negative outer weights

Questions about the implementation and application of the PLS-SEM method, that are not related to the usage of the SmartPLS software.
Post Reply
stefan8
PLS Junior User
Posts: 1
Joined: Mon May 28, 2007 1:17 pm
Real name and title:

negative outer weights

Post by stefan8 »

Hello,

in my PLS path model some weights in formative measurement models are negative.

In a few papers this is seen as a problem.

For example Tenenhaus, Esposito Vinzi, Chatelin, Lauro, “PLS path modeling”, page 165:
„There are no sign constraints on loadings and weights in the PLS algorithm, but unexpected signs of the loadings and/or weights show problems in the data and some actions must be taken.”

Can someone specify what these problems with the data could be? Is it that the data does not fit on the causal hypothesis for the manifest variable?

It is suggested to eliminate manifest variables with negative weights. Is this appropriate even if the negative weight is not significant?

Thanks and regards,

Stefan
User avatar
Diogenes
PLS Super-Expert
Posts: 899
Joined: Sat Oct 15, 2005 5:13 pm
Real name and title:
Location: São Paulo - BRAZIL
Contact:

Post by Diogenes »

Hi Stefan,
probably the problem of negative and nonsignificant weights is the multicollinearity between the MV.
The interpretation of these weights will be problematic too.

The suggestion of take these MV off, probably came from a regression way of thinking.
Another possibility will be group the correlated MV using principal component analysis.
Best regards,
Bido
soumya.ray
PLS Junior User
Posts: 8
Joined: Mon Feb 20, 2006 8:49 pm
Real name and title:

Post by soumya.ray »

Diogenes wrote:Hi Stefan,
probably the problem of negative and nonsignificant weights is the multicollinearity between the MV.
The interpretation of these weights will be problematic too.

The suggestion of take these MV off, probably came from a regression way of thinking.
Another possibility will be group the correlated MV using principal component analysis.
Best regards,
Bido
Thank you Dr. Bido for helping us clarify this issue. I too am facing the problem of negative weights. I am working on a model that contains two formative constructs which are both theoretically multi-dimensional.

My first formative construct is composed of many subfactors (with roughly 20 total MVs). To model this, I am using factor scores from a PCA (oblique rotation) of the subfactors. This gives me a single indicator for each subfactor, two of which have negative weights on the formative higher-order construct. To try and remedy this, I tried getting factor scores from a PCA (orthogonal rotation). The negative weights now become close to zero. I am not sure how to interpret this: the original PCA yielded components with eigenvalues > 1, but the PLS measurement model gives two of them insignificant weights.

For the second formative factor I am simply using all the indicators from the subfactors (a more manageable total of 6 MVs) as formative indicators (the repeated indicators approach). Once again, I have a negative weight.

I am not sure what more to do in either case: dropping indicators/sub-factors seems theoretically unsound because these formative indicators are supposed to compose the higher-order latent phenomenon. The only thing I can think of is to use all the indicators in a reflexive way. This attenuates the exogenous paths but I think thats more theoretically sound than dropping indicators and thereby losing meaning from the molar factors.

thoughts anyone? is there something we can do at the item-development stage to ensure we do not have this problem with formative factors?
christoph1
PLS Junior User
Posts: 3
Joined: Fri Mar 20, 2009 10:03 am
Real name and title:

(negative) outer weights & interpretation

Post by christoph1 »

Hallo!

As my question is related to yours, I post it hear:

I used a 7 point Likert-scale - formative model.

I also get some significant negative outer weights:

- in the first case (LV) this might be because of multicollinearity (condition index of the 14th dimension =33,5).
And this LV also has the majority of indicators (13). Maybe this is also the problem...

How would you interpret a significant negative weight in the case of no collinearity?
Maybe this is related to the fact that I do not fully understand the correct interpretation of significant positve outer weights of MV...
Is it: MV has a great importance/relevance (in explaining the variance) concering the associated LV?
Or is it also: the respondents agree that this is an important indicator concering the associated LV ?

Thanks a lot!
jc
Post Reply