DCM for effective functional connectivity

Based on the time-series extraction in the VOIs, we subsequently applied DCM^25,69 to model effective functional connectivity between these VOIs. DCM tries to explain the observed brain responses in terms of underlying causal interactions between different areas at the neuronal level. DCM estimates the experimental modulation of (intrinsic) self-connections or (extrinsic) forward and backward connections between VOIs that are active during voice and speech processing in a directional manner. We created and estimated DCMs with the DCM12 toolbox (version 7479) as implemented in SPM12. The DCMs were based on five spherical VOIs in the left hemisphere and four VOIs in the right hemisphere, each centered on a peak located in the TVA and low-level AC (Table 1), with a radius of 3 mm to avoid possible overlap between VOIs. Separated models were generated in the right versus left hemisphere as the VOIs that were entered into the models were based on non-symmetrical peaks of activations across hemispheres. We therefore refrained from artificially creating symmetrical models for left and right auditory areas by selecting only symmetrical activation peaks in the left and right AC, as this would serve against our aim to test the functional hierarchy within the local micro-network of each hemisphere independently.

For each participant and brain hemisphere, we first created a full connectivity model (full model) with bidirectional connections between neighboring VOIs in each hemisphere (A matrix) (Fig. 2). Driving input (C matrix) to each node was specified by the experimental condition that elicited the original activity: The all sounds regressor provided input to HG, the voice sound regressor provided input to aST and pST, while the speech sound regressor provided input to mST (and additionally to left pSTS). The driving inputs were mean-centered, causing the parameters of the A matrix to represent the mean connection strengths across conditions. Finally, the modulation of intrinsic and extrinsic connections by experimental conditions (B matrix) followed the activation profile of the VOIs: Intrinsic connections in nodes were modulated by the activation profile of the node (e.g., intrinsic HG connectivity is modulated by all sounds trials), connections from ST/STS regions to and from HG were set to be modulated by the ST/STS activation profile (e.g., HG–aST connections could only be modulated by the voice sound trials), and forward connections from ST/STS regions to other ST/STS region were set to be modulated by the activation profile of the origin of the connections (e.g., connections originating from mST could only be modulated by speech sound trials). We estimated these full DCM models for each participant using Bayesian model inversion.

To estimate group-level parameters in these left and right AC networks, we conducted a second-level PEB analysis, separately for the left and right AC networks. The PEB analysis included a hierarchical model of the connectivity parameters, including connectivity parameters from all participants at the first-level (i.e., DCMs are fitted to each participants data, and posterior probability density over the parameters and the free energy taken to the group level) as well as a GLM modeling at the second-level (i.e., with one regressor per covariate per connection), and the estimation procedure used a variational scheme. After estimating the parameters of the full PEB model, we subsequently pruned away parameters using a BMR approach performed on the A matrix and B matrix parameters. The BMR performs an automatic (greedy) search over the model space to optimize model evidence. We here used the BMR as an exploratory approach with only minimal constraints and performed an automatic search over reduced PEB models. This search was accomplished with the simplifying assumption that these models were all equally likely a priori. The model evidence takes into account both model accuracy (how well the model fits the data) and model complexity (the difference between model parameters and their prior values). The BMR procedure was accomplished in an iterative process. Finally, we applied BMA across all models searched by the BMR, and averaging was performed as a weighted average of parameters across models according to the posterior probabilities of the models. Significant network parameters were determined with a posterior probability of p > 0.99.