Multi group analyses

[KKauwenberghs]

Dear all

I have a question about performing a multigroup anylses in SmartPLS3.

Context
We are conducting a research to validate a model with 8 independent variables and 1 dependent variable. We are looking which succesfactors are important in order for adult students to participate in their educational career and this from the perspective of the students. We have N=191 cases and at the moment or model, as shown below, will be validated.

My interest of research is to see if, depending on which motivational profile the students belong to, certain factors are more important for the students. For this I am going to perform a multigroup analyses where the students will be devided in 3 groups, basing an current research.

Question

My question is, because I don't know which students belong to which group: how I can divide the cases in to the different groups? Or how I can analyse how many groups to divide in? Because motivation is, at the moment, no part of the validaded model.

The research I use for the motivationprofiles, used SPSS and the hierarcical clusteranalyse and K-means analyses to divide the respondents in to their corresponding groups but literature says that:

"In an attempt to account for unobserved heterogeneity with regard to endogenous but also exogenous variables, researchers have routinely used cluster analysis techniques, such as k-means (Hair, Black, Babin, & Anderson, 2010; Sarstedt & Mooi, 2014) on the indicator data, or latent variable scores derived from a preceding analysis of the entire data set. The partition that this analysis produces is then used as input for group-specific PLS-SEM estimations. While easy to apply, such an approach is conceptually flawed because traditional clustering techniques ignore the path model relationships that researchers specified prior to the analysis. But it is exactly these relationships that are likely responsible for some of the group differences. Therefore, it is not surprising that prior research has shown that traditional clustering approaches perform very poorly in identifying group differences in PLS-SEM (e.g., Sarstedt & Ringle, 2010)."

Can someone help me how I can divide the respondents in to groups based on the questions of the questionnaire about motivation using SmartPLS3?

With kind regards,
Kurt Kauwenberghs

[jmbecker]

If you have no variable to classify your respondents into groups a-priori, your need to run FIMIX-PLS or PLS-POS (or potentially both: first FIMIX and then POS) to uncover unobserved groups.
However, these must not allign well to motivaions. They are groups of respondents with similar response behavior, but these must not match motivational profiles. In any case, it will be hard to justify that groups uncovered by segmentation (even k-means) are based on motivations, if you do not have motivations in your dataset (model).

[KKauwenberghs]

Dear Professor

I have two questions:

1. I have divided my respondents in to three groups, based on earlier research, using hierarical clusteranalyses and k-means analyses. Is the procedure for determining differences in the model, based on the different motivational profiles correct?

a. Run standard PLS algorithm of the model for all three groups separately and confirm that it complies with all requirements (i. e. do everything as described in chapter 1 to 6 of Hair et al. (2013).
b. Run the MGA-Analysis in smartPLS
c. Check for measurement invariance of the weights/loadings by verifying that the p-value of PLS-MGA is between > 0.05 and < 0.95. (in smartPLS Reports: Outer Loadings > PLS-MGA). If this is not the case, it means that the constructs may be interpreted differently by the groups. It is not a knock out-criterion but the following analyses should be handled with care (according Rigdon et al., 2010)
d. Compare Path Coefficients, Indirect Effects or Total Effects by checking that PLS-MGA has a p-value SMALLER .05 or LARGER .95 from the Final Results reports to draw your conclusion.

2. Is it possible looking for differences between respondents based on their motivational profile if motivation is not included in the model?

Kind regards