The Anderson-Darling statistic measures discrepancies between an empirical distribution function and a specified continuous distribution function , giving extra weight to the tails.
For a sample of size
,
let
, where
is the
th order statistic. Then
Under the null hypothesis with fully specified, the transformed observations are uniform
on
. Marsaglia and Marsaglia (2004) compute the limiting
distribution using a series with two-term recurrences and give a simulation-based
correction for finite samples.