Detecting Correlational Outliers: Is something other than the Mahalanobis Squared Distance needed ?

• Activity Number: 45
• Type: Contributed
• Date/Time: Sunday, August 11, 2002 : 4:00 PM to 5:50 PM
• Sponsor: Social Statistics Section*
• Author(s): Paul Ritter*+
• Affiliation(s): University of Texas, Austin
• Keywords: Multivariate Outliers ; Mahalanobis Squared Distance ; Comrey's D ; Correlational Outliers ; Outlier Detection ; Principal Components

Abstract

Detecting outliers with quantitative, non-graphical means has received some attention. Specifically, concern has been expressed as to whether the Mahalanobis squared distance (MSD) can adequately identify outliers that will influence factor analyses. This concern is consistent with a theme in some of the literature on multivariate outliers that detection methods should be tailored to specific situations. Andrew Comrey, a pychologist and factor analyst, developed a metric, Comrey's D (CD) that purportedly did a better job at detecting outliers that would influence a factor analysis than the MSD. A couple of Monte Carlo studies have been conducted in which the MSD and CD were compared, including one by Donald Bacon that involved the detection of "correlational outliers." I am in the process of conducting Monte Carlo studies which compare MSD, CD, and other metrics, including principal component scores under more varied and realistic condtions. I hypothesize that MSD will perform well enough in comparison to other more specialized metrics to rethink the concern about its being able to detect outliers that would affect a factor analysis such as "correlational outliers."