Model comparison

We used Bayesian model selection to compare the generative models (Penny et al., 2004) and to determine which model best explained the neural responses for each group (Kiebel et al., 2008; Stephan et al., 2009). We used random effects analysis to accommodate group heterogeneity (Stephan et al., 2010) with adjustment of the model fit for model complexity to reduce over-fitting (Kiebel et al., 2008). The free energy estimates of –log(model_evidence) were used to compare models and estimate the model exceedance probability (xp; that a single model was more likely to have generated the data compared with all other models). We report all models with an xp of >0.05.

Next, we compared eight model families using random effects Bayesian model averaging (Penny et al., 2010). Each model family contained four models differing in the presence or absence of top-down inputs to the system and cross-hemispheric connections, but with the same pattern of connectivity between STG, IFG, and IPC nodes. Specifically, referring to Figure 1, the fully connected family included models 8, 16, 24, and 32, the fully disconnected family models 1, 9, 17, and 25, with the other families being integer increments in between.