scientific article; zbMATH DE number 4043082

From MaRDI portal
Publication:3780294





Consider a linear but heteroscedastic regression model \(X_{ni}=c^ T_{ni} \beta_ n+C^ T_{ni}\gamma_ n\epsilon_ i\), \(1\leq i\leq n\), with i.i.d. residuals \(\epsilon_ i\), parameter vectors \(\beta_ n,\gamma_ n\in {\mathbb{R}}^ p\) and a design matrix \(C_ n=(c_{n1}..c_{nn})^ T\in {\mathbb{R}}^{n\times p}\) which satisfies some asymptotic conditions. As special cases this model contains with \(\gamma^ T_ n=(1,0,...,0)\) a linear homeoscedastic regression model, and with \(p=1\) a location-scale model. Following \textit{R. Koenker} and \textit{G. Basset}, Econometrica 46, 33-50 (1978; Zbl 0373.62038) the author studies \(\alpha\)-regression quantiles \({\hat \beta}_ n(\alpha)\), defined as a solution of \[ \min_{\beta \in {\mathbb{R}}^ p}\sum \phi_{\alpha}(X_{ni}-c^ T_{ni}\beta),\text{ where } \phi_{\alpha}(t)=(1-\alpha)| t| \cdot \chi (t<0)+\alpha | t| \chi (t\geq 0). \] These can be computed as solutions of a linear program. Now consider the solutions \((\hat a_{ni}(\alpha)\), \(1\leq i\leq n)\) of essentially the dual program which are called dual \(\alpha\)-regression quantiles and can be interpreted as rank-scores. The author investigates the asymptotic behaviour of the associated quantile processes and shows e.g. weak convergence to transformed Brownian bridges. Furthermore he considers certain L-estimates, which are defined with help of the quantile functions, studies their statistical and especially asymptotic behaviour and discusses the relations to many special attempts in this context.











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 Q3780294)