Check the variance inflations of your IVs, it may be high. If so apply a factor analysis.

Thank you for the promt response! I am afraid, I need a more explicit information -

1. What variance infllation (VI) has to do with VIP? (mathematically it does not seem to be related).

2. How to check the VI setting in NIPALS?

3. How the factor analysis can substitute for PLS2?

Thanks again for your help!

You can first apply a multiple regression with enter option and check VIF (variance inflation factor). it shouldn't be more than 10. VIF rises when your independent variables are correlated so in my opinion the PLS algorithm confused to whether choose which IV for variance loadings.

Even if your VIF was at good condition you can Analyze them again using factor analysis.

It doesn't substitute the PLS but it tells you the importance of your IVs in another way.

I think this is not what I am asking about. The question is not about how to analyze my dataset but how the PLS with NIPALS algorithm in STATISTICA calculates VIP values and what would be a cutoff for these VIPs. If you are a user of this program, just try to run PLS with any artificial dataset in which Y is calculated as a function of X1 and X2 and see that VIPs for these X1 and X2 are never >0.8 (mind the regression is perfect and all R2 and Q2 are 1).

Again - this suppose to be a standard procedure with many publications saying that they used VIP=0.8 as a cutoff value in STATISTICA...

This is a result of PLS procedure in SAS 9.3 with 15 factors and three different analysis. In first, i entered all 15 factors, in 2nd, i enter 2 factors with least loading based on first analysis, in 3rd, i enter factors with highest loadings on opposite.

The method is same NIPALS.

You see for both 2 factor datasets VIP is different and acceptable for some.

Exactly... I have no problem getting this done in SAS, but the same dataset would give totally different Variable Importance values in STATISTICA. I would like to understand what is going on there, because the equation that StatSoft gives for VIP calculation is the same as everywhere else and people using STATISTICA claim that they get VIPs that are amenable for evaluation with Wold criteria...

I have no idea about whats happening in STATISTICA. Cos i'm satisfied with vast capabilities of SAS in analyzing data.

Thanks, but the question was specifically about this analysis in Statistica:)

So with the problem you've faced, i think you'd better to check other results of the software Which are even met the expectation.

You mean loadings, regression coefficients, R2 and Q2 values? Yes, I use them, of course, but using a cutoff for VIPs would be more straightforward and conclusive. The soft is designed to calculate these values and people seem to use this software all the time with Wold's threshholds (there are papers, figures...), which means that I am missing something...

I asked this question at StatSoft forum, but it is awfully quiet there...

Yes, I meant them. Check them using another software e.g. SAS. I hope they won't be erroneous too.

Hi,

You may want to check out the recent publication on this subject

http://www.researchgate.net/publication/264936800_Interpretation_of_Variable_Importance_in_Partial_Least_Squares_with_Significance_multivariate_correlation_(SMC)

Success !

Thanh Tran

Article Interpretation of variable importance in Partial Least Squar...

I did but I have not understood how this relates to the question.

Thanks for the paper, anyway.

/E

Check with Elena, check with Statsoft in Sweden directly. I have found bugs in older versions of Statistica