Reproducibility and variability of default-mode networks from functional MRI: Comparison between random- and mixed-effect group statistics

Default-mode networks (DMNs) is a part of so-called resting-state networks associated with intrinsic neuronal activations of the human brain. DMNs represent distinct spatial patterns of neuronal activations within anterior cingulate cortex (ACC), medial superior, and middle frontal gyri (i.e., anterior DMN, or aDMN) and posterior cingulate cortex (PCC) along with precuneus (i.e., posterior DMN, or pDMN). In this study, reproducibility and potential variability of the aDMN and pDMN depending on the second-level analysis employing either a random-effect (RFX) that is solely based on inter-subject variability or mixed-effect (MFX) statistic that is based on both the inter- and intra-subject variability. Publicly available group fMRI data were analyzed using temporally-concatenated group independent component analysis (TC-GICA) and DMN-related independent components (ICs) in group-level were automatically selected. Dual-regression approach was adopted to calculate ICs in individual-level via least-square estimation from each subject's fMRI data using estimated group-level ICs as initial parameters. The characteristic traits of the DMNs depending on the adopted second group-level statistics were evaluated based on three performance measures including (1) percentage of overlap, (2) distance of center-of-masses, and (3) Pearson's cross correlation coefficient. The results indicated that the group-level spatial maps from the MFX statistic showed significantly greater level of reproducibility across the subgroups consisted of a part of all the subjects for all three performance measures than these from the RFX statistic (p<10 ^-10 from one-way ANOVA). This may possibly be due to inclusion of the intra-subject variability as a penalty term of neuronal activation. Moreover, for each of the two group statistics, a variability of the DMNs was region-specific, in which the pDMN was consistently showed lower level of variability than the aDMN across all the three performance measures (p<10 ^-10).