Mediation Wars: SEM v. Baron & Kenny

Researchers testing mediation models are likely at least passingly familiar with the Baron & Kenny approach to mediation. The Baron & Kenny (1986) approach involves running several regressions; a test for mediation using Baron & Kenny's approach might involve establishing: a) a relationship between X -- the predictor -- and Y -- the outcome, b) a relationship between X and M -- the mediator, c) a relationship between M and Y, and d) a relationship between M and Y, controlling for X. Of course, not all analyses using the Baron & Kenny approach require all four of these conditions to be met (e.g., X and Y do not necessarily need to be directly correlated), and, to support full mediation, an additional step is required -- establishing that the relationship between X and Y dissapears when controlling for M. Advances in computing power and statistical programs, however, have made another method of mediation testing popular: testing mediation using structural equation modeling (SEM). If you are testing -- or thinking of testing -- a mediation model, you'll definitely want to check out James, Mulaik, and Brett's "A Tale of Two Methods."

James et al. (2006) compare the SEM approach to testing mediation with the Baron & Kenny approach. If you are testing a partially mediated model (e.g., X affects Y both directly, but also through M), the authors argue that these two approaches are essentially equivalent. However, if you are testing a fully mediated model (e.g., X affects Y only through M), these approaches test complete mediation differently. To guide researchers testing fully and partially mediated models, James et al. (2006) supply information regarding how to test for full and partial mediation using each of these techniques.

Of course, the aspiring mediation researcher is primarily interested in the question: "So, which is the best method for me to use when testing mediation?" Fortunately, James et al. (2006) provide pretty clear guidance on this question.

How should researchers test mediation?
1. First, revisit your theory. Is your mediation model of interest full or partial mediation? If your theory is not well-developed enough to guide this question, James et al. (2006) recommend testing for full mediation to satisfy the principle of parsimony -- although they also confess that partial mediation, being common, may be the most practical model to test. Our recommendation? Don't use James et al.'s (2006) guidance as an excuse to not think about your theory. Before you default to testing full mediation, invest time thinking about your theory and about which of these models make more sense.
2. If possible, test your model using SEM rather than Baron & Kenny. What would make this possible? Sample size is one deciding factor -- SEM is typically better suited for analyses of larger samples (perhaps 100 or above) rather than smaller samples (for example, 20-40 observations).

In short, if you want to test a mediation model, James et al. (2006) is a must-read!

How have you tested for mediation in the past?

References
Baron, R. M. & Kenny, D. A. (1986). The moderator-mediator variable distinction in social psychological research: Conceptual, strategic, and statistical considerations. Journal of Personality and Social Psychology, 51, 1173-1182.
James, L. R., Mulaik, S. A., & Brett, J. M. (2006). A tale of two methods. Organizational Research Methods, 9, 233-244.