Interaction effects

Frequently asked questions about PLS path modeling.
iris_afandiphd
PLS Senior User
Posts: 28
Joined: Thu Jun 24, 2010 5:38 am
Real name and title:

Post by iris_afandiphd » Sun Feb 06, 2011 1:43 pm

Hi Christian,

Ok for splitting into 2 groups : male (1) and female (2).
What about other indicators across LVs corresponded to male and female?

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

Post by christian.nitzl » Mon Feb 07, 2011 9:40 am

Hey,

Sorry, I am not sure if I understand your question.

Generally sex is a moderator variable. In other words there is no other variable in the model having an impact on it. Furthermore we only consider the differences in the path coefficients if we use sex as a group variable in SmartPLS.

Best regards,

Christian

iris_afandiphd
PLS Senior User
Posts: 28
Joined: Thu Jun 24, 2010 5:38 am
Real name and title:

Post by iris_afandiphd » Tue Feb 08, 2011 3:26 am

Hi Christian,

I mean that, It's OK for me to split the data into two groups, for example 49% (Male) and 51% (Female) according to GENDER column in the CSV data.

However, the rest of other variables will also divided by using 'filter' in MS Excel.

I understand that only GENDER should be in the model,and the test will be:

1. PLS for Male and other LVs
2. PLS for female and other LVs


Thanks

haslindar
PLS Junior User
Posts: 8
Joined: Sat Jan 30, 2010 3:40 am
Real name and title:

Post by haslindar » Wed Feb 09, 2011 11:08 am

Hi Christian,

I think I might have the same problem but the difference is my sample size is just enough to calculate the model.

Well, I have 2 groups (male and female) which consists of small sample about 104 and the sample for each group is 50 and 54. My question is:

1) since I have a small sample size, do I have to separate the two samples according to the groups as you mentioned in your explanation above and calculate for each model?
2) Can I combine those groups in calculating the model? If I combine those two groups, does it possible for me to put 0=male or 1=female for each of the cases then calculate the model?

Thanks.
Haslinda

haslindar
PLS Junior User
Posts: 8
Joined: Sat Jan 30, 2010 3:40 am
Real name and title:

Post by haslindar » Wed Feb 09, 2011 11:44 am

Hi it's me again.

In this discussion, you mentioned about
"In very complex models it could also be the case that your sample size is to small. Then it could be helpful to use another weighting scheme."

If I'm not mistaken (correct me if I'm wrong as I'm new in this PLS) the normal weighting scheme that we used is the path weighting scheme. But there are other 2 weighting scheme which are factor weighting scheme and centroid weighting scheme. What is the different between these 3 weighting schemes?

Also, did you mean that if I have a small sample size then I can choose other weighting scheme than path? How can I choose? Would there be a criteria on how to choose the weighting scheme?

I've tried to use the path weighting scheme and centroid weighting scheme but the message appeared "A singular matrix occurred......setting another weighting scheme could solve the problem". But if I used the factor weighting scheme then it works. How this happen?

Sorry I have a lot of questions. And millions of thank you in advance.

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

Post by christian.nitzl » Wed Feb 09, 2011 3:28 pm

Hey Haslinda,

Some explanations about weighting schemes you can find in following paper:

TENENHAUS, M.; VINZI, V. E.; CHATELIN, Y.; LAURO, C. (2005): PLS path modeling. Computational Statistics & Data Analysis, Vol. 48, pp. 159-205. https://studies2.hec.fr/jahia/webdav/si ... S_PM_5.pdf

Another weighting scheme as the standard one could be useful if the sample is relative small or a mulitcollinearity problem occure. But differences in path coefficient should not be high if you use a different weighting scheme. Therefore you could use all three in SmartPLS implemented weighting schemes.

For the group size about 50 it is possible to get first result that make sense. In such a case your differences have to be very high (at least 0.4-0.5) if you want to detected significant differences between groups. It also depends on the noise (variance) of your path coefficients.

In my opinion you could also use sex as a control variable. Put your sex variable as a latent variable in your model and connect this variable with all potential variable in your model. In this way you can say that the links in your model are free of the effect of sex.

I hope that helps!

Christian

haslindar
PLS Junior User
Posts: 8
Joined: Sat Jan 30, 2010 3:40 am
Real name and title:

Post by haslindar » Wed Feb 09, 2011 8:53 pm

Hie Christian,

Thanks for your quick reply and being so helpful.

Just another thing that I'm confused is when you suggest that :


"Put your sex variable as a latent variable in your model and connect this variable with all potential variable in your model. In this way you can say that the links in your model are free of the effect of sex"

So, if I put my sex variable as a latent variable and connect it with other potential variable, what score should I put for the sex? Is it 0=male and 1=female for every case? Then calculate the PLS algorithm as normal? Am i got this right?

Million thanks in advance. Sorry for asking many questions as I'm so confused.

Regards.

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

Post by christian.nitzl » Wed Feb 09, 2011 9:20 pm

Hey Haslinda,

You’re welcome!

You can use the dummy variable 0 and 1 for sex. This is your single indicator measurement. After you have implemented the control variable sex you can calculate your model as normal. So you got it right!

Best regards,

Christian

haslindar
PLS Junior User
Posts: 8
Joined: Sat Jan 30, 2010 3:40 am
Real name and title:

Post by haslindar » Wed Feb 09, 2011 10:15 pm

Hi Christian,

Thank you very much. Now I can calculate the model with the factor weighting scheme.

regards.

iris_afandiphd
PLS Senior User
Posts: 28
Joined: Thu Jun 24, 2010 5:38 am
Real name and title:

Post by iris_afandiphd » Fri Feb 11, 2011 5:05 pm

Hi Christian & All

I got error message after split the data into male (174) and female (90)

just filter-out male=1 and female=2 in excel
Help!

Image

Uploaded with ImageShack.us


"may be there are too few observ/ Latent variables Scores?
No problem with a complete / original data

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

Post by christian.nitzl » Sat Feb 12, 2011 1:24 pm

Did you try a different weighting scheme e.g. factor or centroid weight scheme?

Greetings,

Christian

iris_afandiphd
PLS Senior User
Posts: 28
Joined: Thu Jun 24, 2010 5:38 am
Real name and title:

Post by iris_afandiphd » Sat Feb 12, 2011 1:54 pm

Hi Christian,

No.

I'm using a standard scheme.

I try to validate

moderator: gender, age
for testing MALE, YOUNGER vs FEMALE, YOUNGER

Using a select case in original data is Ok now
Beside, I will also try another scheme.

Thanks

jmbecker
SmartPLS Developer
Posts: 1079
Joined: Tue Mar 28, 2006 11:09 am
Real name and title: Dr. Jan-Michael Becker

Post by jmbecker » Wed Feb 16, 2011 2:21 pm

Looking at your figure it seems that you still have gender as a variable in your model, although you split the dataset by gender. That would cause the latent variable gender to have zero variance (always the exact same value in both groups, i.e., 1 or 2) which causes problems during the estimation.

ruchi
PLS Expert User
Posts: 75
Joined: Wed Dec 22, 2010 3:09 pm
Real name and title:

interaction effect

Post by ruchi » Tue Apr 26, 2011 6:33 am

Hi Christian

I read all the your replies posted for interaction effect.
In my model i want to see the effect of my moderator variable (experience) on the relationship of my IV(reflective LV) and dependent variable (reflective).

The variable EXP(experience) was captured for 148 cases and i had categorized and coded it as 1,2,3,4

1 ( 0-5 years exp)
2 (6-10 years exp)
3 (11-15 years exp)
4 (greater than 15 years exp).

Now can i use this EXP variable (which is having either of the values-1/2/3/4 for my 148 cases) as my moderator variable using "create moderating effect" function in SmartPLS. I understand that we cannot use it for dummy variable (0/1 for male and female, I need to run two sepeate models and then compare).

Please let me know as i need to do moderating effects before submitting report

Thanks
Ruchi

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

Post by christian.nitzl » Wed Apr 27, 2011 3:55 pm

Dear Ruchi,
You can use the product indicator approach which means the ‘create moderating effect’ function only for metric (or ordinal) scales. Therefore the question have to be: is your EXP variable metric scaled. If yes than you can use a product term approach for the estimation of your moderating effect (moderating variable EXP is a single item). If not than you have to us a group comparison approach. There is some discussion about the usage of binary variable as moderating variable. But as far as I know there is no clear answer to this question. In my opinion you should use the group comparison approach by building two groups. E. g. you build a group with the cases of 1 for short terms and a group with the cases of 2+3+4 for long terms. How to build this two groups depends on the specific context of your model.

I hope that helps a little bit!
Christian

Post Reply