THE PERFORMANCE OF ROBUST DIAGNOSTIC F IN THE IDENTIFICATION OF MULTIPLE HIGH LEVERAGE POINTS

Midi, Habshah and Abu Bakar, Nor Mazlina (2015) THE PERFORMANCE OF ROBUST DIAGNOSTIC F IN THE IDENTIFICATION OF MULTIPLE HIGH LEVERAGE POINTS. Pakistan Journal of Statistics, 31 (5). pp. 461-472. ISSN 1012-9367

[img] Text
31(5)3.pdf - Published Version
Restricted to Registered users only

Download (428Kb) | Request a copy
Official URL: http://www.pakjs.com/

Abstract

High leverage points have undue effects on the Least Square (LS) estimates. They are responsible for misleading conclusions in regression and also multicollinearity problems. Hence, it is imperative to detect high leverage points and use robust estimator to estimate the parameters of a regression model, so as to arrive at valid conclusions. Several well-known methods have failed to detect multiple high leverage points correctly because of the swamping and/or masking effects. The Diagnostic Robust Generalized Potential (DRGP), is an appealing alternative method that successfully detects high leverage points correctly. However, for small percentages of high leverage points, it has the tendency to identify few low leverage points to be points of high leverage. In this paper, an attempt is made to correctly identify real high leverage point by reducing swamping effects. We propose a method we call Robust Diagnostic-F (RDF), in which robust approach is employed to detect the suspected high leverage points and then F statistics that relates the change in data covariance structure is used to confirm the suspicion. The performance of RDF is evaluated through real data and simulations. Comparisons are also made with existing methods.

Item Type: Article
Keywords: Robust, Diagnostic, Outliers, High leverage points, MM estimator
Subjects: Q Science > QA Mathematics
Faculty / Institute: Faculty of Economics & Management Sciences
Depositing User: Pn Nor Mazlina Abu Bakar
Date Deposited: 12 Dec 2015 03:01
Last Modified: 12 Dec 2015 03:01
URI: http://erep.unisza.edu.my/id/eprint/4081

Actions (login required)

View Item View Item