Normalization models were built in training data only and these models were subsequently applied to adjust the test data. The same normalization strategy was applied for global structural parameters, which were subsequently used to remove the global structural signal from VBM- / FreeSurfer-based features. The resulting data was used for leave-site-out cross-validation analyses. For univariate analyses, as well as for machine learning analyses performed on the entire dataset, data were additionally corrected for a site factor, to account for the impact of site differences (see methods