Screening and selection for quantile regression using an alternative measure of variable importance (Q2274955)

The authors consider a linear quantile regression model. They propose an effective and computationally simple variable screening/selection tool that can be used as a valuable alternative to existing approaches. First, by generalizing the marginal quantile utility intoduced by \textit{Y. Wu} and \textit{G. Yin} [Biometrika 102, No. 1, 65--76 (2015; Zbl 1345.62097)], the authors present a criterion called partial quantile utility (PQU). In contrast to the also used quantile partial correlation criterion, PQU applies to both continuous and categorial data. As a second major contribution the authors introduce a quantile forward regression algorithm (QFR) that uses PQU-based ranking to screen important variables from a potential set. The authors' third major contribution, guided by QFR-screening and adopting the quantile Bayesian information criterion, is to introduce a direct variable selection method to further reduce the QFR-screened set. Using simulation experiments the finite-sample performance of the introduced tool is extensively investigated. Finally, the authors present an empirical application with the goal of identifying risk factors that are associated with acute childhood malnutrition.