Non-significant P-values? Strategies to Better Determine the Importance of Effects and Interactions in GLM

• In Event: Innovations in Quantitative Modeling

Abstract

Previous research has shown that logistic regression with explicitly specified interactions produces difficult to interpret results. The conventional approach to measuring the effects based on significance testing using p-values (e.g. in association with the Wald test or using the deviance to assess model fit) is not always appropriate in the context of GLMs with interactions. The objective of this paper is to demonstrate the limitations of the conventional approach and to explore alternative strategies for determining the importance of effects and interactions in logistic regression.

To illustrate the main points in this paper we provide examples of two studies concerned with different aspects of cyber abuse victimization (a skeletal R commands necessary to conduct the analyses is also provided). Analysis of our example datasets shows that converting binary variables into a multi-level factor works well in small dimension models; Bayesian analysis of the logistic regression with vaguely informative priors is a suitable alternative requiring some expertise in its implementation; Bayesian Model Averaging (BMA) is especially suited for new areas of research regardless the dimension of the model. We recommend that criminologists consider including the BMA in their standard toolbox for analysis of GLMs with interactions.

Authors

• Zarina Vakhitova, Griffith University

• Clair Alston-Knox, Griffith University