On Measuring the Relative Importance of Explanatory Variables in a Logistic Regression

Authors

D. Roland Thomas, Carleton UniversityFollow
PengCheng Zhu, Carelton University
Bruno D. Zumbo, University of British ColumbiaFollow
Shantanu Dutta, University of Ontario Institute of Technology

Abstract

A search is described for valid methods of assessing the importance of explanatory variables in logistic regression, motivated by earlier work on the relationship between corporate governance variables and the issuance of restricted voting shares (RSF). The methods explored are adaptations of Pratt’s (1987) approach for measuring variable importance in simple linear regression, which is based on a special partition of R². Pseudo-R² measures for logistic regression are briefly reviewed, and two measures are selected which can be partitioned in a manner analogous to that used by Pratt. One of these is ultimately selected for the variable importance analysis of the RSF data based on its small sample stability. Confidence intervals for variable importance are obtained using the bootstrap method, and used to draw conclusions regarding the relative importance of the corporate governance variables.

DOI

10.22237/jmasm/1209614580