Mediation analysis, multiple mediators

Comments

• evankesteren

  February 2021

  Hi Sarah,

  that's a great question.

  1. Including all three subscales as mediators would take into account the covariances among the mediators. The indirect effect of the subscale is then interpreted as "indirect effect given that the other subscales have already been taken into account". This "conditional mediation effect" (and its p-value) may be what you want.
  2. Performing three univariate mediation analyses will tell you how strong the mediation effect is for each subscale, not taking into account the other subscales. This "marginal mediation effect" may also be what you want, for example if you want to decide to measure only one subscale later.
  3. There is an alternative option which requires the SEM module: make the mediator a latent variable, with as indicators the three subscales. SEM will then estimate a latent variable which "does its best" to mediate the relation between X and Y, and you will get factor loadings which indicate the contribution of each subscale (and their p-values). I feel that this last approach is the most natural, closest model to the research question you posed. It does require that you know something about SEM / lavaan, but this would be a great reason to learn (if you haven't yet!)

  The model syntax for this last approach would be something like this:

  mediator =~ subscale1 + subscale2 + subscale3

  mediator ~ alpha*X

  Y ~ beta*mediator + direct_effect*X

  indirect_effect := alpha*beta

  
  

  I hope this helps!
• Sarah_Webb

  February 2021

  Thank you so much, this helps a lot!