Student's T Distribution

Student's t distribution is the probability distribution of a sample mean's standardized value when the population standard deviation is unknown and has to be estimated from the same small sample, and it looks like a standard-normal-distribution with fatter tails. It was derived by William Sealy Gosset, a chemist and statistician at the Guinness brewery in Dublin, who needed reliable conclusions from the tiny samples that barrel-scale experiments allowed (barley varieties, brewing yields) and published the result in Biometrika in 1908, in the paper The Probable Error of a Mean, under the pseudonym Student. Ronald Fisher later gave the distribution a rigorous footing and tabulated it, and the resulting t-test became one of the most widely used procedures in empirical science.

The extra width exists because the estimated standard deviation is itself random, and in a small sample it can come out too small by luck, making the apparent z-score too large. The t statistic is (sample mean − μ) / (s / √n), a ratio of a normal quantity to an estimated spread, and the distribution that results depends on a single parameter, the degrees of freedom. For a one-sample mean, degrees of freedom are n − 1: the sample has n values, but once the sample mean is fixed, only n − 1 of the deviations from it are free to vary, which is the same reason the sample variance divides by n − 1. Degrees of freedom therefore measure how much independent information the estimate of the spread rests on, and the fewer there are, the fatter the tails.

As degrees of freedom grow, the t distribution converges to the normal, and at the other extreme it becomes pathological. With 1 degree of freedom it is the Cauchy distribution, which has no finite mean; with 2 degrees of freedom the variance is infinite; for more than 2, the variance is df/(df − 2), always slightly larger than the normal's 1. The practical effect shows in the critical values for a 95% two-sided interval: about 12.706 at 1 degree of freedom, 4.303 at 2, 2.571 at 5, 2.228 at 10, 2.042 at 30 and the normal's 1.960 in the limit. So the normal 1.96 is a fine approximation at moderate sample sizes, but with n = 6 a 95% confidence-interval for a mean must use 2.571, about 31% wider.

The t distribution underlies the workhorse procedures of applied statistics: the one-sample, paired and two-sample t-tests, confidence intervals for means, and the t statistics reported for regression coefficients. Welch's variant handles two groups with unequal variances using an approximate, non-integer degrees of freedom, and the t distribution is the reference behind the typical significance stars in regression output. Beyond testing, the t distribution has become a deliberate modeling choice for heavy-tailed data: robust regression and Bayesian models use a t likelihood so that outliers pull the fit less than a normal would, and t-SNE, the visualization method of van der Maaten and Hinton (2008), uses a t distribution with 1 degree of freedom in the low-dimensional map to let moderate distances be stretched apart and reduce crowding.

Common mistakes are to think the t-test is only for small samples and to forget its assumptions. It is correct for any n when the standard deviation is estimated; the t and normal values simply become indistinguishable as n grows, which is why large-sample practice uses 1.96. The exact t distribution also assumes the data (or the means) are approximately normal, so with tiny samples from strongly skewed or outlier-ridden data, the stated 95% coverage can be off; larger samples lean on the central-limit-theorem to rescue the approximation, and non-parametric or bootstrap methods are alternatives. Another confusion is reading the degrees of freedom as the sample size: they are n − 1 for one mean, n − 2 for simple regression slope tests, and a fractional value in Welch's test.

Read Student's t distribution as the normal curve's honest, small-sample cousin, derived by a Guinness brewer in 1908 to price in the extra uncertainty of estimating the standard deviation from the same few observations, with degrees of freedom as the dial that sets how fat the tails are and which fades to the normal as the data accumulate.