Causal Mediation Approaches for the Pretest-Posttest Control Group Design in Prevention Research

Introduction: The two-wave randomized pretest-posttest design is a common prevention research design. Even in designs with multiple waves of data, research hypotheses often focus on effects at different times after delivery of an intervention. Traditional mediation analysis of this design uses path analysis, difference score, or residualized change scores to investigate the process by which an intervention (X) changes a mediator (M) when then affects an outcome (Y). The purpose of this presentation is to describe modern causal inference approaches for assessing mediation in this design. The causal inference approach for two-wave mediation models differs from traditional two-wave models, primarily in that new methods, such as inverse probability weighting (IPW) and g-estimation remove confounding of the post-test relation of M and Y using the baseline measures of M and Y. These modern methods are important because randomization of interventions does not guarantee causal interpretation of a mediation effect.  The overall goal of this study is to develop first-principle understanding of causal inference for the two-wave model and to develop solutions to enhance interpretation of mediation effects in prevention research. 

Methods: Theoretical, conceptual, and statistical issues in the use of baseline measures to remove confounding will be described in a potential outcomes framework. Inverse probability weighting (IPW) uses a weighting method to remove confounding of the M measure. Similarly, g-estimation provides more accurate estimates of the direct effect and calculates the mediated effect as the difference between total and direct effects. The results of a simulation study investigating sample size, parameter value, and confounding effects will be reported. The methods were applied to mediation analysis of an intervention to reduce intentions to use steroids among high school football players. 

Results: The steroid prevention study application of IPW and g-estimation reduced the size of the relation of the mediated effect to the outcome variable but did not appreciably change research conclusions. Simulation results for the different methods are planned that will vary sample size and effect size of stabilities and mediated effects. 

Conclusions: The potential outcomes methods may improve investigation of mediated effects in the pretest-posttest control group design. The methods are relatively easy to apply but require consideration of model assumptions such as whether measures of all confounders are available and whether there are interactions between baseline measures and the intervention effect.