Session: Advances in Mediation Analysis: Improving Causal Interpretation and Accuracy of Estimates (Society for Prevention Research 23rd Annual Meeting)

Mediation analysis is a critical tool in prevention science to understand the underlying mechanisms of a treatment and to address competing explanations. A randomized experiment solely focusing on group differences in outcomes is generally not sufficient to reveal how a program accomplished its effects. In a single mediator model, X (e.g., a categorical variable indicating treatment versus control groups) causes a mediator (M), and M causes the outcome variable (Y). The goal of this paper symposium is to address challenges and advances in estimating accurate path estimates, increasing statistical power, and enhancing the causal interpretation in mediation analysis. 

The first paper investigates modern causal inference methods as applied to mediation. Only randomizing the treatment is not adequate to achieve causal effects in mediation since the mediator to outcome relationship may still be affected by confounding variables (i.e., the mediator status is not randomly assigned, but rather self-selected by participants). By conducting a simulation study, authors examine the bias in estimates and statistical power when (1) post-treatment confounders are included in the analysis, and (2) some post-treatment confounders are omitted from the analysis. They compare the linear regression approach to mediation with inverse propensity weighting, g-estimation, and doubly robust g-estimation methods. 

The second paper examines the causal inference approaches to two-wave mediation models where pretest measures of M and Y are used to remove confounding for the relation between the mediator and outcome. The authors describe a simulation study to investigate the statistical performance of different causal methods when testing mediation, and apply them to a real prevention dataset. 

The third paper investigates the accuracy of mediation path estimates and statistical power of mediation analysis from a measurement perspective. Authors examine when the violations to measurement invariance in the mediator are a threat to the conclusions of the mediator model. They conduct a Monte Carlo simulation study in which they manipulate the proportion of noninvariant items, the magnitude of the violations, sample size, and the effect size of mediation path coefficients. 

The symposium aims to promote the implementation of modern statistical methods to enhance the statistical accuracy and causal inference in mediation studies in prevention research.