Relationships between directed connectivity and WiN and SBS

After having computed spectrally resolved connectivity measures, we tested whether these explain inter-individual differences in SBS and speech in noise comprehension (WiN). We aimed to discover whether TD connections modulate WiN, SBS, or both. Since we are relating one set of multidimensional variables—connectivity—to another set of multiple variables—WiN and SBS—canonical correlation analysis (CCA) is a natural methodological choice (see for example Smith et al. 2015; Becker and Hervais-Adelman 2020).

For each of the 4 defined TD and bottom-up networks (i.e. the frontal-to-SBS-ROIs and vice versa for each hemisphere, plus primary-auditory-cortex-to-SBS-ROIs and vice versa), directed network connectivity was initially represented by a matrix of either 53 participants × 5 edges × 100 frequency steps (for the frontal-to-SBS-ROI and vice versa networks) or by a matrix of 53 × 2 × 100 for the primary-auditory-to-SBS-ROI (and vice versa) networks. Because of the high dimensionality of this data in comparison to the moderate sample size, we performed an initial dimensionality reduction. This was achieved using PCA of the spectrally resolved data for each node, retaining 3 PCs out of 100 frequency bins, explaining a minimum of 98% variance. These 3 components per node, per participant compose 1 dimension-reduced set of connectivity variables (now being 53 × 5 × 3 for the fronto-temporal model and 53 × 2 × 3 for the primary-auditory-to-SBS-ROI model) fed into the CCA. The other set of variables consisted of SBS scores (from the SBS-ROIs, extracted at syllable rate) and the scores from the WiN task, being normalized to zero-mean unit variance before performing CCA. The CCA, testing the relationship between these two sets of variables—i.e. the spectrally resolved connectivity on the one hand and SBS and WiN on the other—was performed per edge and hemisphere. SBS and WiN variables were normalized to zero-mean unit variance. For each run of the CCA, permutation testing (swapping the subject labels, n = 1,000) was executed to establish significance of the canonical mode linking connectivity and WiN/SBS. To ensure statistical robustness, P-values resulting from all tests here (n = 24, which include the 5 edges on each hemisphere tested for both TD and bottom-up effects of the network involving frontal and SBS-ROI areas, and the 2 edges on each hemisphere tested for TD and bottom-up effects in the network involving primary auditory cortex and the SBS-ROI areas) were subjected to a correction for multiple comparisons using the false discovery rate method (Benjamini and Hochberg 1995). After this correction, in case of a significant canonical mode—which is primarily a set of linear weights for both the connectivity set and the WiN/SBS set that obtain the best possible match between the two sets of variables—post-hoc linear correlation analyses are performed, between the CCA-weighted connectivity patterns and WiN and SBS variables separately, to further characterize and enable interpretation of the observed canonical mode.