Those confounded moderators in meta-analysis: good, bad, and ugly

Abstract

One of the more revealing aspects of meta-analysis is the relationship between effect sizes and moderator variables representing differences among studies in their methods, samples, and interventions. However, interesting moderator variables are generally related to each other as well as to effect sizes. This confounding among moderator variables can make the results of any analysis focusing on a single moderator variable misleading. Despite the ambiguity of such results, this form of analysis is common in meta-analytic studies of intervention. The hazards and complexities of investigating and interpreting confounded moderator variables are illustrated in this article by the author's examining the difference in effect sizes associated with randomized versus nonrandomized designs in a large meta-analytic database of delinquency intervention studies.