forum.smartpls.com

Dear all,

I have tested two moderators in SmartPLS but struggle to confirm my hypotheses correctly.

Case 1: multi group analysis (variable has one indicator which is categorical)
Case 2: moderator effect (variable has 8 indicators which are continuous)

In both cases, the hypothesis is:
"M strengthens the postive relationship between X and Y."

Case 1:
The SmartPLS MGA reveals a very significant path for the "higher group" and a not significant path for the "lower group." Since there is no significant difference between both groups, I would conclude that there is no moderation effect. However, the "higher group" does have an impact. Can I write something like "partly confirmed"? Out of curiosity, I (only as test) run this variable simply as moderator effect, which is positive and highly significant ... I know that both approaches are different but I was then wondering if I should better use the simple moderation effect ... ?

Case 2:
The SmartPLS moderator analysis reveals a highly significant effect. The higher M the better the effect between X and Y. The slope plot tells me that the effect is clearly reversed for a low degree of moderator, just like the classic example in the book of Hair et al., (2017, p. 259). After reading books and papers, I conclude that in this case the hypothesis is simply confirmed and the crossing is just shortly mentioned. But I feel this is not 100% correct ... Or is only the steeper line relevant (significant?) for the confirmation and I am just too picky about the flatter line? In this case, I would have written "partly confirmed" to avoid that the reader thinks that both a low and high level of M is positive for the effect X and Y because without the slope one can only guess ...

I was wondering if my hypothesis formulation is not clear enough but I've seen this style multiple times ...

What do you think?

I hope you know what I mean and would be glad to find someone who can help me!

Kind regards,
Marcella

Case 1: Generally, it sound like that there could be moderation, but you may not have enough power (e.g., because of splitting the data into two smaller groups) when using a MGA to find a significant difference.
If your moderator variables has only two values (0/1) then I would also consider it as a normal moderator variable (interaction term). It works well and is an easy and simple, yet slightly more powerful approach. The advantage is also that you include the mean differences in the DV because of your moderator via your direct effect of the moderator.
One problem, however, could be measurement invariance. If you measures are not equivalent across groups that could pose a problem to the MGA, but also the moderating approach (which forces equivalence, which might not be given).

Case 2: If you interaction term (moderating effect) is significant you have moderation. The effect of X on Y depends on the level of M. So your hypothesis is confirmed.
The next questions is: When is the effect of X on Y significant? --> at which levels of M.
This is a simple slope analysis. The estimated direct effect of X on Y is the simple effect of X on Y, when M is zero (i.e., at its mean, because we have standardized data). That may be significant or not.
In the chart we are also showing the simple slopes of M at +1 and -1 standard deviation. Which is basically the direct effect + and – the moderating effect. Whether these represent meaningful values of M depends on the data (a standard deviation could be outside of the measured range of M, but could also represent only a fraction of the range of M depending on its variability).
You can check whether these simple slopes are significant using the sample estimates from bootstrapping and some easy extra calculations (for example in MS Excel). You may write me an e-mail to get an excel spread sheet that does this for the corporate reputation example.
You could also use adaptations of the Johnson-Neyman point to find the values of M where the effect turns significant/insignificant.
But these are all just additional analysis to better understand the moderation. The initial finding is that you have a moderation. These would all not make sense if the moderating effect is not significant.

Hey, thank you so much for the information.
I am also currently trying to do a moderator analysis. However I don't quite understand how to find out at what level (M+1 SD, M-1 SD) the Moderation is significant. Is there a way to get this information in SmartPLSE? or is it possible and does it make sense to use Process, in SPSS for that, eventhough I am doing the rest of my analysis in SmartPLS? If this is not the case, how do I find at what the moderation is significant ?

Thank you so much.

No it is currently not possible to get this in SmartPLS. We are experimenting with the Johnson-Neyman technique, but this relies on distributional assumptions. This does not really fit the PLS context. If you want to try that (even though it might not be optimal it could suggest you some results) you can write me an e-mail and I can provide you an Excel sheet with an example for the Corporate Reputation model.

Hello,
I'm currently working on a research paper using PLS-SEM. In line with the conversation below I'm wanting to do a johnson and neyman analysis but am not able to find how and what. as I read, you do not think it fits the PLS context could you elaborate on this? and if possible could you provide me the excel that you mention?

Is MGA applicable only for groups which are equally divided? I have a categorical variable gender which consists only two groups male and female. But the number of females and number of males is not equal can I still apply MGA. Similarly I have many other categorical variables which are binary(0/1) in nature for example education. Education has three groups ( bachelors, masters, doctotorate) all three are binary (0/1) . Number of participants in each category is not equal. Can I run simple moderation or MGA can be used to study the effects.

The groups do not need to be equal. However, it can cause problems if the groups are very imbalanced and one group is extremely small (<=10%).
In addition, if you do MGA, your smallest group should still fulfill the minimum sample size requirements.