Hello everyone,
I have a question regarding the analyses of moderation effects with PLS that has previously been answered inconsistently here on the forum (as far as I could find). I would appreciate if anyone could clear this up, ideally with references to published articles.
When testing for moderation in a purely reflective model, is it advisable to use
a) the whole model (e.g., several DVs with differential predictors and moderators) or
b) a reduced model with only predictor, moderator, dependent, and the product-interaction latent?
I have seen both forms advised here on the forum e.g., this in favor of A and this in favor of B
Apparently, however, A results in possible problems with multicollinearity (this has also happened in my model).
Thank you for your help!
Moderation: full vs. reduced model
-
khauschildt
- PLS Junior User
- Posts: 8
- Joined: Wed Dec 09, 2009 10:39 am
- Real name and title:
Moderation: full vs. reduced model
Last edited by khauschildt on Mon Mar 08, 2010 7:42 pm, edited 2 times in total.
-
stephan.kramer
- PLS Expert User
- Posts: 37
- Joined: Tue Mar 25, 2008 9:16 am
- Real name and title:
-
Paolo Pinkel
- PLS Junior User
- Posts: 6
- Joined: Wed Feb 04, 2009 2:23 am
- Real name and title:
- Location: Shanghai
Hey there,
Based on my experience so far, it seems like PLS has a lot in common with hierarchical regression analysis when it comes to moderator effects (Tabachnick, B.G./Fidell, L.S. (2006): Using Multivariate Statistics. 5th ed., Harper Collins College Publishers, New York 2006).
In this respect, you first have to analyze and interprete your main effects model without the moderation effects. This means however, that you also have to include a path from your moderator Z to the DV. This is the case, even if you previously did not hypothesize such a relationship. In the interpretation you can just skip that path.
(see e.g. Sarkar; Echambadi; Harrison (2001) Alliance entrepreneurship and firm market performance. In Strategic Management Journal; Walter; Auer; Ritter (2006) The impact of network capabilities and entrepreneurial orientation on university spin-off performance. In Journal of Business Venturing 21(4))
Then you have to analyze your model including the moderation effect. In this second model you should only interprete the interaction effects. Do not interpret the main effects model.
(see Carte; Russell (2003) In pursuit of moderation nine common errors and their solutions. In MIS Quarterly, 27(3); Eggert; Fassot; Helm (2005) Identifizierung und Quantifizierung mediiernder und moderierender Effekte in komplexen Kausalstrukturen. In Handbuch PLS-Pfadmodellierung)
You can find out the significance of your interaction model by computing F-values (Tabachnick, B.G./Fidell, L.S. (2006): Using Multivariate Statistics. 5th ed., Harper Collins College Publishers, New York 2006, p. 148) and f² values.
Yet, the betas and the t-values of your orginal endogenous DVs should not show high changes in the interaction effect model. THis means, that a DV that was highly significant in your main effects model should remain so in your interaction model. If not, you might have multi-collinearity problems.
I hope this helps.
Based on my experience so far, it seems like PLS has a lot in common with hierarchical regression analysis when it comes to moderator effects (Tabachnick, B.G./Fidell, L.S. (2006): Using Multivariate Statistics. 5th ed., Harper Collins College Publishers, New York 2006).
In this respect, you first have to analyze and interprete your main effects model without the moderation effects. This means however, that you also have to include a path from your moderator Z to the DV. This is the case, even if you previously did not hypothesize such a relationship. In the interpretation you can just skip that path.
(see e.g. Sarkar; Echambadi; Harrison (2001) Alliance entrepreneurship and firm market performance. In Strategic Management Journal; Walter; Auer; Ritter (2006) The impact of network capabilities and entrepreneurial orientation on university spin-off performance. In Journal of Business Venturing 21(4))
Then you have to analyze your model including the moderation effect. In this second model you should only interprete the interaction effects. Do not interpret the main effects model.
(see Carte; Russell (2003) In pursuit of moderation nine common errors and their solutions. In MIS Quarterly, 27(3); Eggert; Fassot; Helm (2005) Identifizierung und Quantifizierung mediiernder und moderierender Effekte in komplexen Kausalstrukturen. In Handbuch PLS-Pfadmodellierung)
You can find out the significance of your interaction model by computing F-values (Tabachnick, B.G./Fidell, L.S. (2006): Using Multivariate Statistics. 5th ed., Harper Collins College Publishers, New York 2006, p. 148) and f² values.
Yet, the betas and the t-values of your orginal endogenous DVs should not show high changes in the interaction effect model. THis means, that a DV that was highly significant in your main effects model should remain so in your interaction model. If not, you might have multi-collinearity problems.
I hope this helps.