SPSS - Working with Moderator Variables 

Generally speaking it should be no problem but you would have to replicate the idea presented here in the context of a logit regression.

It depends on how you want to interpret your results. If you use standardized variables all units are in standard deviations.

@@jensk.perret6794 Sure, thank you Dr.

This is something you have to decide for running the analysis. The standard is to use 5% or 0.05 as with most other tests.

Maybe check out the "Process" Add In by Hayes. It allows for rather designs with moderators and mediators. The handbook describes the different designs possible. Check out this tutorial on installing and using Process: ru-vid.com/video/%D0%B2%D0%B8%D0%B4%D0%B5%D0%BE-r8x7Kcp4sHs.html

In this context, I would recommend using the PROCESS plug-in which allows for the use of multiple moderators, even in more complex designs. Alternatively, all possible interactions can be calculated by hand and be included in a joined regression.

By using standardized variables you can avoid or at decrease multicollinearity that results from the use of the interaction term (which captures the moderating effect). It will however not reduce multicollinearity from other sources. Also standardization only makes sense if the two variables that are interacting are both metric. For binary variables you can skip the standardization. While a VIF of 10 is given as a critical limit I would become a bit more sceptical starting at VIFs of 2. If your results without standardization are reporting VIFs of 2 or below then I would say you're good to go, otherwise I would recommend sticking to the use of standardized variables.

@@jensk.perret6794 Thank you so much Sir! Your explanations help me a lot. I have been wondering why my IV and all interaction variables have high VIF (260+). After watching this video, I know what's wrong with it since both of my moderating variables are continuous variables and I should use standardized values. No wonder there is no problem with multicollinearity when the variable is a binary variable. Thank you so much!

I think the missunderstanding here is, that in the first part I used two binary (or dummy) variables. In this case the interaction term can simply be calculated by multiplying the two variables. In the case that the variables are non-binary (as is the case in the second part) it is better to standardize them first. I hope this answers your question.

@@jensk.perret6794 here in both cases the variables are gender and tall. Is the standardization has been done to fix multicollinearity

@@steenajose1947 Integrating the interaction effect into a regression model may reduce problems of multicollinearity but it not necessarily has to, it could also make the situation worse.

@@jensk.perret6794 Thank you sir, I got it

Either you can calculate the interaction effects for each moderation and then include all independents and all interactions in a single regression model. If you you are thinking about a more complex model structure I would highly recommend using the Process macro by Hayes. You can easily find it if you Google the terms. The following file gives a good overview over the most important models covered by Process: dm.darden.virginia.edu/ResearchMethods/Templates.pdf

@@jensk.perret6794 Hi Sir, how about if I have several moderators and several independents? When I analyze the moderator effect for each relationship, should I put all the rest of ind and mod in the covariates or what should I do? Thanks.

@@jensk.perret6794 Thanks a lot Sir, your caring build confidence in using SPSS and making me among the SPSS users family. With kind regards: Kaspar Milinga Email: kasparmilinga@gmail.com Mobile: +255621149017 and +255756208373 for calls and WhatsApp

@@jensk.perret6794 Hi, I ran the analysis separately for all indepedent variables to see if there is a moderating effect, the results were all significant but then I read this section of someone asking for models with more than one IVs and you suggested to include all ivs and all interactions. Does that include the moderator variable as well? Also did you mean that to do then the same with Zscores and all standardized interaction effects (including the moderator again?). I ran all these variaations and the significance scores are different.

@@matyastorok1925 If you run your analyses separately you assume that they are indeed independent of each other and that the other variables have no effect. Since your results differ if you consider a full model with all variables I guess that your partial models are not as independent of each other as suspected.

Since the minimum number of participants in each sub-group lies at 30, while results become more reliable if you have at least 50 participants in each sub-group (that is each gender) I would in a variance analyse exclude the participant who checked diverse and work with only male and female participants. If you stick with a moderator analysis you might want to check out the add-on PROCESS by Hayes and use it for your moderator analysis as well as it can cope with the fact that your UV is nominal and has more than two categories. But even in this case you might get seriously biased results and I would opt here as well for excluding the participant you checked diverse.

@@jensk.perret6794 Thank you so much for your help!

A negative sign of the moderator (interaction term) means that the moderator is effect diminishing.

Sir if our moderator effect diminished then how and what type of chsnges can increase significant result ?

You can check if the residuals after regression are normally distributed around a mean of zero. Alternatively, to be on the very safe side you could replace the t-test for the moderator variable by a confidence interval generated via Bootstrapping and check whether zero is contained in the confidence interval.

@@jensk.perret6794 how can I do this ?! Do you have video for it?

The answer would depend on how the result are to be interpreted. Theoretically this would also work.

I think want you are looking for is partial correlations. You will find a number on this topic here.

@@jensk.perret6794 Sorry for not asking it clearly. How do I report the output if I use the method shown in this video? For example when a Pearson correlation test is used we report it with (r,n,p) values. Likewise, what are the values that need to be considered for presenting in the output, if found significant? Is there any links you can provide that could help me out?

@@sarveshvasudevan If the Moderator is part of a larger model, I would report it in a table. I usually use the first row with coefficients and asterisks denoting significance and second row reporting standard errors. If you are interested only in the moderating effect you could report it in-text like "...a significant moderating effect of 0.555 (p=0.003)...". The p-value being the significance level. In either case the reader should at least know something about the size of the effect and the significance.

@@jensk.perret6794 Are there any assumptions that needs to be considered before doing this test? Are the assumptions that are done as part of normal linear regression holds good for this one? Also, would love to see any sample on how to write up the report for this test.

@@sarveshvasudevan If you use moderator analysis in the context of a regression requirements are the same as for a normal linear regression. Note, the use of an interaction term can generate multicollinearity, so use standardized values. The same comments hold if you use the approach in the context of partial correlations based on Pearson correlations.

At least for Pearson correlation you can do something comparable. Just check out partial correlations.

This usually does not significantly benefit the multicollinearity issue and in the end it mainly is a question of interpreting the coefficients. If you want to have a result like: a 1 standard deviation increase in the independent leads to an x standard deviation change in the dependent then I would also go with using standardized variables on both ends.

In the case of this video it is just: Tall - Yes or No.

Exactly.

In your argument you are mixing up a number of points. The standard linear regression approach is build around interval scaled variables but allows for binary independent variables (dummy variables) as well. Logit Models are only important if the dependent variable is binary. Thus, the use of gender (binary) and tall (binary) as independent variables is unproblematic. Multiplying both of them is what a moderator analysis is all about. If gender is 1 if it is a woman and tall is 1 if the person is tall, then the resulting variable is 1 for tall women and 0 else. That is also the way to interpret it. The variable accounts for the additional effect of tall women.