Another point that deserves emphasis concerns the different ways of dealing with confounding non-imaging variables (e.g., age and sex) when using neuroimaging data for MVPA classification. Although recent studies were unable to detect differences in predictive performance when comparing different approaches for dealing with confounds in MVPA studies (nor differences in the weights assigned by the models46,47), results from our confound control experiments suggest otherwise. We chose to add the covariates age, sex and data collection site ID directly as features to our model as our initial approach. The underlying principle of this approach is that all relevant variables should be included in the model and that their relative contribution to the final predictive model will be recovered during model training, without the need for manual confound adjustment procedures47. This approach resulted in high classification performance for medication classifications (>0.8 AUC). However, several sites included only medicated patients while in others no patients had received medication, which could suggest that this high performance was achieved through classifiers detecting site-effects directly from covariates (e.g., site ID) rather than brain data. The
