Is Multiple Regression Moribund in the Social Sciences and the Legerdemain of Continuous Variables

• In Event: Modeling and Analytical Issues

Abstract

Advancements in structural equation modeling related programs (e.g., Mplus, AMOS) have improved the estimation and evaluation of indirect effects (i.e., mediation and moderation). However, a review of recent criminological literature shows that some studies apply traditional OLS regression when path analysis is the preferred method. The first goal of this paper is to demonstrate why path analysis is the preferred method when compared to traditional OLS regression techniques, and in particular, to show the advantage of parsing the error term when modeling indirect effects. With regard to creating groups out of continuous variable(s), MacCallum et al (2002) demonstrated the disadvantages of using median-split for a continuous variable to create “arbitrary grouping.” Therefore, the second goal of this is paper is to extend this argument by demonstrating how “arbitrary grouping” of variables commonly used in the field of criminology is problematic for modeling continuous variables within multiple regression. Our analyses demonstrate that when natural groups occur, ANOVA and its counterparts are preferred; otherwise, path analysis is more robust when estimating indirect effect models with continuous predictors due to the ability to model additional error terms.

Authors

• Anthony Vander Horst, Kent State University

• Carlton Mathis, Kent State University