PLS-DA with missing '' values predicted in Y

Analysis
1

Hello,

Firstly, thanks for developing the useful mixomics package!

I am writing to ask about two problems when I was trying to do the sPLSDA analysis. Basically I have two datasets, both of which I am using the first 1440 columns as X and the last column as Y. For the first one, I have some empty predicted values when Y should only have two classes; for the second one, the error message keeps popping up when I am trying to tune the model (in the tune.splsda).

I would really appreciate your help and suggestions in fixing this! Thanks in advance and take care.

Best,

Kathleen

2

Hi Kathleen,

thanks for sharing your data offline.

The main issue you are facing here is the number of zeroes. Basically your first data set is running on empty (you have 76% zero values). By chance during the cross-validation you may end up with a training data set with 0 values and so the PLS-DA struggles to predict anything.

For both cases, note that you have not removed the Y variable in your X data set so you are also overfitting (the PLS-DA will use that information twice to explain Y). This was clear in dataset 2 where the error rate went down to 0 in the second component (not the case once I remove the Y variable:
X2.test <- data2[-index2, -1441]
X2.train <- data2[index2,-1441]

The errors on the tune do not pop out on my end, consider updating the new version of the package, either from GitHub or on BioC (new release due in a few days).

Hope that helps,
Kim-Anh

2 Likes