Session: Modern Methods for Mediation Analysis in Prevention (Society for Prevention Research 24th Annual Meeting)

Key aspects in the study of prevention are the development and evaluation of interventions. The first question we generally ask is whether these interventions work. It is also important to understand how the interventions work, in other words, what are the mechanisms through which these interventions have their effects?  Practitioners often have theoretical models of the pathways through which interventions have their influence on outcomes, and these theoretical models should be evaluated. Understanding the mechanisms of action may also allow us to identify aspects of the intervention that are more and less effective. This can help refine interventions, which is another critical aspect of prevention. These mechanistic questions can be addressed using mediation analysis. Mediation analysis in its simplest form partitions an overall effect of an intervention on an outcome into the part of the effect transmitted via a third variable, the mediator, and the remaining direct effect of the intervention on the outcome. The mediated effect is therefore sometimes called an indirect effect.

Recent research in mediation has focused on sources of bias and more robust methods for analysis. Simple mediation analysis does not take into account measurement error or potential unmeasured confounders of the relationships between the variables. In addition, we often take repeated measures when studying interventions but few studies have applied longitudinal mediation models to investigate mediating mechanisms.

Here we present modern mediation methodologies relevant to the study of mediation in prevention research. The first paper studies the properties of various mediation effect size measures and whether the property of monotonicity holds. In the second paper, Monte Carlo simulation is used to examine the optimal method for creating stable inverse propensity weights, addressing unmeasured confounding of the mediator-outcome relation in the common pretest-posttest design. The third paper summarizes measurement issues in mediation and reports simulation studies elucidating aspects of measurement error and differing factor structures between intervention groups on estimation of the mediated effect. In the fourth paper, investigators used structural equation models to study mediation of the effects of a randomized health and safety intervention for law enforcement officers on exhaustion. The fifth paper describes the application of longitudinal structural equation mediation models allowing for measurement error and unmeasured confounding to data from a randomized trial of interventions for chronic fatigue syndrome.  In the final paper, meta-analytic path analysis is employed to assess mediation across several randomized anxiety and depression intervention studies in youth.

This symposium provides examples both of cutting edge mediation methodologies and the application of such methods to answer substantive mediation questions. We showcase critical advancements in the field of modern mediation analysis and how innovations in the methods can help practitioners address interesting mechanistic questions in the field of prevention.