Mediation: Direct path gets stronger through mediation?

Frotee · Post by **Frotee** » Thu Apr 25, 2013 10:35 am

Another interesting result has popped up in my model:
I have two mediating variables in my model.
Testing one of them for the difference of the direct effect with and without the mediator variable, I noticed that the direct effect got stronger when I included the mediator variable – whose effect, when included as a mediator, is insignificant. If I Include the mediator variable as an external variable, its effect becomes significant.

How does that work and what does it tell me? Shouldn't a mediator variable usually reduce the direkt effect when included into the model?

christian.nitzl · Post by **christian.nitzl** » Fri Apr 26, 2013 9:26 am

…seems to me that it is a correlation problem between your latent constructs. In a regression model a coefficient can be also become higher when you introduce a new variable. This is because a high correlation between the variables. Therefore we test VIF (variance inflation factor).
I ask myself how do you measure your latent construct (formative or reflective; likert-scale or binary)? This could be a reason why this problem appeared.
Greetings,
Christian

Frotee · Post by **Frotee** » Sat Apr 27, 2013 11:24 am

The effect occurred more strongly when I measured the latent construct formatively on the basis of a 5-point likert scale (however, it also happens when measuring reflectively). There was indeed a correlation problem between the variables, which is why I have since decided to measure it reflectively after all.

I also noticed that the path loadings as well as the R^2 of my endogeneous latent construct changed depending on whether I measured the latet construct meantioned in the beginning formatively or reflectively.
I'm curious: How is that, and what does it mean?

christian.nitzl · Post by **christian.nitzl** » Wed May 08, 2013 10:18 am

…when using reflective measurements you control the measurement error. In contrast, for formative measurements you don’t control this. This could be one reason for the differences. There are many articles who discussing the differences between reflective and formative measurements. You can find also some information in German in my working paper on pp. 5:

http://papers.ssrn.com/sol3/papers.cfm? ... id=2097324

Greetings,
Christian