(a) Highly similar associations between behaviour and the brain can be found across a wide range of different measures derived from fMRI. We included a set of network matrices, spatial maps and amplitudes (node timeseries standard deviation) derived from several distinct group-average spatial parcellations/decompositions: from ICA decompositions at two scales of detail (dimensionalities of 25 and 200); a PROFUMO decomposition (PFM; dimensionality 50); an atlas-based hard parcellation (108 parcels [Yeo et al., 2011]); task contrast spatial maps (86 contrasts); and MSM warp fields from native space to MSMAll aligned data (from estimate_metric_distortion; https://github.com/ecr05/MSM_HOCR_macOSX/blob/master/src/MSM/estimate_metric_distortion.cc). Each row reports a separate CCA analysis, performed against behaviour/life-factors. A very similar mode of variation is found across most of the parcellation methods and different fMRI measures. rU-V is the strength of the canonical correlation between imaging and non-imaging measures (confidence intervals estimated using surrogate data), PU-V is the associated (family-wise error corrected) p-value estimated using permutation testing, taking into account family structure, and rU-V CI is the 2.5–97.5% confidence interval estimated using surrogate data. rU-Uica is the correlation of a CCA mode (subject weights) with the positive-negative mode of population covariation obtained from ICA200 partial network matrices as used in Smith et al. (2015), and is therefore defined to be one in the row containing the results from that CCA. The rU-Uica result was included because it shows whether different metrics are associated with similar or distinct behavioural modes of population covariation (one may expect different rfMRI measures to be associated with distinct aspects of behaviour). The final column contains the total number of CCA modes with PU-V <0.05 (results in other columns correspond to the most significant CCA mode, except for rU-Uica, which relates to the maximum correlation across all CCA modes). (b) The rU-V results here are inflated in comparison to the results presented in Supplementary file 1a (due to increased overfitting as a result of the parcellation only being available in 441 subjects compared with 819 subjects included for the other CCAs), but the associated PU-V can (to some extent) be used for comparison. Therefore, this Table compares PFM (d = 50), HCP_MMP1.0 (d = 360), and fractional surface area (the fraction of cortex occupied by each area in the multimodal HCP_MMP1.0 parcellation) on the same set of 441 subjects (only considering subjects with a complete set of 4800 resting state timepoints). (c) Results from simulated datasets in which one or more of the network matrices, amplitudes and spatial maps are fixed to the group average to remove any subject variability associated with it. Results in each row were driven by variables in which subject variability was present, as indicated with ✓ (variables with - were fixed to the group average). Results are shown for within-subject correlations between simulated and original z-transformed network matrices (Znetwork matrix), across-subject correlations between simulated and original subject correlation matrices (Rcorrelation), and for results obtained from the CCA against behaviour. Note that comparable CCA results from the original data can be found in Supplementary file 1a. This Table presents results from full correlation network matrices. (d) Results from simulated datasets in which one or more of the network matrices, amplitudes and spatial maps are fixed to the group average to remove any subject variability associated with it. Results in each row were driven by variables in which subject variability was present, as indicated with ✓ (variables with - were fixed to the group average). Results are shown for within-subject correlations between simulated and original z-transformed network matrices (Znetwork matrix), across-subject correlations between simulated and original subject correlation matrices (Rcorrelation), and for results obtained from the CCA against behaviour. This Table presents results from partial correlation network matrices. Note that the results flagged with * are poorly estimated as a result of the low rank of the PFM subject network matrices (containing 50 PFM modes) used to drive these simulations. The reason for this is that the PFM 50-dimensional subject network matrices were added into the data (to keep the simulation pipeline identical). This approximated 50-dimensional network matrix is too low rank to allow accurate estimation of partial connectivity across a much larger number of nodes. The full correlation results in Supplementary file 1c are estimable, and support the 25-dimensional ICA results. (e) Modulating the subject spatial maps by thresholding and binarizing retains the shape and size aspects, but removes any relative amplitude information from the spatial maps. Binarised % results are binarised after applying a percentile threshold, and therefore only retain shape aspects (while fixing the size). The results reveal that even after thresholding and binarizing the spatial maps, remaining spatial variability strongly drives the cross-subject information present in the resulting network matrices. See earlier Tables for a description of the measures.