Fig. 1: Training of the multiplex stochastic block models (SBMs).
A diagram of the data pipeline is shown in (A). We integrated participant data spanning1 resting state functional connectivity (FC) within the salience brain network and2 responses to the psychosis risk symptom assessment items. Pair-wise similarity distance measures were computed between participants to produce a weighted graph (network) with a neuroimaging layer and a (psychosis risk) symptom layer. A multiplex SBM was fit to the multi-layer network via a variational expectation-maximization (EM) algorithm. This process was conducted twice, once for the youth and again for the early adults. The selection process for the optimal number of blocks (Q) is shown in (B). We selected the optimal Q for each subject group (youth and early adults) using the Integrated Completed Likelihood (ICL) criterion. We evaluated values of Q ranging from one to seven and selected the models that converged with the highest ICL. Convergence was not reached when Q was greater than five for the youth and three for the early adults. We found Q : = 2 to be optimal for both the youth and early adults, producing ICLs of –231,381.11 and 40,699.73, respectively. A connection parameter estimates table for the multiplex SBMs is shown in (C). We report the mean and variance of the within- and between-block estimates for each layer in each age group’s model. We also report the log-likelihoods of the models.
