Fig. 6. CCAs reported in the population neuroimaging literature might often be unstable.
a Between-set correlations and the number of samples per features are extracted from the literature and overlaid on predictions from the generative model for various true between-set correlations rtrue. The reported between-set correlation can be predicted from the used number of samples per feature alone using linear regression (R2 = 0.83). We also estimated the weight error (encoded in the colorbar) for each reported CCA (details in Supplementary Fig. 16). b The distribution of estimated weight errors for each reported CCA is shown along the y-axis. For many studies weight errors could be quite large, suggesting that conclusions drawn from interpreting weights might not be robust. See Supplementary Fig. 16 for the procedure for estimation of weight errors.