Session: Advanced Mediation Analysis (Society for Prevention Research 27th Annual Meeting)

SESSION INTRODUCTION: Statistical mediation analysis is important in understanding how an intervention changes behavior. The goal of this organized poster forum is to present advanced topics in mediation analysis that help researchers optimize their prevention studies. The theme of this poster forum fits within the general theme of advancing prevention research through the use of innovative methods and statistics. Presenters will be coming from different institutions from around the world to present their individual and collaborative researcher on mediation analysis in prevention science.

The first paper, “Total effect decomposition in mediation analysis with a binary outcome” extends recent work on the non-collapsibility of odds ratios in logistic regression in the presence of confounder variables to non-collapsibility of odds ratios in logistic regression in the presence of a mediator variable. Insights from the potential outcomes framework for causal inference provide a decomposition of the total effect that helps distinguish between non-collapsibility and the mediated effect.

The second paper, “Causal direct and indirect effects: The link between traditional and logistic regression and potential outcomes framework for binary outcomes” demonstrates the similarities and differences between the traditional estimators and potential outcomes estimators for mediation analysis with binary outcomes and exposure-mediator interaction, and describes the implications of these similarities and differences for the application of mediation analysis in practice.

The third paper, “Statistical evaluation of tests for suppression as a case of inconsistent mediation” describes how a third variable effect, suppression, may increase the magnitude of effect estimates when a suppressor variable is adjusted for in a statistical analysis. The paper discusses and investigates several methods for detecting the presence of suppression in prevention studies.

The fourth paper, “Potential contributions of machine learning to statistical mediation analysis” describes how to use cutting-edge machine learning algorithms such as the random forest algorithm to correctly identify a mediator and outcome model for a mediation analysis in prevention data. Total, direct, and mediated effects are then defined using the potential outcomes framework for causal inference. Theoretical and practical implications of using machine learning approaches to mediation are discussed.

It is expected that advanced topics in mediation analysis will be appealing to the audience of the 2019 SPR Annual Meeting. At the end of the presentations, the discussant will summarize the main points and theme and moderate a discussion between the attendees and the presenters.