How to extract latent factors in PLSR model

Question

Hi everyone,

I was going through the basic example of baltic .jmp (spectral data) given on the link "Example of Partial Least Squares".

after selecting number of factors in jmp and analyse it, several reports are generated. To my understanding NUMBER OF LATENT FACTORS are related to those variables (out of 27 in this example) which explain most of variance. If that is true then, how to identify positions/names of those variables (from v1 to v27 in this case).

Regards,

Khan.

marshadmit0 · Answer

Thanks Ian and Peter for the advice. first of all PLSR has been widely used in variable selection in literature. For the moment it is worth trying what Ian suggested, going with VIP and Coefficient plots which make some sense too.

Peter_Bartell · Answer

Just to add a bit to Ian's wise advice...my experience from a practical point of view is to, if I can, follow up any PLS related analysis where the focus is on variable identification, with designed experiments to bring more credence to the cause/effect relationships you may be seeking.

ian_jmp · Answer

Given that PLS is a multivariate approach that relies ultimately on correlations between variables, there is some debate about how, or even if, it should be used for variable selection (pricking an 'active subset' of v1 to v27 in this case). In PLS each X variable plays a dual role (for both dimensionality reduction and regression), and the 'VIP vs Coefficients' plots show this. So if you do want to try what you suggest, this might be a reasonable place to start.

There are many more details here, or in other books on PLS.