When attempting to use select a subset of vertices to do surface-based decoding in an ROI, the screening_percentile parameter does not change the results at all.
I am using logistic regression with L2 penalty, in a nested cross-validation setup (i.e., I have 5 runs of data, 1 run is manually excluded, then the remaining 4 run is fed into decoder.fit() with leave-one-run-out CV).
The resulting accuracy in the inner CV fold (decoder.cv_scores_) and the outer CV fold (on the manually left-out run) did not change at all when the screening_percentile was set to 5 or 100.
Am I missing something?