Interaction effects

Frequently asked questions about PLS path modeling.
Schnuffel84
PLS Expert User
Posts: 31
Joined: Mon Jun 14, 2010 7:35 pm
Real name and title:

Interaction effects

Post by Schnuffel84 »

Dear all,

actually I am writing my diploma thesis and I am analyzing my model in SmartPLS.
I have already done the "normal" testing and measuring of my model, but now I want to include moderators like age, sex, income, size of household and involvement. Now I have some questions about the proceedings.
-Can all variables be measured with interaction effects? I have read somewhere that sex has to be measured differently because of being a categorial variable, is that true? Do I have to measure it with multiple group analysis?
- If I want to analyse the other moderators for example the size of household and I want to compare households with less money and households with more money. Do I have to split my variable, and then only analyse it seperately, e.g. first I analyse households with less money and then with more money. Is this the rifht way?
- If I create a moderating effect in SmartPLS, then I see the interaction effects. Does SmartPLS create them on his own or do I have to do it in SPSS or Excel?
- How do I have to measure the interaction effects? I have read somewhere that it is usually that only the moderator, dependent, independent and the interaction term are used and that the other variables in my model which are not affected by the moderator are eliminated. Is that true? Do I have to measure it by oneself?

I would be very grateful if somebody could help me!

Thank you in advance.

Best regards,

Larissa
christian.nitzl
PLS Expert User
Posts: 248
Joined: Sat Jul 25, 2009 1:34 pm
Real name and title:

Post by christian.nitzl »

Dear Larissa,

Especially for your situation I have wrote my working paper (only in German), because I have had the same question as you at the beginning ;-)

On pp. 40 you can find some hints for handle interaction effects with PLS and the most cited literature in that context.

You can easily download the working paper with following link:

http://papers.ssrn.com/sol3/papers.cfm? ... id=2097324

I hope this will help!

Christian
Last edited by christian.nitzl on Thu Jul 05, 2012 1:37 pm, edited 3 times in total.
Schnuffel84
PLS Expert User
Posts: 31
Joined: Mon Jun 14, 2010 7:35 pm
Real name and title:

Post by Schnuffel84 »

Hi Christian,

thank you very much for your help. I will read it now :)


Best regards,
Larissa
Schnuffel84
PLS Expert User
Posts: 31
Joined: Mon Jun 14, 2010 7:35 pm
Real name and title:

Post by Schnuffel84 »

Hi Christian,

now I have read your working paper, but there are still some questions remaining...
-Can all variables be measured with interaction effects? I have read somewhere that sex has to be measured differently because of being a categorial variable, is that true? Do I have to measure it with multiple group analysis?

--> sex has to be measured with multiple group analysis, but how do I have to do it in SmartPLS? How is the proceeding? Can I do it with the "creating moderating effect" button?
- If I want to analyse the other moderators for example the size of household and I want to compare households with less money and households with more money. Do I have to split my variable, and then only analyse it seperately, e.g. first I analyse households with less money and then with more money. Is this the rifht way?

--> Do I need to split the variable, or do I have to do the analysis with the whole variable? The variable hosuehold is for example only measured with one indicant. Or can I see in the output whether the household is "poorer" or "richer"?
- If I create a moderating effect in SmartPLS, then I see the interaction effects. Does SmartPLS create them on his own or do I have to do it in SPSS or Excel?
--> I have read that it is very easy to create moderating effects in SmartPLS, so I don´t need to transfer the indicants into SPSS or Excel? Can SmartPLS do the whole analysis?
- How do I have to measure the interaction effects? I have read somewhere that it is usually that only the moderator, dependent, independent and the interaction term are used and that the other variables in my model which are not affected by the moderator are eliminated. Is that true? Do I have to measure it by oneself?
--> I think that it is OK to measure the interaction effects by oneself without the whole model

I am sorry that I have so many questions, but I have a lot of problems with the proceedings in SmartPLS.

Best regards,
Larissa
christian.nitzl
PLS Expert User
Posts: 248
Joined: Sat Jul 25, 2009 1:34 pm
Real name and title:

Post by christian.nitzl »

Hey Larissa,

There is no ‘creating moderating effect button’ in SmartPLS for a binary group comparison. Therefore you have to calculate it by yourself. As descript in the paper you have to split up your data analog your moderator variable (e.g. sex). Afterward you run the calculation for every model. Then you compare the path coefficients as descript in the paper.

Is your moderator variable a reflective measurement you can use ‘create moderating effect’ function in SmartPLS. Some introduction to do that you can find in SmartPLS (Help -> Welcome). This could be useful for your household variable.

Kind regards,

Christian
Schnuffel84
PLS Expert User
Posts: 31
Joined: Mon Jun 14, 2010 7:35 pm
Real name and title:

Post by Schnuffel84 »

Hey Christian,

I have finished my estimations with the interaction terms of my reflective numerical constructs and it worked very well...thank you :)
But there are still some questions remaining for the binary group comparison of the categorial variable.
I have to split the sex variable into male and female, right? I have done it in SPSS, but now the Indicator Set in SmartPls is read, so I can´t work with it. When I splitted the sex variable, the variable for male exists now of 1 and 0 in the dataset and the same for the female variable. Is it the right way I splitted up the variable? I am a little bit confused :)
Then you said here and in your article, that I have to run the calculation for every model, but my question is how to do it?!
Do I have to run the calculation with the whole model with all variables or with just the endogenous variable, the exogenous variable and the moderator variable?
Do I have to put the arrow form the moderator variable to the endogenuous variable (like in the other interaction calculation) or where do I have to put it?

Sorry, that I have so many questions. I would be very grateful if you could help me one again :)

Best regards,
Larissa
christian.nitzl
PLS Expert User
Posts: 248
Joined: Sat Jul 25, 2009 1:34 pm
Real name and title:

Post by christian.nitzl »

Hey Larissa,

No Problem and your are welcome!

The best way for the calculation of the moderation effect in this case is to build a separate data table for every parameter value e. g. one for female and one for male. But pay attion on that every table have the same variable structure (for ease). Thereafter you have two different datasets which you have to import to SmartPLS. That is easy to manage by right click on your project (left windows above) after you have open SmartPLS. Then you have to choose ‘import indicator data’. After you have imported the two new datasets you can choose every set by right click on one of the new dataset and click on ‘use data for calculation’. Now, you have to run the model and the bootstrapping for every dataset to get the necessary figures for the binary multigroup comparison. Last but not least you can insert these figures (path coefficient and variance) in the mentioned formulas in the working paper below on p. 44-46 for the path of interest.

Hope these explanations will help you!

Best regards,

Christian
Last edited by christian.nitzl on Fri Feb 03, 2012 8:57 am, edited 2 times in total.
Schnuffel84
PLS Expert User
Posts: 31
Joined: Mon Jun 14, 2010 7:35 pm
Real name and title:

Post by Schnuffel84 »

Hi Christian,

thank you very much for your answers and the new working paper, they are very helpful.
But I am still confused about how to create the models for the binary group comparison.
When I split up the dataset in two datasets in SPSS, for example I have cases in my dataset where stands 1 (for male or female) and 0 (if it isn´t male or female in the dataset), do I have to delete the cases where stands a 0 in SPSS? By hand or casewise deletion in SmartPLS? Sorry, if it is a stupid question :)
On page 44 of your new working paper I see a figure of the moderator effect of the variable sex. Is this the way I need to model the effect? Do I just have to draw the independent variable (male) and the dependent variable (customer loyalty)? Or do I have to draw the independent variable (in my example convenience), the dependent variable (customer loyalty) and the moderator variable (male) which has an arrow on the dependent variable? And then compare the both measurements?

Then if I want to measure t (on page 45), the size of my sample is how many for example men are in my sample?Where do I find the standard error in SmartPLS, in the bootstrapping calculation? How can I see in SmartPLS that the variances differ from each other in order to use the other t formula?
And then I noticed that the formula on page 50 of your new working paper differs from the one in your old.
-For effect size in the new paper : f2= R2(with interaction variable)-R2(without interaction variable)/1-R2(with interaction variable)
-in the old paper : f2=R2(with interaction variable)-R2(without interaction variable)/1-R2(without interaction variable)
Which one is now the right? :)

Thank you very much for your help. I really appreciate it!

Best regards,
Larissa
christian.nitzl
PLS Expert User
Posts: 248
Joined: Sat Jul 25, 2009 1:34 pm
Real name and title:

Post by christian.nitzl »

Hey Larissa,

you have to build for every group (e.g. male) a table. That mean, in one table/data file are only the information/parameter values (‘Ausprägungen’) for male and in a other separate table/data file for female. Afterward you have two separated data files. Therefore you need not longer the binary code 0 and 1 for your calculation.

Every data file you have to treat in SmartPLS as his own model. I would use for calculation all paths which you have in the base model. What is your research question? If you want to know how the different groups influence your moderator effect or would you know, how the different groups are influence your direct effects. Maybe you can show both. But in the moderator model with e. g. the product indicator approach only the path between product variable and target variable can be interpreted (p. 48). The example on page 44 is only to show in a simplified example how the binary group comparison works. For each of the separated model you can run the bootstrapping procedure and so you get the necessary path coefficients and standard deviation (bootstrapping -> path coefficients) for the difference test. In footnote 270 in the working paper there is an e-mail-address where you can get an example MS Excel for the t-difference test. It is also tested, if the variances are significant different (that is a simple F-test).

The formula for the effect size on page 50 in the new working paper is the right one. The often used formula of Chin et al. (2003) has a little bug. See the comment in footnote 295 on page 50. All in all there are often some nice hints in the footnotes especially for calculations. But the next time the author should be do these hints in the text above :-)

Kind regards,

Christian
Last edited by christian.nitzl on Fri Jul 23, 2010 1:35 pm, edited 1 time in total.
Schnuffel84
PLS Expert User
Posts: 31
Joined: Mon Jun 14, 2010 7:35 pm
Real name and title:

Post by Schnuffel84 »

Hey Christian,

thank you for your answer. But I still don´t understand how to handle the variable sex.
you have to build for every group (e.g. male) a table. That mean, in one table/data file are only the information/parameter values (‘Ausprägungen’) for male and in a other separate table/data file for female. Afterward you have two separated data files. Therefore you need not longer the binary code 0 and 1 for your calculation.
You mean that I have to create two datasets where in one there are all variables of my model and the male variable and in the other dataset there are also all variables of my model but now the female variable, is this right?
I have splitted up my dataset into two, but didn´t knwo how to handle the 0 in the data, so I deleted all the cases with a 0 by hand. Was this the right way? I don´t really think so :( Because now for the variable male or female just 1 remains in the dataset. What do you mean with I don´t need the binary code anymore? But what needs to stand in the variables?Which number?
Now I have tried to measure my model, but I had only the relevant variables left because I thought that I only need them for the measurement. But now in SmartPLS the following errors occur:
-if I use path weighting scheme : A singular matrix occurred during the estimation of the path coefficients using the path weighting scheme.Setting another weighting scheme could solve the problem.
- if I use the other weighting schemes : An error occurred during the outside estimation. Maybe there are too few estimations.
What does this mean? I am so sorry that I am too stupid to split up a dataset, but I really dont´ know how to handle it in SPSS?
If I have the male dataset, do I have to delete the female cases by hand? And the other way round?
I hope you understand my stupid questions.

Best regards,
Larissa
Schnuffel84
PLS Expert User
Posts: 31
Joined: Mon Jun 14, 2010 7:35 pm
Real name and title:

Post by Schnuffel84 »

It´s me again :)

Now I created a new data file with only female or male in it, but didn´t delete the cases where 0 stands and now the calculation works.
But you said before that I don´t need the binary code 0 and 1 anymore, but I left it in the data file. So what do I exactly do wrong? :)
Is it wrong not to delete cases? Or is it wrong to have still the binary code?
I am so confused :)
You have asked before what I want to analyse, but I didn´t answer it yet. My model is about convenience and how convenience effects customer loyalty, customer satisfaction and share of wallet. I measured the direct effects before, but now I want to know how the sex effects the effects between convenience and the factors for success. So do I need the base model or just the relevant variables for my calculation?
When I measured the interaction terms like involvement I just used the relevant variables and not the whole model.

Best regards,
Larissa
christian.nitzl
PLS Expert User
Posts: 248
Joined: Sat Jul 25, 2009 1:34 pm
Real name and title:

Post by christian.nitzl »

Hey Larissa,

Let us say in your columns are the different measured variables (the different questions) and in the rows are the information for cases (persons who you asked). Then in one group data file you have only the cases for e. g. man with all columns and in the other one you have all rows fore e. g. female. In other words, you have to delete the not necessary cases by yourself as you mentioned above. Therefore you do not need anymore 0 and 1 in your file. In one table you have only the ones and in the other one only the zeros. But you can leave it there if you want. It seems for me easier to manage such transformations in MS Excel than in SPSS. But that is only a matter of taste :-)

The cited errors typical occur if you have a sample which is to small for calculation in PLS. For every sample you should have at least 30 – 50 cases (p. 47). As far as I know, the PLS algorithm could work ex 10 - 20 cases. How many cases do you have for each sub sample?

I hope you understand my stupid answers :-)

Best regards,

Christian
Schnuffel84
PLS Expert User
Posts: 31
Joined: Mon Jun 14, 2010 7:35 pm
Real name and title:

Post by Schnuffel84 »

I still don´t get it :)

I have tow data files:
1)all variables and male
2)all variables and female

Now I see in the first data file that in the male column there are 0 and 1, is it right? And in the second data file it is the same.
Then I delete the not necessary cases (the 0) by hand and for the male data file 274 cases and for the female data set 224 cases are left. Is this the right proceeding, because the whole data file before had 498 cases...?
Therefore you do not need anymore 0 and 1 in your file. In one table you have only the ones and in the other one only the zeros. But you can leave it there if you want.
So when I deleted the 0 in each data file only 1 are left. Is it right? But what do you mean that I have in one table only 0 and in the other only 1? What needs to stand in the male/female columns? Not 1?

I am so sorry :)

Now I take all variables and the calcualtion error doesn´t occur anymore.

Best regards,
Larissa
christian.nitzl
PLS Expert User
Posts: 248
Joined: Sat Jul 25, 2009 1:34 pm
Real name and title:

Post by christian.nitzl »

Hey Larissa,

say 1 = male and 0 = female then after the right modulation analog this group variable/column it would be only 1 for the male data file and 0 for the female data file in tthe group variable/column.

As far as I understand that is right the way as you have do it now, isn’t it?

Best regards,

Christian
Schnuffel84
PLS Expert User
Posts: 31
Joined: Mon Jun 14, 2010 7:35 pm
Real name and title:

Post by Schnuffel84 »

Hey Christian,

so in the male data file with all variables there are only 1´s in the male column and in the female data file with all variables there are only 0´s in the female column? :)

So, it is OK, that the male data file exists of 274 and the female data file exists of 224 cases?

Thank you so much for your help. I am sorry that it has been so hard with me :)

Best regards and a nice evening,

Larissa
Post Reply