Multicollinearity in Temporal Generalization Analysis in the case modif moving time window

đź’¬ Support & Discussions
1
  • MNE version: e.g. MNE 1.6.1
  • operating system: e.g. macOS 12

Hello everyone,

I am currently using MNE-Python version 1.6.1 on macOS 12 and have a question about enhancing temporal generalization analysis. I am looking to modify the cross-time validation decoding analysis by implementing a moving window of 20 ms, as opposed to the standard 1 ms. My reference for this task has been the “Motor Imagery Decoding with CSP” tutorial available here: Motor Imagery Decoding with CSP.

Previously, I successfully adapted this approach for temporal decoding where the output is a one-dimensional AUC. However, my current goal is to produce a generalization across time with a two-dimensional AUC matrix that includes [train time, test time]. I use Linear Discriminant Analysis for the classifier, but I’m facing a challenge with very strong multicollinearity along the diagonal of the results matrix.
I have been looping the classifier over time to generate the 2D matrix and done the analysis without the function GeneralizingEstimator since I need moving time window per 20 ms with window length 1 ms and with sfreq=1000. I have no idea how MNE can tackle this multicolinearity in their function or maybe my method is wrong.
How can I tackle this multicollinearity issue in this analysis?

Any insights or guidance would be greatly appreciated! I’m also happy to share my script if it helps in providing more targeted advice.

Best,
Risa

**

2

Hi,

I’ve never used GeneralizingEstimator so I don’t know how it works under the hood. However I have implemented a similar analyses myself. You have no control on the time window and step parameters when using it?

I’m not sure what you mean by that. Maybe that your results along neighboring time windows have similar AUC values? If so, given that you are sliding only by 1ms I wouldn’t be surprised.

In any case, could you share some code and maybe a figure? This will probably clarify things and allow more people to give you some feedback.

Cheers,

1 Like