Efficiency Robust Two-Sample Rank Tests
Journal of the American Statistical Association · 1967 · Volume 62 · Issue 320
DOI: 10.1080/01621459.1967.10500929
Abstract
In the classical two sample problem, the null hypothesis of identical distribution functions is tested. Rank testing procedures have the property that they are valid (have Type I error identically equal to α) for any absolutely continuous distribution function. Choices among rank tests are usually based upon local power calculations for a given specified null distribution against specific parametric alternatives. A rank test is said to be efficiency robust when no other has a uniformly better local power performance (risk point) for a specified class A of distribution function alternatives F. Such rank tests may be constructed using formal Bayes procedures resulting in convex mixtures of the locally most powerful rank tests T F for F εΛ. Asymptotic relative efficiency (ARE) calculations uncover the symmetry relation ARE ( T F , T F *; F ) = ARE ( T F *, T F ; F *). Examples involving Logistic, Normal and Double Exponential are considered.
Enable JavaScript for citation tools and additional paper interactions.