PVCA message: boundary (singular) fit: see ?isSingular

Question

[Sebastian Hesse]

Running the pvcaBatchAssess from the pvca package I get the message: "boundary (singular) fit: see ?isSingular". Unfortunately I do not understand this message and also "?isSingular" doesnt help me much. Can you advice me on what to do?

The function runs through though and I get a graphical output as expected.

Thanks for your help! Sebastian

Answer 1

[Jean-Karim Heriche]

isSingular is a function from the lme4 package. This indicates that some effects are linear combinations of each other or that there's a variance of 0 somewhere. This can also happen when p>>n (more parameters than samples). This also means that the model is likely overfitted and/or suffers from numerical stability issues.

[Sebastian Hesse]

Is there something I could/should do about? What I noticed in this particular combination of samples (I am performing proteome analysis with different groups of patients agains a large healthy comparison group) is that "age" is a very strong variance influence. I suppose this has to do with the particular combination of patients and controls.

Using LIMMAs removeBatchEffect I corrected for the three main influencers (age, date-processed and sample_prep) and it went fine. These data I am using for visualisation and clustering analysis.

During LIMMAs diff ex analysis I block for the known batch effects and it also runs smooth with biologically sound results.

Is there anything to worry about or tale care of using this approach or should I simply not worry about it? Would it be better to change the healthy comparison group (one option would be to use family members of patients only instead of my large healthy cohort).

Any suggestions are most welcome, thanks!

[kanat.yermekbayev]

I wonder if I could ignore the warning message, if I have no interest in the differences in those "some linear combinations" which were specified as random effect in the model? Thanks

[Jean-Karim Heriche]

Usually one should investigate such warnings to understand if the output of the analysis is reliable.