scientific article; zbMATH DE number 3909538

From MaRDI portal
Publication:3685841





[For the entire collection see Zbl 0542.00012.] Assume that in a linear regression model \(Y_{ij}=\theta_ i+\beta_ ix_{ij}+\epsilon_{ij}\), \(i=0,...,k\); \(j=1,...,n\) (sample size) with non-specified error distribution \({\mathcal L}(\epsilon_{ij})\), it is known that \(\theta_ 1\leq...\leq \theta_ k\). It is desired to determine those i's for which \(\theta_ 0\leq \theta_ i.\) Nonparametric rank-based estimates \({\hat \theta}{}_ 0\) of \(\theta_ 0\) and \({\tilde \theta}{}_ i\) of \(\theta_ i\), for \(1\leq i\leq k\), are considered such that \({\tilde \theta}{}_ 1\leq...\leq {\tilde \theta}_ k\). It is shown that for a given \(0<P^*<1\), constants \(d_ 1,...,d_ n\) may be found such that with \[ I=\min \{1\leq i\leq n : {\tilde \theta}_ i\geq {\hat \theta}_ 0-n^{-1/2}d_ i\} \] the selection \(\{I,I+1,...,k\}\) (asymptotically, as \(n\to \infty)\) yields \(P^*\) as the probability of correct selection. The method of proof uses results from \textit{J. N. Adichie}, Ann. Math. Stat. 38, 884-893 and 894- 904 (1967; Zbl 0152.371).











This page was built for publication:

Report a bug (only for logged in users!)Click here to report a bug for this page (MaRDI item Q3685841)