A Small Sample Power Study of the Anderson-Darling Statistic and a Comparison with the Kolmogorov and the Cramer-Von Mises Statistics

The Anderson-Darling goodness-of-fit procedure emphasizes agreement between the data and the hypothesized distribution in the extremes or tails. An improved table of the quantiles of the Anderson-Darling statistic, useful for small sample sizes, was constructed using the Cray-2 supercomputer. The power of the Anderson-Darling test is compared to the Kolmogorov and the Cramer-von Mises tests when the null hypothesis is the normal distribution and the alternative distributions are the Cauchy, the double exponential, and the extreme value distributions.