HookThe brewer who wasn't allowed to say his name
In 1899 the Guinness brewery in Dublin hired a young Oxford chemistry graduate named William Sealy Gosset to bring science to beer. His problem was chronic small samples: a field trial of a new barley variety might yield four plots, a hops assessment eight resin measurements — and the statistical theory of the day assumed you knew the population standard deviation \(\sigma\), or had hundreds of observations from which to estimate it flawlessly. Gosset worked out exactly how much extra uncertainty a small-sample estimate of \(\sigma\) injects, checked his curve by shuffling 3,000 recorded height measurements into 750 samples of four and tallying the results by hand, and in 1908 published the answer. But Guinness, burned by an ex-employee leaking trade secrets, forbade its staff from publishing under their own names — so the paper appeared under the pseudonym Student. The distribution you are about to use is still called Student's \(t\).
The problem it solves is the honest version of every test on a mean. With \(\sigma\) known, \(\frac{\bar{X}-\mu}{\sigma/\sqrt{n}}\) is a standard normal and the critical value is 1.96. Replace \(\sigma\) with an estimate \(s\) computed from the same handful of observations and the statistic picks up a second source of randomness — the denominator now wobbles too. The result is a distribution with fatter tails than the normal: extreme values happen more often, so the evidence bar must sit higher. How much higher depends on one number, the degrees of freedom \(\nu=n-1\), and this section teaches you to run the whole test — hypotheses, statistic, critical value, conclusion — with that correction in place.
ModelWhy estimating \(\sigma\) changes the distribution
Suppose \(X\sim\mathrm{N}(\mu,\sigma^2)\) and you draw a random sample of size \(n\). If \(\sigma\) were known, \(Z=\frac{\bar{X}-\mu}{\sigma/\sqrt{n}}\sim\mathrm{N}(0,1)\) — that is A-level Maths. In real problems \(\sigma\) is a fiction; you must estimate it from the sample using the unbiased estimator with divisor \(n-1\):
\[s^2=\frac{1}{n-1}\sum(x_i-\bar{x})^2=\frac{1}{n-1}\left(\sum x_i^2-\frac{(\sum x_i)^2}{n}\right).\]
Substituting \(s\) for \(\sigma\) gives the test statistic \(t=\frac{\bar{x}-\mu}{s/\sqrt{n}}\), which follows Student's \(t\)-distribution with \(\nu=n-1\) degrees of freedom. Why \(n-1\)? The \(n\) deviations \(x_i-\bar{x}\) are not free: they must sum to zero, so only \(n-1\) of them carry independent information — one degree of freedom is spent estimating the mean. The \(t\)-curve is symmetric and bell-shaped like the normal but with heavier tails, and the smaller \(\nu\) is, the heavier they are: the two-tailed 5% critical value is 2.776 at \(\nu=4\), 2.262 at \(\nu=9\), 2.045 at \(\nu=29\), sliding towards the normal's 1.96 as \(\nu\to\infty\). The fat tails are not a technicality — they are the honest price of not knowing \(\sigma\), and using 1.96 with a sample of eight quietly pretends the price is zero.
MechanismSetting up — hypotheses, statistic, degrees of freedom
The machinery has a fixed order, and AQA marks each cog. One: state the hypotheses about the population mean in symbols — \(H_0:\mu=\mu_0\) against \(H_1:\mu\neq\mu_0\) (two-tailed) or \(H_1:\mu>\mu_0\) or \(H_1:\mu<\mu_0\) (one-tailed). The direction comes from the wording: 'has the mean changed?' is two-tailed; 'is the machine over-filling?' is one-tailed. Two: state the assumptions the test needs — the population is normally distributed and the sample is random. With a small sample there is no central limit theorem to hide behind, so normality is a genuine assumption, not a formality, and examiners ask for it by name. Three: compute \(\bar{x}\) and \(s\) (divisor \(n-1\) — on a calculator that is the \(s_x\) or \(\sigma_{n-1}\) key, never \(\sigma_n\)). Four: assemble the statistic and its distribution under \(H_0\):
\[t=\frac{\bar{x}-\mu_0}{s/\sqrt{n}}\sim t_{n-1}.\]
Write the degrees of freedom down explicitly; a correct \(t\)-value tested against the wrong row of the table scores like a wrong answer.
A bakery's loaves are sold as having mean mass 500 g. Trading standards weighs a random sample of 8 loaves and finds \(\bar{x}=494.2\) g with \(s=6.8\) g (divisor \(n-1\)), and the masses may be assumed normally distributed. Test at the 5% level whether the mean differs from 500 g. Hypotheses: \(H_0:\mu=500\), \(H_1:\mu\neq 500\) — 'differs' means two-tailed. Statistic: the standard error is \(\frac{s}{\sqrt{n}}=\frac{6.8}{\sqrt{8}}=2.404\) g, so \(t=\frac{494.2-500}{2.404}=\frac{-5.8}{2.404}=-2.41\) (3 s.f.), with \(\nu=8-1=7\) degrees of freedom. The setup is complete — the verdict belongs to the next block.
MechanismThe verdict — critical values, comparison, conclusion in context
Under \(H_0\), the statistic is a random draw from \(t_{n-1}\); the test asks whether the observed value is too extreme to believe. From the tables, find the critical value for your tail arrangement: for a two-tailed test at significance level \(\alpha\), each tail gets \(\frac{\alpha}{2}\); for one-tailed, the whole \(\alpha\) sits in one tail. The comparison is then mechanical — reject \(H_0\) if the statistic lands beyond the critical value (for two tails, if \(|t|>t_{\text{crit}}\)) — but the conclusion is not. AQA wants two sentences, in this order: the statistical decision ('reject \(H_0\)' or 'insufficient evidence to reject \(H_0\)'), then the meaning in context, hedged correctly: 'there is significant evidence at the 5% level that the mean mass differs from 500 g'. Never 'this proves', never 'the mean is 494.2' — a hypothesis test weighs evidence, it does not certify facts. And note the asymmetry: failing to reject \(H_0\) is not evidence that \(H_0\) is true, merely a verdict of 'not proven' on this sample.
Completing the bakery test: two-tailed at 5% with \(\nu=7\), the critical values are \(\pm 2.365\). Since \(|t|=2.41>2.365\), the statistic falls in the rejection region: reject \(H_0\). In context: there is significant evidence at the 5% level that the mean mass of the loaves differs from 500 g — and since \(\bar{x}<500\), the direction of concern is under-filling. Two instructive reruns. At the 1% level the critical values become \(\pm 3.499\) and \(|t|=2.41\) is comfortably inside: the same data no longer justify rejection — the conclusion is bought at a significance level, not absolute. And had we wrongly used the normal critical value 1.96 (pretending \(s\) were \(\sigma\)), rejection would look far more emphatic than seven degrees of freedom actually warrant: with \(n=8\), the gap between 1.96 and 2.365 is exactly the price of estimating \(\sigma\) from eight numbers.
CaseGuinness, small data, and when \(t\) is the right tool
Gosset's employer cared about a ruthlessly practical question: does this barley variety, this kiln temperature, this hop schedule actually change the beer, on the evidence of a handful of trials? That is still the exact shape of every \(t\)-test question: a stated benchmark mean \(\mu_0\) (the label value, the historical average, the manufacturer's claim), a small random sample, no outside knowledge of \(\sigma\), and a population plausibly normal. Run the checklist before committing: is \(\sigma\) unknown? If \(\sigma\) is given, the test is the ordinary \(z\)-test and reaching for \(t\) is wrong. Is the sample small? For large \(n\) the \(t\)-distribution is practically normal — \(t\) is still valid, but the distinction stops mattering. Is normality reasonable? Masses, lengths, times from a stable process usually pass; heavily skewed data (waiting times, incomes) fail, and with \(n=8\) there is no large-sample rescue.
The deeper lesson Gosset left is about intellectual honesty with small data. His curve concedes more to chance precisely when you know least — the critical value at \(\nu=4\) is 2.776, forty percent sterner than the normal's 1.96 — and relaxes only as evidence accumulates. Ronald Fisher, who tabulated the distribution in the form you use and christened the statistic \(t\), called Gosset's work one of the most important publications in the history of statistics. Guinness never did let him take the credit while he worked there; the tables in your formulae booklet are his monument.
VocabularyKey terms the mark scheme pays for
TrapsMisconceptions that cost marks
ExamWhat examiners want
The mark scheme for a one-sample \(t\)-test is a fixed skeleton — hypotheses in symbols with \(\mu\) defined in context; assumptions (normal population, random sample) stated; \(s\) computed with divisor \(n-1\); the statistic \(t=\frac{\bar{x}-\mu_0}{s/\sqrt{n}}\) evaluated with working shown; degrees of freedom \(\nu=n-1\) written explicitly; the critical value quoted from the tables for the correct tail arrangement; comparison; conclusion. Rehearse it as a skeleton and every question becomes filling in eight blanks. The two habitual mark-losers are calculator variance keys (\(\sigma_n\) instead of \(s_{n-1}\) — check by recomputing \(s^2\) once from the formula) and tail confusion — a two-tailed 5% test uses the 2.5% column, and mixing that up shifts the critical value a whole column.
Conclusions carry their own marks and their own language. Two sentences: the decision about \(H_0\), then the contextual meaning with the significance level attached — 'there is significant evidence at the 5% level that the mean mass of loaves is below 500 g'. Keep the verbs non-assertive: 'evidence that', 'suggests', never 'proves' or 'shows the mean is'. If the question asks why a \(t\)-test rather than a \(z\)-test, the creditworthy answer names both triggers: \(\sigma\) is unknown and the sample is small, so the extra variability of \(s\) matters and the population must be assumed normal. And if asked what changes with a larger sample, say the degrees of freedom rise, the critical value falls towards 1.96, and the test gains power — three consequences, one sentence each.