Replicability

Similarly, we assessed whether the group patterns observed in significant clusters identified in discovery analyses were observed in two relatively large-scale, independent datasets selected from (a) the EU-AIMS Longitudinal European Autism Project (LEAP), a large multi-site European initiative aimed at identifying biomarkers in autism [51, 52] and (b) the Gender Explorations of Neurogenetics and Development to Advance Autism Research (GENDAAR) dataset collected by the GENDAAR consortium and shared in the National Database for Autism Research [53]. For details on autism and NT inclusion and exclusion criteria for these samples, as well as our selection process, see Additional file 1: Supplementary Methods [52, 53]. The resulting EU-AIMS LEAP (N = 309) R-fMRI datasets comprised N = 133 males and N = 43 females with autism as well as N = 85 NT males, and N = 48 NT females (see Additional file 1: Table S1); resulting GENDAAR (N = 196) R-fMRI datasets comprised N = 43 males and N = 44 females with autism, as well as N = 56 NT males and N = 53 NT females (see Additional file 1: Table S2). For a comparison of demographic and clinical information between ABIDE, EU-AIMS LEAP and GENDAAR, see Table S3, S4 and S5 in the Additional file 1. After applying the same ComBat and pre-processing pipeline as used in the ABIDE-based discovery analyses, we extracted each of the R-fMRI metrics means from the Z-maps. As for robustness, we extracted values for each R-fMRI metric from the masks corresponding to the clusters showing statistically significant effects in discovery analyses and computed the corresponding effect sizes and residuals using the same methods described above.

For both robustness and replicability, discovery findings were determined to be robust and/or replicable (R+) based on two criteria: (1) the group mean difference(s) observed were in the same direction as those identified in the findings from discovery analyses [68] and (2) their effects were not negligible as defined by partial eta squared (i.e,.η_p² ≥ 0.01 any small, medium or large effects were R+) which is also consistent with prior work [41]. Finally, for consistency across analyses we also computed cluster-level effect sizes of the discovery findings using the same approach described above.