Question about Nilearn's Decoders "screening_percentile" parameter

Hello everyone,

I am using nilearn’s Decoder to train a model and specifying the screening_percentile parameter to be e.g. 20. As far as I understand, this does ANOVA based feature selection before training the model.

However, when I plot the weights of the trained model, it results in a whole brain plot. So what I do not understand is: Shouldn’t there be less weights/voxel on the resulting weight plot, as some features should have been removed by feature selection?

When I use a model with L1 penalty, which internally basically also does feature selection but differently, the weight plot actually shows this. There are much less voxels/weights visible on it. This makes much more sense to me.

I would be very happy, if someone could explain to me why this is different with the ANOVA based feature selection.

Thank you in advance,