Portal de Periódicos da CAPES

Confidence Intervals for Regression Quantiles

1994; Linguagem: Inglês

10.1007/978-3-642-57984-4_29

2628-8966

Roger Koenker,

Statistical and numerical algorithms

Several methods to construct confidence intervals for regression quan-tile estimators (Koenker and Bassett (1978)) are reviewed. Direct estimation of the asymptotic covariance matrix requires an estimate of the reciprocal of the error density (sparsity function) at the quantite of interest; some recent work on bandwidth selection for this problem will be discussed. Several versions of the bootstrap for quantile regression will be described as well as a recent proposal by Parzen, Wei, and Ying (1992) for resampling from the (approximately pivotal) estimating equation. Finally, we will describe a new approach based on inversion of a rank test suggested by Gutenbrunner, Jurečková, Koenker, and Portnoy (1993) and introduced in Hušková(1994). The latter approach has several advantages: it may be computed relatively efficiently, it is consistent under certain heteroskedastic conditions and it circumvents any explicit estimation of the sparsity function. A small monte-carlo experiment is employed to compare the competing methods.