Hello, Thorhildur! Unfortunately, I have not worked with PLS directly. However, there are several others methods for variable selection. For instance, Successive Projections Algorithm (SPA) and Firefly Algorithm (FA) may be very helpful tools. If you are stuck with PLS, I suggest you to try using SPA or FA. On my profile there are some papers that describe them. Please, feel free to contact me any time if you need.

Best wishes and good luck with your research!

Hi Lauro - many thanks for your answer and good suggestions. I will have a look at those papers and algorithms, I have not heard of those before, but they could indeed prove helpful!

Best regards, Thorhildur

Dear Thorhildur,

I can propose you two papers dealing with the topic:

Statistical tests for comparison of quantitative and qualitative

models developed with near infrared spectral data (Journal of Molecular Structure 654 (2003) 253–262)

Combining near and mid infrared spectroscopy for heavy oil characterisation (Fuel 01/2014; 133:310–316).

Regards.

Ludovic.

Many thanks! I will take a look at them

Best regards,

Thorhildur

If you try to build a PLS model from a data set, the cross valiadation is the standard way to estimate the number of components.

Dear Weilu Lin,

Thank you for your answer! Could you elaborate on what you mean with cross validation? Is it EFA you are referring to?

Best regards,

Thorhildur

Comparing the model errors on cross validation and on prediction of an independent data set is the best way to compare them - i.e. do they predict well on new samples.

The problem is that I have a very small cross-country dataset, I can try prediction but not sure how valid it is for such low numbers. Thanks all for your suggestions!

Kærar þakkir Sveinn :)  Great help, excactly what I need.

Hi Thorhildur

I agree with Sveinn on GoF. There also other fit  and quality indices that can be very helpful. 

I suggest you go for WarpPLS software that calculates many indices for a given model. The user manual is also neatly written explaining meanings and ideal values for each indices.

read more about warpPLS here: http://www.scriptwarp.com/warppls/#User_Manual

Hope this is helpful.

-Saiyidi

Hi Thorhildur, there is a nice paper by Sharma and Kim on model selection in PLS ("Model Selection in Information Systems Research Using Partial Least Squares Based Structural Equation Modeling," Proceedings of the 2012 ICIS), which suggests that the Q^2 works reasonably well across most model constellations. Unfortunately, this does not hold for the GoF - see also our paper on the GoF in a related context: http://link.springer.com/article/10.1007/s00180-012-0317-1

Many thanks Marco, I will take a look at this paper!

Doing somewhat exploratory research I feel like I should be able to defend why I chose the model presented and not something else.

Best regards,

Thorhildur

Here you go!

Thank you Theo, i see I need a bit more depth to my understanding - look forward to reading!

:) to lecture or teach someone is a good and valuable thing -  I would not have enrolled in a PhD (relaviely) late in life if I did not want to learn -  so many thanks for all your good suggestions and help. Please not have your helpful nature and teaching skills removed!

Sounds promising with method, simple is always good. Will be in touch regarding data :)

Dear all,

I would support the views of theo here on Stone-Geisser. It is a good lecture from my perspective theo!

It was used a lot more in the 1980s for model comparisons and selections.

I would also warn you away from GoF measures in PLS at this time as Marko suggests.

There was a Journal of Academy of Marketing Science article "brand relationship quality or BRQ" in the title a couple of years back that conceptually rationalised whether the second order model should be preferred to a further analysis with just using the first order constructs. Check that out..

best regards,

Brad

Thank you very much for your answer Brad, I will definitely check the article! I report  R2 and Q2 in the paper. 

Perhaps I should create a new thread, but as I have gotten so many good responses here, I want to squeeze in one additional but related question:

If I compare two models, one of them with a larger number of explanatory variables than the other, then R2 would be higher in the latter if both models "make sense", right? Would the same happen to Q2 ?

The reason I ask is that I have a "two stage" model with four independent,  two mediating and one dependent variable. I want to compare the model with and without the mediating variables, i.e. structural vs. reduced versions of the model. As far as I understand, to be able to do that I have to capture the impact on the whole model (not just the predictive relevance of the independent variable as I know there is a recursive element I cannot really capture with this method) and I am not sure if either of those indicators are suitable for that.  If anyone understands where I am coming from then many thanks for all inputs :)

Best regards, Thorhildur

hello..

i understand where you are coming from. you should have a look at some of my papers using the Confirmatory Vanishing Tetrad model.

If the smaller model is nested within the larger model in your study then this might be the way to accomplish what you are talking about in a data driven way.

the SAS macro by John Hipp is great.

Hipp, J. R.; Bauer, D. J., and Bollen, K. A. Conducting Tetrad Tests of Model Fit and Contrasts for Tetrad-Nested Models: A New SAS Macro. Structural Equation Modeling. 2005; 12(1):76-93.

Hipp, J. R. and Bollen, K. A. Model Fit in Structural Equation Models with Censored, Ordinal, and Dichotomous Variables: Testing Vanishing Tetrads. Sociological Methodology. 2003; 33:267-305.

Bollen, K. A. and Ting, K. F. A Tetrad Test for Causal Indicators. Psychological Methods. 2000; 5(1):3-22.

c.f., the last example in the bollen and ting paper I think!

also my paper at researchgate might also help

Enjoy this new area

Brad

Excellent, many thanks Brad!

john hipp is a really nice guy and he could tell you who has used tetrad analyses this way.

i have.

Ken bollen will also respond. They are really generous with their support

I started following you. I see what comes out of it.

brad.