2.5.1. Temporal group ICA with dual regression

Group-level FNs were computed by means of a tGICA approach [Fig. 2(a)]⁴² by temporally concatenating participants’ datasets after time series normalization to zero mean and unit variance—producing a single-group dataset with dimensions [channels (46) × Hb chromophores (2)] × [time points (5000) × participants (99)]. The FastICA algorithm⁶⁶ was applied to the group dataset to extract 15 ICs. This number corresponds to the number of principal components explaining 60% of group data variance, which was established previously using principal component analysis (PCA). The number of ICs was determined using three criteria based on the consistency of the components across different initializations of the ICA algorithm, the anticorrelation between the Hb chromophores, as well as the percentage of data variance explained by each IC (see the Supplementary Material for additional details on ICA model order selection). The subject-specific spatial maps associated with each independent FN were obtained using a dual regression approach. This two-step method involves an initial spatial regression of the tGICA spatial maps to the individual fNIRS datasets to obtain the subject-specific time courses associated with each group-level IC. A linear model fit is computed between the estimated subject-specific time courses and fNIRS datasets to estimate the subject-specific spatial maps.

(a) Processing pipeline for tGICA and (b) connICA methods.

Statistical analyses were performed channelwise for each FN using a one-way random effects ANOVA with language background as a factor (i.e., BIL, SP, and BQ), resulting in 15 spatial maps of between group differences (i.e., channelwise F-test). Statistical tests were corrected for multiple comparisons at the channel level using the false discovery rate (FDR, q<0.05) method.⁶⁷ Bayesian hypothesis testing was also performed to estimate the relative likelihood of the data under the null and the alternative models.⁴⁵ To quantify the plausibility of the absence of an effect (i.e., evidence of absence), we repeated our group-level statistical analyses for each FN based on a Bayesian ANOVA using the R package BayesFactor.⁶⁸