MerQuR+Statistics Glossary

Statistics Glossary → Anderson-Darling test

Anderson-Darling testconcept

A goodness-of-fit test of whether a sample comes from a specified distribution (usually the normal), based on a comparison of cumulative distributions that gives special weight to the tails.

The test sums the squared differences between the empirical cumulative distribution function of the data and the theoretical cumulative distribution over the whole range of values. The Kolmogorov-Smirnov test looks only at the largest difference and is more sensitive in the middle of the distribution; Anderson-Darling divides the differences by F(x)(1 − F(x)) and so gives more weight to departures in the tails. The null hypothesis is 'the data come from a normal distribution'; the larger the A² statistic, the worse the fit.

When the mean and standard deviation are estimated from the data, the critical values must be adjusted accordingly; software uses approaches such as Stephens' corrections for this. For testing normality, Anderson-Darling usually has power close to Shapiro-Wilk and higher than Kolmogorov-Smirnov. Like all normality tests, it can flag trivial departures in large samples and miss serious ones in small samples, so the result should be read together with a Q-Q plot.

An everyday example

A laboratory wants to know whether the errors of a measuring device are normally distributed. Most errors look normal, but very large errors occur now and then. Because the fit in the middle is good, Kolmogorov-Smirnov may miss the problem, whereas Anderson-Darling, which weights the tails, is more likely to detect these extreme errors.

In MerQur

It is offered in Normality Tests as one of the tests of whether a variable, or each group level, departs from the normal distribution.

Analyses where this concept appears

Normality Tests

Related concepts

Shapiro-Wilk testNormality assumption