Three Sensitivity-Analysis Methods to Assess Unmeasured Pretreatment Confounding Bias in Experimental Mediation Analysis

Statistical-mediation analysis is a widely used method in psychological research that helps understand the intermediate variables, known as mediators ( M ), by which an independent variable ( X ) causes an outcome variable ( Y ). A major contribution to statistical-mediation analysis has been the incorporation of causal methods because it allows a clear definition of the causal direct and mediated effects and the specification of the assumptions to interpret such effects as causal. Modern causal approaches to mediation analysis encourage routinely investigating the extent to which unobserved confounders may explain the observed mediated effects. The recommendation acknowledges that even when X represents random assignment, participants are not usually randomly assigned to levels of M ; hence, unobserved confounders may bias the M to Y relation ( b -path). In this article, we describe unobserved pretreatment confounding of the M to Y relation in experimental mediation studies and three sensitivity-analysis methods to assess unmeasured pretreatment confounding of the M to Y relation: the correlated-residuals method, the left-out-variables-error method, and the phantom-variable method. We report the results of a simulation study that compares the routine application of the three sensitivity-analysis methods. Results generally indicate that larger effect sizes of the population b -path are less susceptible to confounding bias for all sensitivity methods. Thus, an initial approach to investigating confounding bias in experimental mediation studies is to assess the effect size of the path relating M to Y , and more details can be obtained by applying one of the three sensitivity-analysis methods.