Addressing Confounder Bias in Mediation Models

Introduction:  Causal inference continues to be an important area of research with regard to statistics and methods in prevention research.  Recent research in causal inference with regards to statistical mediation has focused on addressing the sequential ignorability assumption; specifically, that there is no confounding between the mediator and the outcome variable. This poster focuses on comparing and contrasting three different methods for assessing sensitivity to confounding presented in Imai et al. (2010) VanderWeele (2010) and Mauro (1990) as well as the graphical depiction of these methods. 

Methods:  Using data from a large intervention study to decrease steroid use among high school athletes, we demonstrate two situations in single mediator models, where confounding is likely and unlikely to occur. Data consisted of 1506 adolescent football players from 31 high schools in Oregon and Washington.  Mediation models assessed whether team as an information source was a significant mediator of: 1) the relationship between the intervention and strength training self-efficacy for the first example and 2) the relationship between the intervention and nutrition for the second example. 

Results:  In examples one and two, significant mediation occurred (ab = 0.184, ab = 0.065 respectively), but for example two, the three methods for assessing confounder bias indicated one or more confounders could cause the mediated effect to become zero. The graphs depicting these two examples are examined and the interpretations of these graphs are explained.

Conclusions:  We discuss potential confounders for the given examples and we discuss broadly how these how these methods could be implemented in other future mediation studies.  We also talk about the limitation of using this method including dealing with more complex mediation models such as multiple mediator models or nonparameteric models.  Additionally, we address situations where the three methods may provide conflicting results; such as when one method suggests that confounding is likely and another suggests confounding is not likely.  Lastly, we briefly discuss several methods for dealing with confounding in mediation models such as the instrumental variable method and propensity score analysis.