Cramér-type moderate deviations for Studentized two-sample \(U\)-statistics with applications (Q342663)

Two-sample \(U\)-statistics are important quantities for nonparametric statistical inference, in particular when two populations shall be compared with respect to a statistical functional. If variance heterogeneity among the groups has to be assumed, Studentized versions of these statistics are commonly used. The authors analyze the asymptotic distributional behavior of a Studentized two-sample \(U\)-statistic \(\widehat{U}_{\bar{n}}\) by establishing a Cramér-type moderate deviation theorem which quantifies the (asymptotic) accuracy of the normal approximation to the distribution of \(\widehat{U}_{\bar{n}}\) in tail areas. Two examples of special interest are the two-sample \(t\)-statistic and the Mann-Whitney \(U\)-statistic. The results are applied to analyse the false discovery rate (FDR) of the linear step-up (LSU) test by \textit{Y. Benjamini} and \textit{Y. Hochberg} [J. R. Stat. Soc., Ser. B 57, No. 1, 289--300 (1995; Zbl 0809.62014)] for high-dimensional two-sample comparisons. The authors consider the cases that this multiple test is carried out by means of \(p\)-values which result from applying a normal or a bootstrap approximation to the null distribution of (marginal) two-sample \(t\)- or Mann-Whitney \(U\)-statistics, respectively. It turns out that in such cases the LSU test asymptotically controls the FDR whenever the number of tests is of asymptotic order bounded by \(\exp\{o(n^{1/3})\}\), where \(n = n_1 \wedge n_2\) denotes the minimum of the group-specific sample sizes. Finally, a numerical study illustrates the authors' theoretical findings.