Non-null semi-parametric inference for the Mann–Whitney measure

Abstract

A simple method is introduced for finding large sample, boundary-respecting confidence intervals (CIs) for the two-sample Mann–Whitney measure, θ=Pr{X>Y}−Pr{X<Y}. This natural separation measure for two distributions occurs in stress–strength models, receiver operating characteristic curves, and nonparametrics generally. The usual estimate of θ is a centred version of the well-known Mann–Whitney statistic. Previous Wald-type CIs are not boundary-respecting. The difficulty is typically nonparametric, whereby appealing exact distributions hold only for one null parameter value, preventing the formulation of true distribution-free inference for non-null values. Here, the rank method setting and a result, that stochastic ordering is equivalent to monotone transformation of location shift, allow the assumption that data derive from a smooth location shift family. A suitable class of location shift families then model the asymptotic variance, leading to a rapidly converging iterative CI method based on roots of quadratics. Simulations show that the proposed method performs at least as well, or better, than competing CI methods.

Keywords:

• boundary-respecting confidence interval
• extended logistic family
• generalised hyperbolic secant distribution
• location shift
• stochastic ordering

AMS Subject Classification :

• 62G10
• 62G15