Experimentally manipulating mediating processes: Why and how to examine mediation using statistical moderation analyses

Statistical mediation analysis is commonly used to examine mediation, but it is not the default paradigm; researchers also test for mediation through experimental mediation analysis, such as the two randomized experiments design, the experimental-causal-chain design, the moderation-of-process design, and the parallel design, all of which differ considerably in terms of procedures and requirements. Which requirements are genuinely necessary, and which are not? This paper compares the effectiveness of these research designs in examining mediation. Three constitutive requirements for supporting a mediational hypothesis were identified: (A) a significant interaction effect of the independent variable (X) and the manipulation of the proposed mediating process (M) on the dependent variable (Y); (B) a significant effect of X on the measured M within the control group whose M is not manipulated and can function naturally; and (C) a significant effect of the manipulation on the measured M. Using these criteria, existing designs all have drawbacks, so this paper proposes a manipulation-of-mediation-as-a-moderator (MMM) design to fulfill all three requirements. MMM provides strong evidence for the causal inference M → Y, avoids false alarms in many cases, and provides direct evidence for the relationships between X and M and between manipulated M and measured M. The paper presents a step-by-step example of MMM for interested practitioners. In its discussion of the relationships among X-caused M, manipulated M, and measured M and the distinction between mediation and moderation, this paper enriches the understanding of the nature of mediation analyses in psychology.