issue with testing of mediation

[adrifont]

Hello everyone,

I am having an issue with the modeling and interpretation of mediation effects.

My model has 4 constructs, all reflective in nature, which you can see in the attached files. These are namely Humanlikeness(5 indicators), warmth(3 indicators) , verisimilitude (5 indicators) and state empathy (3 indicators).

Based on the litterature, I would expect warmth to fully mediate the relationship between the IV Humanlikeness and the DVs state empathy and verisimilitude . When I run the bootstraping analysis without the direct path from humanlikeness to verisimilitude, the indirect effect through warmth is significant (p=0,02), in support of my hypothesis. However, inclusion of the direct effects,which turn out to be nonsignificant as expected, renders the indirect effect through warmth marginally nonsignificant (now p=0,06). Besides nonsignificance, the path of the direct effect is negative, which cannot really be accomodated by any of my theories and makes interpretation difficult.

My questions therefore:
1)Why does inclusion of the direct effect, render an otherwise significant indirect effect insignificant? my sample size is N=550
2)In this case: do direct effects need to be included in the model, or is hypothesis testing and modeling without the direct effects warranted in the light of what I have just explained?

So you can better see what I mean: attached you can find the models with and without direct effects, together with the bootstrapping output for indirect effects when direct effects are not included in the model.


I would be extremely thankful for any guidance,

Best regards,

Adrian

[cringle]

You may want to take a look at the updated mediation chapter in the 2nd edition of the PLS-SEM book:
https://www.smartpls.com/documentation/ ... s-sem-book

@1: When you include the direct relationship, it may explain variance that has before been explained by the indirect effect; hence the latter one becomes smaller. For the mediation analysis, you may want to include the direct effect from the beginning.

@2: Keep the direct effect in the model.

Looking at your results, I would not worry about the significance. Much more important: the coefficients are so small that they almost do not matter (i.e., they have no relevance). This is a key problem (which I have not idea how to solve given your model and data).

Best regards
Christian