A robustification of the sign test under mixing conditions (Q2266317)

For many test statistics, when the observations from which the statistic is derived are not independent, the variance of the statistic is changed. The authors consider the case of the two-sample sign test when the observations are \(\psi\)-mixing as defined by \textit{J. R. Blum}, \textit{D. L. Hanson} and \textit{L. H. Koopmans}, Z. Wahrscheinlichkeitstheor. Verw. Geb. 2, 1-11 (1963; Zbl 0117.356). They obtain a 'robustified' version of this test, the test statistic being modified by the use of a strongly consistent estimator for the variance. The asymptotic power function of this modified test against translation alternatives is obtained and the asymptotic relative efficiency (ARE) against the Neyman-Pearson test computed for \(\psi\)-mixing Gaussian observations. This shows that for \(\psi\)-mixing Gaussian observations the ARE can either increase or decrease when compared to the ARE for independent observations. The modified sign test is also compared with \textit{R. J. Serfling's} robustified version of the Wilcoxon test [Ann. Math. Stat. 39, 1202-1209 (1968; Zbl 0162.219)] against alternatives of the form \(F_{\Delta}=(1- \Delta)F+\Delta F^ 2\). (The Wilcoxon test is optimal on these alternatives for independent observations.) Examples show that also in this case the ARE may change drastically when the observations are not independent.