The t-distribution is the reference curve a t-test compares its statistic against. It resembles the normal curve but has heavier tails, which is how it accounts for the extra uncertainty created by estimating the spread of the data from the same small sample.
If you knew the true standard deviation of your data, comparing two means would be simple: standardise the difference and read it off the normal curve. You never know it. You estimate it from the same sample that produced the means, and that estimate is itself uncertain — sometimes too small, which makes an ordinary difference look impressive. The t-distribution is the correction for exactly that, and it is the reason a t-test is not just a z-test with different letters.
The correction shows up as heavier tails. A value of 2.5 standard errors from zero is a once-in-eighty event on the normal curve; on a t-distribution with five degrees of freedom it is closer to once in eighteen. That extra tail weight is not pessimism for its own sake — it is the honest probability once you allow for the possibility that your variance estimate came in low. The practical consequence is that a small study needs a larger observed difference to clear the same significance bar, which is precisely the discipline a small study ought to be held to.
The shape is controlled entirely by one parameter, the degrees of freedom, and it converges on the normal curve quickly. At 5 df the two-sided 95% critical value is 2.571 against the normal's 1.960 — a penalty of about 31%. At 30 df it is 2.042, a penalty of 4%. By 200 df it is 1.972, and the difference has stopped mattering to any decision anyone would make. This is why the choice between t and z is nearly irrelevant at online-experiment sample sizes and genuinely important in a survey of forty customers.
It is worth being clear about what the distribution describes, because the name invites a misreading. It is not a claim about the shape of your data. It is the distribution of a *statistic* — the standardised difference between means — under the assumption that nothing is going on. Your revenue figures can be as skewed as they like; what the central limit theorem and this curve together handle is the behaviour of the summary computed from them.
The same curve does more work than significance testing alone. Every confidence interval around a mean is built by multiplying a standard error by a critical value taken from it, which is why a 95% interval from a small sample is wider than the 1.96 multiplier most people carry in their heads. Getting that multiplier from the normal curve when the sample is small produces intervals that are too narrow and a false sense of precision.
The density itself is rarely worth writing out — what matters in practice is the critical value it produces and how fast that value settles down as the sample grows.
t = ( x̄ − μ ) / ( s / √n )The `s` is the giveaway: an estimated standard deviation rather than a known one. Replace s with a known σ and this is a z statistic.
df = 5 → 2.571 df = 30 → 2.042 df = 200 → 1.972 normal → 1.960The whole practical content of the distribution. The gap is 31% at five df and under 1% by two hundred.
Var(T) = df / (df − 2), for df > 2Larger than the normal's 1, and undefined below three df. That excess is the heavier tails stated numerically.
x̄ ± t_(df, α/2) · ( s / √n )The multiplier comes from this curve, not from 1.96 — see the two-sample t-test calculator.
A B2B team surveys 12 enterprise accounts on hours saved per week by a new workflow. The sample mean is 6.4 hours with a standard deviation of 3.1. They want a 95% confidence interval, and the question is what multiplier to use — the familiar 1.96, or the one this sample actually earns.
The correct 95% interval is 6.4 ± 2.201 × 0.895, or 4.43 to 8.37 hours. Using 1.96 would have given 4.65 to 8.15.
The difference looks modest and is not cosmetic. The normal-curve interval is 12% too narrow, which means an interval sold as 95% confidence is delivering closer to 91% — and every claim built on it inherits that overstatement. At twelve accounts the penalty is real; had this been 120 accounts it would have been under 1% and not worth a sentence. The lesson is not that small samples are unusable but that they must pay for their own uncertainty. Note also what the interval does not fix: twelve self-selected enterprise accounts may not represent the customer base at all, and no critical value corrects for [selection bias](/glossary/selection-bias).
Applying it to a live measurement problem is the part that goes wrong. If you are designing an experiment, reading a result you do not trust, or trying to work out what your marketing actually caused, that is the work we do.
