problem with BETA model coefficients using PLSREGRESS

I am using the PLSREGRESS function to build a model to predict the Y response values from a future X matrix of predictors. I am using the syntax: [XL,yl,XS,YS,beta,PCTVAR,MSE,stats] = plsregress(X_Data,Y_Data,50,'cv',....); with a variety of cross validation methods such as random 10-fold, venetian blinds and contiguous blocks etc. The problem I have is that while the Mean Squared error changes with each different cross validation method the beta coefficients never change and are in fact identical to the unvalidated model coefficients. Thus I cannot use my PLSR model to predict Y responses for a new X matrix of predictors. Has anyone come across this before? and if so how did they resolve it. Thanks

[Tom Lane]

You are right, the 'cv' option is intended to estimate the MSE by cross-validation. It doesn't affect the coefficient estimates. But I would think that would make your problem easier. You can use this single set of coefficients, estimated on all the data, to make predictions. What would you do if the coefficients changed?

[ardeshir moinian]

Hi,

I was also looking into using CV to improve PLS model predictions. However, the CV predictions don't change at all when I change the CV type. Which means that I don't even need to perform CV. This is very odd!