On the selection of variables in criminology

Abstract

Objectives: The inclusion of variables that: (1) are descendants of unobserved confounders and (2) do not theoretically or empirically cause variation in the constructs of interest in a multivariable regression model can potentially adjust for the bias generated from unobserved confounders. Nevertheless, the validity and utility of descendants for criminological research has yet to be evaluated. Methods: Two studies were developed to address the gap in the literature. First, a randomly specified directed equation simulation analysis was performed. Second, using data from the Pathways to Desistance study, the technique was implemented to observe if the association between gang involvement and criminal involvement was attenuated after adjusting for exposure to violence (a potential descendant of an unobserved confounder). Results: The simulation analysis demonstrated that adjusting for the descendants of unobserved confounders can reduce bias in key estimates. The magnitude of the association between gang involvement and criminal involvement was approximately half of the bivariate association after introducing exposure to violence into the model. Conclusions: The findings suggest that adjusting statistical models for variables that are a descendant of an unobserved confounder and do not cause variation in the association of interest can reduce the bias generated by an unobserved confounder.