Power and reproducibility in the external validation of brain-phenotype predictions.