Page 1 of 1

3-Way Interaction of Moderation

Posted: Mon Jul 20, 2015 3:37 am
by alfa.ryano
Hi Everybody,

I made 3-way interaction of moderation using SmartPLS 2. The relationship between ID and LO is moderated by variables G and CS. So, the moderation formula is like this, ID + G + CS + (ID x CS) + (ID x G) + (G x CS) + (ID x CS x G). You can find my model below.

Since SmartPLS doesn't support 3-way moderation directly (I mean we have to compute it manually, for an example, using Excel), I change ID x G, ID x CS, and G x CS into variables and I manually input them into the *.csv file. The result is as follow and it seems 'good', indicated by the 'normal' path coefficients.
1-Multiplication as Variable-BEFORE.jpg
1-Multiplication as Variable-BEFORE.jpg (91.05 KiB) Viewed 21026 times
Moreover, I computed ID x CS x G using Excel and added it as a variable. So, the model looks like as follow.
3-Multiplication as Variables-AFTER.jpg
3-Multiplication as Variables-AFTER.jpg (109.82 KiB) Viewed 21026 times
Surprisingly, the path coefficients get wild. The values are very high. Is this condition is possible?

I then did an experiment. I changed the variables that represent ID x G, ID x CS, and G x CS, not as variables that I manually input into the *.csv file, but as the moderation effects calculated by SmartPLS. Only ID x CS x G acts as a variable. The result is as follow.
4-Multiplication not as Variables-AFTER.jpg
4-Multiplication not as Variables-AFTER.jpg (109.28 KiB) Viewed 21026 times
Surprisingly again, the result, specifically the path coefficients, is not equal to the previous one. Actually, when I removed the ID x CS x G variable, the path coefficients of the model that uses manually-computed multiplication variables and the model that uses SmartPLS-computed multiplication variables are NOT different. They show the same values. Only when I include ID x CS x G variable, the path coefficients of the two models show different values.

So, two questions here:
1. Why are, in the 3-way interaction of moderation, the path coefficient values very big?
2. Why are the two models, the model that use the manually-computed multiplication variables and the model that use the SmartPLS-computed multiplication variables, different in their path coefficient values?

Thanks for the help.

Regards,

Alfa

Re: 3-Way Interaction of Moderation

Posted: Sat Jul 25, 2015 3:32 pm
by cringle
Could you please replicate things using SmartPLS 3. We considerably improved the moderator analysis in SmartPLS 3.

Best regards
Christian

Re: 3-Way Interaction of Moderation

Posted: Mon Jul 27, 2015 3:07 pm
by alfa.ryano
Thank you for your Reply, Christian. I will try it using SmartPLS 3. By the way, still related to 3-way moderation, if I have three variables A, B, and C, and if I want to multiply them all, which one of these approaches is the best:

Standardize( Stadardize(A) x Stadardize(B) ) x Stadardize(C)

or

Standardize( Stadardize(A) x Stadardize(B) x Stadardize(C) ) ??

Thank you very much.

Regards,

Alfa

Re: 3-Way Interaction of Moderation

Posted: Tue Jul 28, 2015 2:04 pm
by jmbecker
You do not want to standardize any of the product terms. So it would be just: Standardize(A) x Standardize(B) x Standardize(C)

In SmartPLS3 we make sure that the product term LV is not standardized after multiplication. This, however, will only work for two-way interactions. If you do your caclulations outside of SmartPLS and import the product terms, SmartPLS will not know that it should not standardize the LVs and hence it will do it as it always standadizes normal LVs.
Your only solutation is to take the LV Scores from the direct effects model and calculate the three-way interaction model outside of SmartPLS (two-step approach like).

Re: 3-Way Interaction of Moderation

Posted: Wed Jul 29, 2015 6:30 am
by alfa.ryano
Thank you for the explanation. It helps me much. :)

Regards,

Alfa Ryano

Re: 3-Way Interaction of Moderation

Posted: Mon May 30, 2016 11:02 am
by Sara b
jmbecker wrote:You do not want to standardize any of the product terms. So it would be just: Standardize(A) x Standardize(B) x Standardize(C)

In SmartPLS3 we make sure that the product term LV is not standardized after multiplication. This, however, will only work for two-way interactions. If you do your caclulations outside of SmartPLS and import the product terms, SmartPLS will not know that it should not standardize the LVs and hence it will do it as it always standadizes normal LVs.
Your only solutation is to take the LV Scores from the direct effects model and calculate the three-way interaction model outside of SmartPLS (two-step approach like).
Hello, Do you mean using ANOVA to calculate the three-way interaction?

Re: 3-Way Interaction of Moderation

Posted: Tue Dec 18, 2018 3:54 am
by 592210373@qq.com
jmbecker wrote: Tue Jul 28, 2015 2:04 pm You do not want to standardize any of the product terms. So it would be just: Standardize(A) x Standardize(B) x Standardize(C)

In SmartPLS3 we make sure that the product term LV is not standardized after multiplication. This, however, will only work for two-way interactions. If you do your caclulations outside of SmartPLS and import the product terms, SmartPLS will not know that it should not standardize the LVs and hence it will do it as it always standadizes normal LVs.
Your only solutation is to take the LV Scores from the direct effects model and calculate the three-way interaction model outside of SmartPLS (two-step approach like).
Dear Professor Becker, for three-way interaction, I can understand from your above explanation that "to take the LV Scores (of first meoderator(s) from the direct effects model, like what we do using two-step approach', right? However, I do not understand what you mean of "calculate the three-way interaction model outside of SmartPLS". You mean we can not do three way interaction in Smartpls,am i right?

So where and how to do it?
I hope we can do it in Smartpls, if so, in the three way interaction model , all the effects and moderator(s) are also needed, as the above example,right?

I am still new to statists and also Smartpls, and has confused for sometime.

Thank you for your elaboration and your patience!

Best,
Marina

Re: 3-Way Interaction of Moderation

Posted: Tue Dec 18, 2018 8:48 am
by jmbecker
You cannot do a three way interaction in SmartPLS.

You need to do it in a software capable of doing regression analysis and bootstrapping such as SPSS. You need to export the latent variable scores from the main effects model (without the interaction) and then create the product scores in that software (by multiplication). You then do the structural model regression with the software and look at the unstandardized coefficients.

Re: 3-Way Interaction of Moderation

Posted: Wed Dec 19, 2018 1:35 pm
by 592210373@qq.com
Ok, I will do the three way interaction in SPSS.
Thank you very much. I really appreciate it!

Best,
Marina

Re: 3-Way Interaction of Moderation

Posted: Tue Jul 12, 2022 3:43 am
by ferry.h.f@gmail.com
Hi,
Can you please share if you have done the 3-way interaction for moderators on SPSS? Is there any step by step guideline for it?
Regards,
Ferry

Re: 3-Way Interaction of Moderation

Posted: Fri Jul 29, 2022 11:33 am
by Hwsm
592210373@qq.com wrote: Wed Dec 19, 2018 1:35 pm Ok, I will do the three way interaction in SPSS.
Thank you very much. I really appreciate it!

Best,
Marina
Hi Marina, did you manage to do the three-way interaction?
If so, could you please share how?
Greetings,
Hannah

Re: 3-Way Interaction of Moderation

Posted: Mon Aug 01, 2022 8:07 am
by jmbecker
The new SmartPLS 4 will support three-way interactions (and even higher interactions) without the need to go outside the software. The pre-release is already available for testing: https://www.smartpls.com/smartpls4/

Re: 3-Way Interaction of Moderation

Posted: Tue Aug 02, 2022 8:33 pm
by Hwsm
jmbecker wrote: Mon Aug 01, 2022 8:07 am The new SmartPLS 4 will support three-way interactions (and even higher interactions) without the need to go outside the software. The pre-release is already available for testing: https://www.smartpls.com/smartpls4/
Dear Mr. Becker,

I am desperately trying to calculate higher order interaction terms (SmartPls4).
But how can I identify if the interaction is significant... with path coefficients or conditional direct effects?

Is it correct to perform this analysis using the PLS model rather than the process model?
Unfortunately, the results I get are not significant except for a few interaction terms, leaving me confused, as I had expected the majority to be significant.




I would be very thankful, if you could help me.
Greetings,
Hannah
Bildschirmfoto 2022-08-02 um 22.38.11-1.png
Bildschirmfoto 2022-08-02 um 22.38.11-1.png (148.19 KiB) Viewed 94 times

Re: 3-Way Interaction of Moderation

Posted: Wed Aug 03, 2022 11:57 am
by jmbecker
You can determine if the interaction is significant by looking at the three-way product terms in bootstrapping.

Assessment of conditional direct effects can give you an additional idea of how and where the interaction has its effects, i.e., how it changes the effects and which of these specific conditional effects are significant.

However, your model seems quite complex; with so many interactions in one model, it is not surprising that not many (any) are significant. You would need a lot of data to robustly estimate such a model. Regardless of using PLS or Process.

I would recommend to rethink whether all paths need to be moderated. I would also expect that your theoretical development must be quite complex, theorizing a three-way interaction for all of these effects.

Re: 3-Way Interaction of Moderation

Posted: Thu Aug 04, 2022 12:40 pm
by Hwsm
Hello Dr. Becker,

thank you so much for your reply!
These informations were very helpful for me.
Regarding the complexity of my model – I underestimated it, but unfortunately I am already in the midst of my bachelor thesis.

After reducing my model, I am getting some significant results in the PROCESS in conditional indirect effects.
Could you maybe advise me to which extent I can interpret these values?
Because the individual effects are found to be significant in conditional indirect effects, but they are not significant in path coefficients.

One of my hypotheses is:
Age and experience will moderate the effect of habit (GE) on behavioral intention (NA), so that the effect will be stronger for younger people with less experience.

Am I correct in accepting this hypothesis?
Because for GE -> NA conditioned on EXPERIENCE at -1 SD and m_age at -1 SD the path coefficient is largest, and also significant. (highlighted orange).
I have added a screenshot of the results for habit (GE).

I would be very thankful if you could help me.
Process Bootstrap-screen.png
Process Bootstrap-screen.png (233.69 KiB) Viewed 72 times
Geetings,
Hannah