Probability & Statistics · Student’s t-distribution

Student’s t-distribution

The t-distribution with ν degrees of freedom is the distribution of the standardized mean (x̄ − μ)/(s/√n) of n = ν + 1 values drawn independently from the same normal population N(μ, σ²), when σ is replaced by the sample’s s. It looks like the standard normal but has heavier tails; as ν grows it approaches the normal.

t-distribution in this visualization

When nn values are drawn independently from the same normal population N(μ,σ2)N(\mu, \sigma^2), the standardized mean (xˉ−μ)/(σ/n)(\bar{x} - \mu)/(\sigma/\sqrt{n}) is standard normal. In practice σ\sigma is unknown and the sample’s own standard deviation ss takes its place; the result T=(xˉ−μ)/(s/n)T = (\bar{x} - \mu)/(s/\sqrt{n}) follows Student’s t-distribution with ν=n−1\nu = n - 1 degrees of freedom. Its density is fν(t)∝(1+t2/ν)−(ν+1)/2f_\nu(t) \propto (1 + t^2/\nu)^{-(\nu+1)/2}. Here it is drawn in red over the standard normal in ink, with the tails beyond a cutoff you choose and the two 95% critical values.

Why the tails are heavier

The sample standard deviation varies from sample to sample, and when it happens to come out small the ratio is blown up. Large values of TT are therefore more common than large values of a normal variable, and the gap grows the further out you look: the normal density falls like e−t2/2e^{-t^2/2}, the t density only like a power of tt. So the 95% critical value t∗t^* is larger than 1.96, and a confidence interval or a test that uses 1.96 with ss claims more confidence than it has.

As the degrees of freedom grow

With more data, ss settles close to σ\sigma, and the t-distribution approaches the standard normal. The approach is gradual: t∗−z∗t^* - z^* (with z∗≈1.96z^* \approx 1.96) shrinks roughly in proportion to 1/ν1/\nu, so once ν\nu is large each doubling about halves it; no particular sample size makes the two the same. The exact result needs independent values from one normal population; for other data the t-distribution is only an approximation, which improves as the sample mean becomes closer to normal.

Things to notice

  • With one degree of freedom the t-distribution is the Cauchy distribution, whose mean does not exist.
  • Thirty is not a magic number: at ν = 29 the critical value is still about 2.045.

Controls

  • Drag the triangle on the axis to move the cutoff; the left and right arrow keys move it by 0.01.

Related