Comparison of 3 categories/conditions

[hcben53]

Hello everyone,

I´m in the act of conducting the analysis for my master thesis with smartPLS and I´m facing the following problem:

I collected data via performing an experiment with 96 subjects and randomly assigned them into 3 groups. Each group has the identical number of observations. The model is attached. Now I intend to compare the investigated condition (vinyl record) with 2 controlled conditions (CD & MP3). Is it possible to compare the investigated condition (vinyl record) with one control category by means of including a categorical predictor variable (the 1st exogenous IV) as a dummy coded variable? Therefore, I will include the data of i=32 accounting for the observations of the vinyl record condition (coded as 1) and the data of i=32 observations accounting for the observations of the CD condition (coded as 0 as the reference category). The same comparison will be applied for the MP3 condition (Vinyl=1, MP3=0).

Is that a suitable approach or do I have to conduct a multigroup analysis?

Kind regards,

Benjamin Beiersdorf

[jmbecker]

It is a valid approach. However, when you interpret your results you need to account for the fact that PLS routinely standardized all variables including your dummy. Hence, the interpretation will not be the difference between reference category and treatment category, but more similar to a effect coding scheme.

If you use a multigroup analysis you only account for interaction effects of your treatment onto the paths of the structural model, but not mean differences in your DVs.

It certainly depends on what is more interesting to your research question, but interactions could also be added to the first model.

[hcben53]

Thanks for your reply Dr. Becker.

You said it´s a valid approach but I don´t get your point regarding the interpretation of the difference between the reference category and the treatment category.
To make the purpose of my research more comprehensible: I´m investigating the value of vinyl for consumers. For that purpose I am comparing vinyl (treatment) with 2 substitutable music formats (CD & MP3) each seperately as a reference category. In order to test my hypothesis I need to ascertain whether vinyl has a stronger/weaker significant effect onto a construct with respect to the reference category. For example: The interaction with a vinyl record leads to a greater haptic stimulation compared to the interaction with a CD.

I am looking forward to your reply.

Kind regards,
Benjamin Beiersdorf

[jmbecker]

Significance testing and direction of effects will be the same as when using normal dummies. However, if you have a normal dummy, the estimated effect is the mean change in the dependent variable when switching from reference to treatment category. However, if you standardize the dummy, the effect has not such a straightforward interpretation.

[hcben53]

Dear Mr. Becker,

tanks for the information. I have another question regarding your last sentence. The PLS algorithm draws on standardized latent variable scores and uses them automatically as the estimations for the paths in the structural model. You said the interpretation of the Dummy variable will be very similiar to as a effect coding scheme. That makes sense - the latent variabel table of PLS vomits the values of 1 and -1. How does PLS generates unstandardized values for a more straightforward interpretation? IPMA modeling rescales latent variable scores in an unstandardized form. Could that be applied in my particular case?

Kind regards,

Benjamin

[jmbecker]

Yes, I think IPMA could be a solution to the rescaling problem regarding dummy coded variables.

[hcben53]

Thank you very much Dr. Becker! I wish you a merry Christmas season and a happy New Year!