Abstract
Mediation analysis has become a very popular approach in psychology, and it is one that is associated with multiple perspectives that are often at odds, often implicitly. Explicitly discussing these perspectives and their motivations, advantages, and disadvantages can help to provide clarity to conversations and research regarding the use and refinement of mediation models. We discuss five such pairs of perspectives on mediation analysis, their associated advantages and disadvantages, and their implications: with vs. without a mediation hypothesis, specific effects vs. a global model, directness vs. indirectness of causation, effect size vs. null hypothesis testing, and hypothesized vs. alternative explanations. Discussion of the perspectives is facilitated by a small simulation study. Some philosophical and linguistic considerations are briefly discussed, as well as some other perspectives we do not develop here.
Highlights
Without respect to a given statistical model, mediation processes are framed in terms of intermediate variables between an independent variable and a dependent variable, with a minimum of three variables required in total: X, M, and Y, where X is the independent variable (IV), Y is the dependent variable (DV), and M is the mediator variable that is supposed to transmit the causal effect of X to Y
The aim of a mediation study can either be to find ways to change the level of the dependent variable, or the aim can be to understand the process through which the independent variable affects the dependent variable, or the purpose of the research may be prediction
The approach further focuses on effect sizes over Null hypothesis significance testing (NHST), and states that causal inferences should not be drawn from observational data for reasons similar to those we provide in the discussion of the hypothesized vs. alternative explanations section
Summary
Without respect to a given statistical model, mediation processes are framed in terms of intermediate variables between an independent variable and a dependent variable, with a minimum of three variables required in total: X, M, and Y, where X is the independent variable (IV), Y is the dependent variable (DV), and M is the (hypothesized) mediator variable that is supposed to transmit the causal effect of X to Y. As still further evidence of the difficulty of making mediation claims, parameter bias, and sensitivity have emerged as common concerns (e.g., Sobel, 2008; Imai et al, 2010; VanderWeele, 2010; Fritz et al, 2016), as has statistical power for testing both indirect (e.g., Shrout and Bolger, 2002; Fritz and MacKinnon, 2007; Preacher and Hayes, 2008) and total effects (Kenny and Judd, 2014; Loeys et al, 2015; O’Rourke and MacKinnon, 2015)
Sign-in/Register to view highlights & summary 
Published Version (
Free)
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a call