forum.smartpls.com

Have you ever tried to create a second-order formative index?

Right now, I am trying to do so for a construct "level of sports sponsorship".

However, I am facing several problems:
* I do not know which inner weighting scheme is most adequate (in this case, there are significant differences between the schemes, especially between path weighting scheme and the other two schemes).
* There are several indicators with negative weights, most probably because of multicollinearity. While Diamantopoulos/Winklhofer (2001) recommend to be aware of multicollinearity, they do not explicitly tell what to do under these circumstances.

Has anybody got tips for me?

When I create second order latent variables--in fact, for all latent variables, first order or second order--I typically set the outer model to be outer directed. Latent variables in PLS are always formative composities, regardless of how the outer model is specified (McDonald, 1996). As long as the indicator has a positive correlation to the dependent variable(s) the weight in this scheme will always be positive.

I do this to side-step the issues with multicollinearity among indicators, which I have found to almost always be a problem in applied marketing research (where I have most often used PLS).


McDonald, R.P. (1996). Path analysis with composite variables. Multivariate Behavioral Research, 31(2), 239-270.

Also, consider what the multicollinearity of the indicators is telling you: perhaps the construct should not be measured via formative indicators. It would seem more suited to a reflective, true score model (ala LISREL).

In this case, the only rationale I can see for using PLS (becuse, again, PLS latent variables are always defined by formative measures, no matter how we specify the model) is if you are in a situation where you need to be able to calculate the impact of the indicators on your dependent measure of interest.

In this case you must use PLS, in my mind, but need to have the latent variables to have outer directed indicators.

Let's look for example at a construct like service quality.

If you regard service quality as a formed attribute like Rossiter (2002, p. 314) does, you might want to create a second-order formative index.

Multicollinearity may arise from a halo effect, e.g. the overall satisfaction with the service.

Especially from a managerial perspective, one is interested in the impact of each first-order component (which may be a perception of a specific marketing instrument) on the overall construct and endogenous latent variables. Under these circumstances, a formative measurement model is preferable.


Rossiter (2002): See Forum "Literature"

My apologies, I wasn't following the model you were describing.

I assume you're measuring the second order latent variable with all of the indicators of the first order latent variables?

If so, then what I was suggesting should still work, I think. specify the measurement model for each latent variable to be outer directed. What you will find, I expect, is that the weighting scheme no longer has much impact and that the first order to second order structural coefficients--the second order constructs compent weights, if you will--are positive (assuming they theoretically should be!).

Hello, what do you mean by "set the outer model to be outer directed"? I have the same problem that the PATH WEIGHTING SCHEME leads to quite different results than the other both alternatives (centroid and factor weighting scheme). What could be the reason for this and how could I solve this?
Thank you very much.
Marco