the F-distribution

Are two groups equally variable? Does adding predictors to a model explain significantly more variation than chance? These questions are really about comparing two estimates of spread — a ratio of variances. The F-distribution is the sampling distribution of exactly such a ratio, and it powers the analysis of variance (ANOVA) and tests for comparing model fits.

An F with d1 and d2 degrees of freedom is built as F = (U/d1) / (V/d2), where U and V are independent chi-squared variables with d1 and d2 degrees of freedom. Since each chi-squared measures summed squared deviations (a variance, up to scale), their ratio compares two variances. The F lives on the positive axis (a ratio of positive things), is right-skewed, and has two separate degree-of-freedom parameters — one for the numerator and one for the denominator, and the order matters. Under the null hypothesis that the two underlying variances are equal, the ratio hovers around 1; an F-value far above 1 is evidence the numerator variance is genuinely larger.

The F-distribution is the backbone of ANOVA, where it compares the variation BETWEEN group means to the variation WITHIN groups: a large F says the groups differ by more than random noise. It also appears in regression overall-significance tests and in comparing nested models. Two honest notes: it connects to the others (the square of a t-variable with k degrees of freedom is an F with 1 and k), and its validity, like the t and chi-squared, leans on the underlying data being approximately normal.