "Student" — and the name stuck forever. To this day it is Student's t-distribution, one of the great pseudonyms in the history of science, and a reminder that some of the sturdiest tools in statistics were forged to check the consistency of a pint.

See it explained