Fig. 1. Multivariate brain-behavior associations analysis pipeline.

a ABCD was the discovery set and Generation R as the external validation set. The discovery set was divided into training and test sets 30 times, resulting in 30 train-test pairs in ABCD. The eigenvectors of PCA from the ABCD_Training set were applied to ABCD_Test set to calculate the principal components, then the canonical loadings obtained from the ABCD_Training set were projected to ABCD_Test set to compute the out-of-sample generalizability. Similarly, weight vectors of SCCA from the ABCD_Training set were then directly applied to Generation R to assess the out-of-study generalizability of the model. We also implemented the qualitative replication approach, in which we train the CCA model independently in Generation R and compare the results across the two cohorts. Note that the sample size in ABCD is an example from one train-test split. b Out-of-sample generalizability within ABCD and out-of-study generalizability in Generation R.