Optimal Rank Tests
Lehmann and Stein (1949) and Hoeffding (1951b) pioneered the development of an optimal theory for nonparametric tests, parallel to that of Neyman and Pearson (1933) and Wald (1949) for parametric testing. They considered nonparametric hypotheses that are invariant under permutations of the variables in multi-sample problems so that rank statistics are the maximal invariants, and extended the Neyman-Pearson and Wald theories for independent observations to the joint density function of the maximal invariants. Terry (1952) and others subsequently implemented and refined Hoeffding’s approach to show that a number of rank tests are locally most powerful at certain alternatives near the null hypothesis. We shall first consider Hoeffding’s change of measure formula and derive some consequences with respect to the two-sample problem. This formula assumes knowledge of the underlying distribution of the random variables and leads to an optimal choice of score functions and subsequently to locally most powerful tests. Hence, for any given underlying distributions, we may obtain the optimal test statistic.
- Hoeffding, W. (1951b). Optimum non-parametric tests. In Proceedings of the Second Berkeley Symposium on Mathematical Statistics and Probability, pages 83–92, Berkeley, Calif. University of California Press.Google Scholar
- Royston, I. (1982). Expected normal order statistics (exact and approximate). Journal of the Royal Statistical Society Series C, 31(2):161–165. Algorithm AS 177.Google Scholar
- Sen, P. (1968a). Asymptotically efficient tests by the method of n rankings. Journal of the Royal Statistical Society Series B, 30:312–317.Google Scholar
- Sen, P. (1968b). Asymptotically efficient tests by the method of n rankings. Journal of the Royal Statistical Society Series B, 30:312–317.Google Scholar