Jörg,
Because PLS is a regression based technique our models can be estimated using covariance or correlation matrices as input, but that does not mean we do not have exact latent variable scores that are--or could be--calculated.
As long as we have the raw data (regardless of whether we use it as input to our PLS estimation), and the weights that come from PLS, we are able to calculate exact latent variable scores. Always.
Just because we don't have to use raw data as input does not mean that we lose any of the properties of PLS estimation.
In contrast, true constructs--theorectical abstractions--are never exactly known. Moreover, true constructs always imply reflective measurement.
In PLS we never have reflective measurement, regardless of mode of estimation. And, therefore, in PLS we never are truly dealing with constructs or latent variables; we just (imprecisely) have borrowed that language. This is fine, but we should remember that we are not being precise.
I want to again state, because I know this can be a sensitive area, that this is not a criticism of PLS--in fact, being able to exactly calculate latent variable scores is often a huge postive aspect of PLS, for example, in applied marketing research.
Formative constructs with problematic indicator weights
-
- PLS Junior User
- Posts: 3
- Joined: Wed Jan 02, 2008 2:28 pm
- Real name and title:
Dear all,
I am not quite sure whether I got things right. I have the same problem (negative weights for formative constructs when measured in the "formative"-mode). I followed the advice to switch the direction of the arrows (i.e. to calculate formative constructs in the "reflective"-mode).
My question is, which values I should take for the formative constructs when calculated in the "reflective"-mode: Outer loadings (which equal outer model [weight or loadings] in case of the "reflective"-mode) or outer weights?
I did not understand the argument concerning the values of the latent variables as the Index Values for Latent Variables differ between the two modes of calculation.
Furthermore, is there any literature claiming that it is okay to calculate formative constucts in the "reflective"-mode?
Thank you very much.
Kind regards,
Volker
I am not quite sure whether I got things right. I have the same problem (negative weights for formative constructs when measured in the "formative"-mode). I followed the advice to switch the direction of the arrows (i.e. to calculate formative constructs in the "reflective"-mode).
My question is, which values I should take for the formative constructs when calculated in the "reflective"-mode: Outer loadings (which equal outer model [weight or loadings] in case of the "reflective"-mode) or outer weights?
I did not understand the argument concerning the values of the latent variables as the Index Values for Latent Variables differ between the two modes of calculation.
Furthermore, is there any literature claiming that it is okay to calculate formative constucts in the "reflective"-mode?
Thank you very much.
Kind regards,
Volker
Hi @all,
I just wanted to bring up Volker's question again since I am in the same situation right now.
If I have a "mode B" construct with negative weights and/or weights greater than 1 - what should I do? Would it be okay to use Mode A instead and report the weights? One result would be that the path coefficient in my inner modell change quote strong...
In my case all quality criteria for reflective constructs are more than fulfilled by my formative construct (which is truely formative by content), but if I use "mode B" I get problems due to multicollinearity (VIF about 3) that lead to negative weights...
Prof. Bido recommends to perform a PCA for the formative construct in case of multicollinearity. Has anybody experiences in this?
best regards,
Carsten
I just wanted to bring up Volker's question again since I am in the same situation right now.
If I have a "mode B" construct with negative weights and/or weights greater than 1 - what should I do? Would it be okay to use Mode A instead and report the weights? One result would be that the path coefficient in my inner modell change quote strong...
In my case all quality criteria for reflective constructs are more than fulfilled by my formative construct (which is truely formative by content), but if I use "mode B" I get problems due to multicollinearity (VIF about 3) that lead to negative weights...
Prof. Bido recommends to perform a PCA for the formative construct in case of multicollinearity. Has anybody experiences in this?
best regards,
Carsten
-
- PLS Junior User
- Posts: 10
- Joined: Fri Jun 24, 2011 3:01 pm
- Real name and title:
Hi,
I would like to rewake the discussion.
I also have negative weights for my formative indicators. I switched the direction of the arrows in the formative measurement model in SmartPLS as suggested above. However...
1. Not all negative weights disappear
2. The values of the latent constructs are quite different to the results before the change of the direction of the arrows in my formative measurment models. How come?
Does anyone have an idea / solution?
Best regards
I would like to rewake the discussion.
I also have negative weights for my formative indicators. I switched the direction of the arrows in the formative measurement model in SmartPLS as suggested above. However...
1. Not all negative weights disappear
2. The values of the latent constructs are quite different to the results before the change of the direction of the arrows in my formative measurment models. How come?
Does anyone have an idea / solution?
Best regards
-
- PLS Junior User
- Posts: 10
- Joined: Fri Jun 24, 2011 3:01 pm
- Real name and title: