|
|
Bimonthly Since 1986 |
ISSN 1004-9037
|
|
|
|
|
Publication Details |
Edited by: Editorial Board of Journal of Data Acquisition and Processing
P.O. Box 2704, Beijing 100190, P.R. China
Sponsored by: Institute of Computing Technology, CAS & China Computer Federation
Undertaken by: Institute of Computing Technology, CAS
Published by: SCIENCE PRESS, BEIJING, CHINA
Distributed by:
China: All Local Post Offices
|
|
|
|
|
|
|
|
|
|
Abstract
The quest for unified solution for handling multicollinearity and autocorrelation jointly has resulted in a great deal of interest in two-stage approaches within the linear modelling paradigm recently. This paper contributes to this growing interest with Two-Stage Ridge Quantile Regression (TRQR) and Two-Stage Lasso Quantile Regression (TLQR) methods. The two-stage methods fundamentally involve the application of appropriate transformation to linear regression data followed by regularization to control respectively, autocorrelation and multicollinearity. The aim is to determine if regularized and robust regression methods reduce total model error, and find which method is most effective for dealing with autocorrelation and multicollinearity problems. The utility of the methods were assessed using simulations based on models with small (2) to relatively large (8) predictors, in comparison with other regression methods using the Mean Squared Error criterion. The results indicate that the TRQR method, at 0.5 quantile level, is most suitable for handling multicollinearity and autocorrelation problems with many (8 or more) predictor variables. However, the Two-Stage Ridge Regression method performs better with few (2 or less) predictor variables. The findings show that each method is affected by sample size, number of predictors or multicollinearity level.
Keyword
Autocorrelation, Multicollinearity, Two-Stage method, Ridge regression, Lasso regression, Mean squared error
PDF Download (click here)
|
|
|
|
|