CFA and Interpretation Matters

[u-ing]

Hi,

I try to identify success factors by using PLS and right now, I just have a test data sample to experiment with, but smartPLS won't accept the import of this sample. (I've already discussed this in the "Bugs"-section of this forum)

Since I won't have much time, once the "real" data is generated and my knowledge about PLS is exclusively theoretic, I've got two questions about applying smartPLS.

First, I heard that the interpretation of the PLS Output would be a bit tricky. Are there any main problems about interpreting the influence from one latent variable to another or about interpreting the weights and loadings of the manifest variables? Or is it as "easy" as in linear regression, where you just have a "plus" or a "minus" and some values, that show you the strength of the influence? I think I still can't completely imagine how to interpret the weights practically...

Second, I want to test if I contributed all indicators to the right LV's in my model using a confirmatory factor analysis (CFA). Will that be possible with smartPLS?

I hope someone can help me here, since my deadline really isn't far away and I still don't have my data...:-/

Best regards

Sebastian

[jjsailors]

First, I heard that the interpretation of the PLS Output would be a bit tricky. Are there any main problems about interpreting the influence from one latent variable to another or about interpreting the weights and loadings of the manifest variables? Or is it as "easy" as in linear regression

Two comments. The first is that you should keep in mind that weights are what you want to look at to indicate how much each indicator contributes to each latent variable; the indicators cause the latent variables, not the other way around, and the loadings are useful to indicate how much each indicator has in common with the others in a block (high correlations among indicators is not a requirement nor even preferred in PLS). The second is that the path coefficients are exactly analogous to regression coefficients; by that I mean that you're dealing with prediction, not causality.

Second, I want to test if I contributed all indicators to the right LV's in my model using a confirmatory factor analysis (CFA). Will that be possible with smartPLS?

As PLS is analogous to a principal component analysis and not a true factor analysis, you are probably better off using covariance-based SEM (e.g. LISREL, EQS, AMOS) for your CFA. And, if a true factor model underlies your theoretic framework, you may be better off using covariance-based SEM for all of your anlaysis.

[u-ing]

Thanks for your reply John!

I thought about using LISREL or AMOS for the CFA in the first time as well.

But the problem with using LISREL for all of my analysis is that
1. I will use formative indicators and
2. my data sample will be really small
These two reasons made PLS my first choice.

Best regards

Sebastian

[jjsailors]

I thought about using LISREL or AMOS for the CFA in the first time as well.

But the problem with using LISREL for all of my analysis is that
1. I will use formative indicators

Conceptually I'm not sure what the point of CFA is if your model is using formative indicators. With formative indicators you are defining your constructs as weighted sums of your indicators rather then the indicators reflecting some common underlying construct. Since that is the case, by what criteria would you say that an indicator fits another block/construct better? With formative indicators it would seem to me that none of our conventions regarding CFA apply.

So given that and the small sample size (but primarily the formative indicators--difficult but not impossible to do with covariance-based SEM), I think you are correct to use PLS.

I would be interested in contrary opinions regarding CFA and PLS.

[u-ing]

Alright, yesterday - again - I read some papers with propositions and hints about applying SEM with formative indicators with PLS and I think my confusion slowly disappears, if not completely...

So, John, you would say that, due to the specific characteristics of formative indicators (they don't necessarily need to be correlated), it is not possible - or not necessary or useful - to use a CFA on my model? Is that right? Please tell me if something's wrong in my formulation.

Instead I might be better off to just compare the weights related to the indicators, to see, which indicators contribute how much to a construct??!! In addition to that I will have to test about multicollinearity to assess indicator reliability.

Best

Sebastian

[joerghenseler]

Correlation among indicators is by no means a criterium for evaluating formative measurement models. CFA does not make any sense, because its output does not imply anything, i.e. it does not tell you anything about the reliability and validity of your formative measurement.

You may want to check the following papers:
Diamantopoulos/Winklhofer (JMR, 2001) and
Krafft et al., in: Bliemel et al. (eds.): Handbuch PLS-Pfadmodellierung, Schaeffer-Poeschel 2005

[u-ing]

OK, I worked myself through another pile of papers (including those you proposed to me, Jörg) and I am indeed much smarter now.

Guess I'm prepared to get done with my analysis. All that's missing until now is the data...

Thanks for your help, John and Jörg!

Best wishes

Sebastian

[geddar8]

Maybe a stupid question, somewhat related to this area. If I want to predicts the effect of a manifest exogenous variable (reflecting an underlying latent variable) on a latent endogenous variable, do I just multiply all the way through (loadingXpath weight)?

Thanks,

Peter