combining indicators / multicollinearity

[thilo38]

Dear Forum,
thanks or this great place. Unfortunately I couldn't find answers to my questions in other threads.

1) in my success factor analysis with smartPLS I have several endogenous formative constructs.
If VIF is > 3 I' ll have to treat multicollinearity, therefore I would combine indicators with a high bivariate correlation into a single new composite indicator by using their average values (HAIR 2014).
If the Outer-VIF for an indicator is < 3 but the respective indicator correlates high (0.72) with another one from the same construct would you combine these indicators, too ?
If yes, which treshold would you set? When is a correlation of two indicators to high? I would like to avoid combing too many indicators due to the loss of information, but on the other hand i am facing some unexpected negative weights...

2) After assessing the measurement models, I would like to compare two groups via MGA and perform an Important-Performance-Analysis for the two groups as well
(Group 1: Nonprofits with low turnover from donations vs Group 2: Nonprofits with high turnover from donations)

To perform the MGA the measurement models should be the same for both groups (same set of indicators for all constructs), so if I combine an indicator due to high correlation with another one in group 1, this have to be done also for the second group, right?

Thank you very much for your help! (question one is the more important one right now ; )
Thilo

[jmbecker]

1)
I would not only blindly combine indicators, but also assess theoretically (and practically) if it makes sense to combine the items. Do they really measure the same (or a very similar) thing? What is the interpretation of the combined item? Does that have any meaning? What was the intention to measure the items separately in the first place?

In addition, you could also try to estimate your model with Mode A (reflective scheme). We show in Becker, J.-M., A. Rai and E. E. Rigdon (2013): Predictive Validity and Formative Measurement in Structural Equation Modeling: Embracing Practical Relevance, in: Proceedings of the International Conference on Information Systems (Milan, Italy), that using Mode A can be a good alternative to using Mode B for formative constructs if multicollinearity is high within some of the indicators of the construct. Both weights and paths are better estimated, especially if your sample size is not very large (i.e., grouping further reduces the available sample size per group).


2) Yes, if you combine indicators in one group, you should also combine them in the other group. Configural invariance is the first requirement for MGA, which requires that the same basic model structure exists in all the groups (in terms of number of constructs and items associated with each construct).

Some information on invariance testing in PLS can be found in
https://www.researchgate.net/publicatio ... st_Squares

The procedure will be available with the next release of SmartPLS 3 (3.2.2), within the next couple if weeks.

[thilo38]

Dear Dr. Becker,
thank you very much for your answer and the provided literature.
You helped me a lot!
All the best from Bonn to Cologne
Thilo