modeling moderating effect when measurement models are mixed

[saab]

I have a basic question here:

I want to model a moderating (not mediating) effect. The MV of the exogenous LV are formative, the MV of the endogenous LV are reflective. The MV of the moderating variable is also reflective.

How do I have to build the interaction term (if this is possible at all), since the particular MV's are in a mixed-mode?

[swende]

I´m interested in this issue too. At the moment, we are planning some convenience functionality for modeling moderating effects in the next SmartPLS version and a clarification on this question could improve our developments.

[stefanbehrens]

I am interested in the same issue. However, despite extensive search, I have not been able to identify any literature that addresses this problem.

Looking at the way interaction terms are created in purely reflective and purely formative models, two possible alternatives seem reasonable:

A) Creating a number of interaction product-indicators by multiplying each indicator of the reflective LV with the LV-score of the formative LV, or
B) Creating a single interaction indicator by multiplying the LV-scores of both LVs (like in the purely formative case).

Given that LV-scores in PLS are simply weighted linear combinations of the LV's indicators, both alternatives should yield the same result. Hence B should be easier and produce less model "clutter".

Following Chin's recommendation, I guess, this approach would also benefit from some monte-carlo-simulation to confirm its validity.

Any thoughts?

Stefan

[e09356]

Looking at the way interaction terms are created in purely reflective and purely formative models, two possible alternatives seem reasonable:

A) Creating a number of interaction product-indicators by multiplying each indicator of the reflective LV with the LV-score of the formative LV, or
B) Creating a single interaction indicator by multiplying the LV-scores of both LVs (like in the purely formative case).

Following Chin's recommendation, I guess, this approach would also benefit from some monte-carlo-simulation to confirm its validity.

This is exactly where new work is required. monte carlo simulations need to be completed here. The work of Robert Ping in CBSEM has shown very high correlations between different derivations of scores for measurment models. You would be expecting very high correlations between methods.

[MichaelKeuper]

The method A seems to be easy to understand and to carry out in excel or SPSS. But could somebody please clarify how to actually carry out method B?

PLS is relatively new to me, so I wonder how I can read out the LV-Scores casewise.

I think (and hope) there must be a convenient way, somehow with the output options and the .lst file, I guess. Is ist the eta-values?

I can't find any literature explaining the procedure in detail for a newby, so maybe somebody here is willing to explain the method in detail to me (and some others).

And I have an similar question to the initial one of this thread: what, if you have everything mixed:

formative as well as reflective exogenous constructs, formative as well as reflective moderators.

And is there a problem to put everything into the model at one time? (I've got 8 exogenous, 3 endogenous and 5 moderator variables) Or is it better to test each moderating effect at once?

Thanks a lot..

[Paolo Pinkel]

Dear Community,

the issue of a mixed moderator model was already discussed a couple of times in this forum, but it seems like nobody has found a bullet-proof solution yet.

For the case that both, the reflective and formative variable have two or more indicators, up to now two approaches have been recommended (See Dr. rer. pol. Stefan Behrens' post from above):

A) Creating a number of interaction product-indicators by multiplying each indicator of the reflective LV with the LV-score of the formative LV, or
B) Creating a single interaction indicator by multiplying the LV-scores of both LVs (like in the purely formative case).

Unfortunately, as to my best knowledge there is no proper solution in the literature. Even the Chin et al. paper does not deal with this special case.

Has anybody experiences with this issue or knows an appropriate solution?

Greetings

[shady]

Hi All,

I am experiencing the same problem as above but could not find a solution. My independant variable is formative, dependant variable is reflective, one moderator is reflective (continuous) and the other moderator is categorical. I am completely new to PLS with zero statistics background. Is their any way to do the analysis?

Additionally, consider that in the above model I only have one moderator (the categorical one, z, with binary values 0 and 1) then how to go about it? Is it possible and valid to divide the data set into two- one data set with all moderator z value= 0 and the other data set with all moderator z value=1, then run them separately (without moderator ) and do group comparison between the two?

[stephan.kramer]

Dear Community,

the issue of a mixed moderator model was already discussed a couple of times in this forum, but it seems like nobody has found a bullet-proof solution yet.

For the case that both, the reflective and formative variable have two or more indicators, up to now two approaches have been recommended (See Dr. rer. pol. Stefan Behrens' post from above):

A) Creating a number of interaction product-indicators by multiplying each indicator of the reflective LV with the LV-score of the formative LV, or
B) Creating a single interaction indicator by multiplying the LV-scores of both LVs (like in the purely formative case).

Unfortunately, as to my best knowledge there is no proper solution in the literature. Even the Chin et al. paper does not deal with this special case.

Has anybody experiences with this issue or knows an appropriate solution?

Greetings

Just wanted to dig this thread out because I am encountering the same problem now...does anybody of the PLS-expert-users have a solid opinion on that topic?

Thanks in advance.

Best regards, Stephan