Community Detection in Partial Correlation Network Models Online Appendix

This section contains an illustration of the Blockbuster algorithm on simulated data. We draw a single sample from the SBPCM with R = 1 common factors and k = 2 communities for T = 500 and n = 100. We draw a GSBM where each community has size n/k. The edge probabilities are set to ps = p = 0.5 for all s = 1, . . . , k and qvr = q = 0.01 for all v, r = 1, . . . , k, v 6= r. The network-dependence parameter φ is set to 20 while the network variance σ is 1. We draw [Θ]ii from a power law distribution f(x) = αxm/x α+1 for x ≥ xm with xm = 0.75 and α = 2.5. The edge-weights are drawn uniformly in the interval [0.3, 1]. We draw the data identically and independently from a multivariate Gaussian with covariance matrix given as in (9), where the factor loadings q are generated from a standard normal. The rst panel of Figure OA-1 displays a heatmap of the correlation matrix of the panel conditional on the factor, with the series ordered by the true community partition. The second and third panels show the corresponding sample correlation matrix when the series are randomly shu ed and when they are ordered by the estimated community