RANK TESTS FOR MULTIVARIATE TREND

From MaRDI portal





Suppose that bivariate observations \((x_ j,y_ j)\), \(j=1,...,n\), are of the form \[ x_ j=n^{-1/2}a_ 1(j/n)+e'_ j,\quad y_ j=n^{- 1/2}a_ 2(j/n)+e_ j \] where \(a_ 1\), \(a_ 2\) are unknown drift functions and \((e'_ j,e_ j)\) are independent pairs of identically continuously distributed bivariate error random variables. The problem in this paper is to test the hypothesis \(H_ 0: a_ 1=a_ 2=0\) against alternatives which are linear in space, but not necessarily in time: \[ a_ 1(t)=u(t)\cos \gamma,\quad a_ 2(t)=u(t)\sin \gamma \] where \(u(t)\) is a positive monotone function and \(\gamma\) is a fixed angle. By projecting the bivariate data onto a line of direction of \(\theta\) \((0<\theta \leq 2\pi)\), the authors defined \[ N(\theta)=\#(concordant\quad pairs\quad in\quad direction\quad \theta)- \#(disconcordant\quad pairs\quad in\quad direction\quad \theta)= \] \[ \sum_{1\leq i<j\leq n}sgn\{x_ j\cos \theta +y_ j\sin \theta -(x_ i\cos \theta +y_ i\sin \theta)\} \] and considered test statistics \(S_ b=\int^{2\pi}_{0}N(\theta)b(\theta)d\theta\) for some appropriate bounded continuous function b. For instance, \(b(\theta)=\cos (\theta -\gamma)\). The standard test is based on the least squares (LS) regression coefficient of \(x_ j\cos \gamma +y_ j\sin \gamma\) on j. It is shown that \(S_ b\) is a U-statistic with non-null mean and null variance. For the normal case, the asymptotic efficiency of the \(S_ b\) test \((b=\cos (\theta-\gamma))\) relative to the LS test is found to exceed \(3/\pi\). The case when \(\gamma\) is unknown, is also discussed. A numerical example is provided.











This page was built for publication: RANK TESTS FOR MULTIVARIATE TREND

Report a bug (only for logged in users!)Click here to report a bug for this page (MaRDI item Q3223701)