Bug found in the moderation effect using Smart PLS 4

Questions about the implementation and application of the PLS-SEM method, that are not related to the usage of the SmartPLS software.
Post Reply
Beishanto
PLS Junior User
Posts: 1
Joined: Tue Dec 20, 2022 12:11 pm
Real name and title: Dr. Victor Temprano-Garcia

Bug found in the moderation effect using Smart PLS 4

Post by Beishanto »

Good afternoon,
We have found a bug in the moderation effect using Smart PLS 4.0.8.7. This bug occurs in both PLS-SEM and PROCESS.
When we run the moderation, the moderating effects coefficients are correct. However, the coefficients of the direct relationships influenced by the moderator variable are wrongly calculated.
In the case of a continuous moderator variable, the value of this coefficient is at -1 SD within the mean slope, meanwhile, it should be the value in the mean slope corresponding to zero.
In the case of a dichotomous moderator variable, the value of this coefficient is at -1 SD within the zero group slope, meanwhile, it should be the mean value in both groups (0 and 1).
For further information regarding our data, please contact us. We appreciate it if you can help us out with this problem.
Best Regards,
Dr. Hector Perez-Fernandez and Dr. Victor Temprano-Garcia
Bistyca
PLS Junior User
Posts: 1
Joined: Mon Dec 26, 2022 7:37 am
Real name and title: Habe

Re: Bug found in the moderation effect using Smart PLS 4

Post by Bistyca »

Has anyone responded yet, I'm curious how to fix this.
tunnel rush
jmbecker
SmartPLS Developer
Posts: 1282
Joined: Tue Mar 28, 2006 11:09 am
Real name and title: Dr. Jan-Michael Becker

Re: Bug found in the moderation effect using Smart PLS 4

Post by jmbecker »

Thank you very much for the detailed information. We also already had a direct exchange in the SmartPLS support tool about this, pointing out that you are too quick to come to the alarming conclusion that it is a bug. That's more likely a misunderstanding.

(1) In the SmartPLS data view (double click on the dataset and click on the Setup button in the menu), you can flag variables as metric, ordinal, categorical, and binary. If you flag a variable as binary, SmartPLS will always assume dummy coding. In addition, SmartPLS 4 has changed the way how it treats binary variables (i.e., dummy coding) since version 4.0.8.3. Now, we always normalize binary variables to 0/1 coding, even if the original coding is different (for example 1/2). This has a reason. When you use a binary variable with 1/2 coding as moderator, you get a conditional effect result for the binary variable being 0 – also see our explanation under (2). But 0 is not a value in the 1/2 coded binary variable, which makes the results interpretation quite difficult. To help users with easier interpretation of their results, we now normalize all variables flagged as binary to dummy coding (0/1). Hence, versions before 4.0.8.3, may report unintuitive results for conditional direct effect when the variable was not appropriately coded as 0/1 by the user (i.e., the result are statistically correct, but not meaningful).

Note: SmartPLS leaves variables that are flagged as binary always unstandardized to get correct dummy coding even if you use standardized results otherwise.

(2) Moderated regression analysis and likewise PLS-SEM always provide the conditional effect of the predictor when a moderator is included. The effect is conditional on the moderator being 0. Thus, in your example, the effect of the REL (being X) on CHAN (being Y) is always the conditional effect when the moderator GENDER (being M) is zero.

This comes from the following regression equation:
Y = a + (b1 + b3 M) X + b2 M which equals Y = a + b1 X + b2 M + b3 X*M

Thus, you can see that the simple effect b1 of X on Y as estimated by regression is the conditional effect of X on Y when M is zero (and thus no addition or subtraction of b3).

(3) If you like to use a different reference point – instead of the 0 binary category - such as the mean value of the binary categories (whatever that ends up being interpreted by the user, e.g. mean gender as a reference point instead of male or female), you should not use dummy coding, but effect coding or something similar. You will get this result in SmartPLS when you do not flag a variable as binary in the SmartPLS but leave it metric (double click on the dataset and use the Setup button in the menu). Then SmartPLS uses your original variable values for your computations, which may be effect coded or contrast coded or whatever you like. However, metric variables will always be standardized if you request standardized results in the algorithm. In these cases, careful interpretation is warranted.

Please also refer to our latest publication where we also have a chapter on moderation: Becker, J.-M., Cheah, J. H., Gholamzade, R., Ringle, C. M., & Sarstedt, M. (2022). PLS-SEM’s Most Wanted Guidance. International Journal of Contemporary Hospitality Management, forthcoming. https://doi.org/10.1108/IJCHM-04-2022-0474

Or get this article on ResearchGate: https://www.researchgate.net/publicatio ... d_guidance

Does this explanation sound reasonable and explain the results in your example? We are happy to hear back from you on this issue.
Dr. Jan-Michael Becker, BI Norwegian Business School, SmartPLS Developer
Researchgate: https://www.researchgate.net/profile/Jan_Michael_Becker
GoogleScholar: http://scholar.google.de/citations?user ... AAAJ&hl=de
Post Reply