General composite quantile regression: Theory and methods. Issue 9 (2nd May 2020)

• Record Type: Journal Article
• Authors: Wu, Yanke; Tian, Maozai; Tang, Man-Lai
• Is Part Of: Communications in statistics. Volume 49:Issue 9(2020)
• Page Start: 2217
• Page End: 2236
• Publication Date: 2020-05-02
• Subjects: Asymptotic relative efficiency -- general composite quantile regression -- oracle property -- weight function; Mathematical statistics -- Periodicals; Mathematics; Statistics; 519.2
• DOI: 10.1080/03610926.2019.1568493
• Languages: English
• ISSNs: 0361-0926

Abstract

Abstract: In this article, we propose a new regression method called general composite quantile regression (GCQR) which releases the unrealistic finite error variance assumption being imposed by the traditional least squares (LS) method. Unlike the recently proposed composite quantile regression (CQR) method, our proposed GCQR allows any continuous non-uniform density/weight function. As a result, determination of the number of uniform quantile positions is not required. Most importantly, the proposed GCQR criterion can be readily transformed to a linear programing problem, which substantially reduces the computing time. Our theoretical and empirical results show that the GCQR is generally efficient than the CQR and LS if the weight function is appropriately chosen. The oracle properties of the penalized GCQR are also provided. Our simulation results are consistent with the derived theoretical findings. A real data example is analyzed to demonstrate our methodologies.