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 latter is supported by the finding that classifications using covariates only (age, sex, and site) also resulted in high AUC (>0.8 AUC), whereas regressing these covariates out from brain data resulted in lower performance, with only classifications for medicated vs. control classification remaining significant (supplementary Tables S9). Our control experiments also show that the FreeSurfer data itself is likely to be confounded by site-effects as well, as classifications using brain data only (without regressing out covariates) resulted in relatively high classification performance for medicated OCD vs. HC and medicated vs. unmedicated OCD classifications (with 0.70 and 0.75 AUC, respectively). This could be explained by classifiers being able to identify sites through specific sample characteristics (demography and inclusions criteria used) resulting in different brain anatomy, and methodical differences such as types of scanner and imaging protocols used. Interestingly, control experiments for main diagnosis classifications showed that these were relatively unaffected by different ways of dealing with confounds (supplementary Table S8). This
