All network simulations concerned a stationary network, i.e., assuming that mutual interactions between phenotypes have over time resulted in a stable variance-covariance matrix. Assuming m phenotypes, stationary network data were created according to the model:(4)where Σ denotes the m×m variance-covariance matrix between the phenotypes, I is a m×m identity matrix, and B is a full m×m matrix containing the regression parameters β of all the phenotypes on each other (e.g., element B[i,j] contains the regression parameter β of phenotype i on phenotype j). The diagonal of the matrix B was set to 0, implying absence of self-activation of the phenotypes (i.e., the phenotypes do not affect themselves). Ψ is a m×m diagonal matrix containing the variances of all phenotypes conditional on the effects of the other phenotypes. In all network simulations, the GV was only associated to the first phenotype in the network (Figure 1f). Note, however, that the GV effect spreads throughout the network as all phenotypes in the network were directly or indirectly interrelated. To assure convergence of our network models (i.e., simulation settings result in stable systems), we checked the sufficient condition that the largest eigenvalue of B*B t is smaller than 1 [33].
