Interpretation of interaction term / sign changes

Questions about the implementation and application of the PLS-SEM method, that are not related to the usage of the SmartPLS software.
Post Reply
stephan.kramer
PLS Expert User
Posts: 37
Joined: Tue Mar 25, 2008 9:16 am
Real name and title:

Interpretation of interaction term / sign changes

Post by stephan.kramer »

Dear all,

I have three important questions concerning moderation in PLS. I have two reflective independent LV's (A,B) that predict one reflective (dependent) LV (C). I created an interaction term (A*B) and used the standardised scores for term creation as recommended by Chin et al. (2003). My questions now are:

(1) Can the path coefficients be interpreted like betas in regression analysis? In my model, I have a negative path coefficient from the interaction term (C) and two positive path coefficients (main effects) of the independent LV's (A,B). So, if my moderating variable increases about 1 std.dev, my predictor-variable decreases about the path coefficient of the path coefficient of the interaction term? Is that correct?

(2) Is PLS similar to moderated regression analysis in the way that the main effects of the predictor- and the moderating variable (A,B) can not be interpreted meaningful?

(3) Another interesting effect I observed is that the sign-change-option has a huge impact on the significance of the interaction term. While for the LV's the sign change option usually had a decent effect in my previous models, it possessed a huge effect on the interaction term. For example, in one model it boosted the t-value from 0.8 to 3.xx when changing from construct-level to individual change... Does anybody have an explanation for that? And, even more important, is individual sign change a suitable approach for these kinds of models?

Any comments would be highly appreciated!!!

Thanks a lot & best regards,

Stephan
stephan.kramer
PLS Expert User
Posts: 37
Joined: Tue Mar 25, 2008 9:16 am
Real name and title:

Post by stephan.kramer »

Hi all,

I just wanted to bring this topic back on the agenda, because I am really concerned about it.

My big problem is the following: I have a model which consists of the IV's A and B, an interaction term A*B, and a DV C.

What I oberserve is that in my model there is a positive path coefficient from the interaction term A*B on the dependent variable C.

However, if I perform a subgroup-test by half-splitting my sample according to the value of the latent variable score of B, and perform an analysis of A on C in both subgroups, under B(high) the path coefficient from A to C is slightly negative (but non-significant), while the path coefficient from A to C under B(low) is positive and significant. As far as I understood, this would indicate a negative interaction.

In addition to these observations, if I remember correctly, Christian Ringle once taught us in the PLS-course he held at WHU that you have to be cautious with the interpretation of the sign of the interaction term. But I am not really sure about this anymore...

If somebody has an opinion on that, I would really appreciate any comments on that topic, please!!!

Thank you and best regards,

Stephan
stephan.kramer
PLS Expert User
Posts: 37
Joined: Tue Mar 25, 2008 9:16 am
Real name and title:

Post by stephan.kramer »

In case anybody is interested in this topic:

1) The sign and value of the path coefficient of the interaction term works out fine.

2) The effects described in the previous post can be driven by outliers in the dataset (in my case, one (!!!) outlier-datapoint was responsible for these weird results).

3) Lessons learned: PLS, like all regression methods, is very susceptible to outliers. Outlier analyses should be performed with great care before calculating models in Smart-PLS-

Best regards and thanks to Christian Ringle for the very helpful suggestions,
Stephan
Post Reply