Question

OPLS-DA: values of pR2Y and pR2Q

0

Entering edit mode

Eddie • 0

@f2dcf917

Last seen 20 months ago

Belgium

Hi all, I'm running an OPLS-DA models. The Y is the bivariate response

```oplsda <- opls(dataMatrixEF, Y, predI = 1, scale="none", orthoI = NA, subset=c(1:100))

The issue I have is I don't get the p-value for R2Y and Q2Y in either my output or the diagnostic plot although I don't get an error message that model was not build because of predictive component was not significant. My question : not obtaining te p-values is it an indication of a not significant model? Thank you

100 samples x 228 variables and 1 response none scaling of predictors and standard scaling of response(s) R2X(cum) R2Y(cum) Q2(cum) RMSEE RMSEP pre ort Total 0.874 0.401 0.343 0.395 0.366 1 2 Warning message: 'permI' set to 0 because train/test partition is selected

```

Metabolomics • 2.1k views

ADD COMMENT • link 2.1 years ago • updated 22 months ago Eddie • 0

0

Entering edit mode

Hi, The ropls::opls modeling does not currently compute the pR2/pQ2 in case of a train/test partition is selected with the subset parameter, as is the case in your example (see the warning message: the number of permutations permI is automatically set to 0 in this case). You may obtain the pR2/pQ2 value by restricting your dataMatrixEF and Y vector to the 1:100 samples first, and then re-running the ropls::opls (keeping subset = NULL this time). Note that running (O)PLS(-DA) without centering is very rare and the model should be interpreted with caution. Best wishes, Etienne.

ADD REPLY • link 2.1 years ago etienne.thevenot ▴ 40

0

Entering edit mode

Thank you that was useful information. About the issue of centering you mentionned I meant to put a question about it but you have already brought it up. I have confusion about scaling of class response. I didn't understand how a class response can be scaled. And that's why I did the scaling outside the method then chose no scaling when applying the method, however the oupput mentions that the predictors are not scaled but response is. For example I fitted a model under UV scaling and worked but under pareto scaling the predictive component was not significant and I was afraid that this result is because of the scaling applied to the response. Below is part of the outcome. I would appreciate it if you could expalin the scaling of class response in the method. Thank you.

100 samples x 228 variables and 1 response pareto scaling of predictors and response(s)

ADD REPLY • link 2.0 years ago Eddie • 0

score 0 · Answer 1 · 2022-04-17

0

Entering edit mode

etienne.thevenot ▴ 40

@etiennethevenot-8285

Last seen 15 months ago

France

Hi, The (O)PLS(-DA) core method is a regression approach. In case of a classification, classes are converted to dummy (binary) variables. Unless you have specific reasons to do so, you can rely on the proposed scaling (e.g. pareto, or standardized = mean centering and unit variance scaling) without extra scaling outside the method. I would not expect that the choice of one of these two scaling approaches or the other would have a major impact on the significance of your model in most of the case when you work in with spectral data (i.e. acquired with the same instrument). Best, Etienne.

ADD COMMENT • link 2.0 years ago etienne.thevenot ▴ 40

0

Entering edit mode

Hi. Yes I agree with that for model signifiance but for features selection and biomarkers scaling has an impact. I get different results for top most important variables. My data has over 100 highly correlated variables.

ADD REPLY • link 22 months ago Eddie • 0