Thanks for your answer.

I will explain with more detail the points you ask me.

First, here are the function parameters I have used:

Using DiscriMiner:

`plsDA(X, Y, autosel = FALSE, comps = 3, cv = "LOO")`

Using MixOmics:

`plsda(X, Y, ncomp = 3, mode = "regression")`

`perf(plsda, validation = "LOO")`

I do not know whether discriMiner scales variables, the documentation of the package is poor. I can only link the **plsDA** function for more info…

According to X and Y loadings, I have calculated Pearson correlation for each component:

Although correlations are almost 1 (I don’t understand why component 1 and 3 have inverse correlation), Y loading values differ a lot, maybe because of scaling, but it makes a huge difference when plotting X and Y loadings in a PCA.

Here I present Y loadings:

Classification error rate is difficult to compare between the two packages, as yours presents a complete table with overall and BER error rates with different distances, and discriMiner only outputs an error rate number:

`With mixOmics:`

$overall |
max.dist |
centroids.dist |
mahalanobis.dist |

comp 1 |
0.7446809 |
0.7234043 |
0.7234043 |

comp 2 |
0.4680851 |
0.5319149 |
0.4468085 |

comp 3 |
0.4468085 |
0.4893617 |
0.4680851 |

$BER |
max.dist |
centroids.dist |
mahalanobis.dist |

comp 1 |
0.8000000 |
0.7585470 |
0.7585470 |

comp 2 |
0.5185897 |
0.5365385 |
0.4514957 |

comp 3 |
0.4651709 |
0.4950855 |
0.4792735 |

`With discriMiner:`

`error_rate = 0.1702128`

Lastly, if R2 and Q2 are not available, is there any estimate indicating the model prediction accuracy and predictive relevance of my variable I could use?

Thanks a lot for your help.

Adrián López