Fig. 4. Effect of global structural covariates on classification.
a Comparison of associations between global structural features and the first principal components determined from the 14 selected VBM-based (orange; used to order the x-axis) and the 11 selected FreeSurfer-based (blue) features (see also Supplementary Table 1,0). b Effect of residualization against global structural features on classification performance and classification performance obtained from global features only. Notably, AUC values obtained from analyses with permuted diagnoses showed mean values > 0.5, which was due to chance associations in the comparatively small datasets. Furthermore, surface based features showed an increase in performance after residualization against permuted global features. This suggests features with poor cross-site reproducibility were coincidentally prioritized for classification in the original data and this was remedied in the residualized data. The two sets of global features were identical except for the addition of either a median VBM- or FreeSurfer-based feature