Comparison and Application of Traditional and Potential Outcomes Framework Methods for Estimating Mediated Effects in a Three-Wave Mediation Model

2^ND PAPER WITHIN ORGANIZED POSTER FORUM

ABSTRACT BODY: Randomized interventions involving a treatment and control group are used to study intervention effects on hypothesized mediators and subsequent effects of hypothesized mediators on drug-use outcomes over multiple measurement waves. For randomized interventions, it follows that intervention precedes both the mediator and outcome but will often be unclear if the mediator precedes the outcome or the outcome precedes the mediator without making strong assumptions or using longitudinal data (MacKinnon, 2008). Additionally, cutting edge causal inference methods from the Potential Outcomes Framework conceptualize mediated effects as the difference between two potential outcomes (i.e., “What is the mediated effect had all participants from intervention group been in the control group instead?”) which aligns more naturally with how prevention scientist think about causal inference. Potential outcomes methods are also the only way to handle time-varying confounders. Therefore, the application of the potential outcomes framework to longitudinal mediation models emphasizes both the temporal nature and modern causal inference interpretation of mediating processes. The study uses a Monte Carlo simulation to compare the performance of traditional regression-based methods (e.g., Analysis of Covariance) and potential outcomes-based methods for estimating longitudinal mediated effects and applies the methods to the Athletes Training and Learning to Avoid Steroids dataset (ATLAS; Goldberg, et al., 1996).

A Monte Carlo simulation study was used to investigate the statistical performance of traditional regression-based methods (e.g., Analysis of covariance, difference scores, residualized change scores) and potential outcomes methods (e.g., inverse propensity weighting, sequential G-estimation) for estimating longitudinal mediated effects across three waves with a randomized intervention. The Monte Carlo study had a sample size 200 and parameter sizes of the mediated effect corresponding approximately to zero, small, medium, and large effect sizes.

Simulation results suggest similar performance between traditional and potential outcomes methods in many simulation conditions. The differences between methods were observed when the mediator and outcome had high stabilities and there were cross-lagged relations (e.g., mediator at pretest affects outcome at follow-up). Application to the ATLAS dataset revealed some overlap between traditional and potential outcomes methods across mediated effect estimates.

Potential outcomes methods perform well, and similarly to traditional methods, in estimating longitudinal mediated effects but also provide the only way to adjust for time-varying confounders in mediation models.