For Gaussian data, variables j and k are independent given all the others exactly when their partial correlation −Kjk/√(KjjKkk) is zero, K being the inverse covariance matrix. The sample partial correlation r has an exact test: t = r√((n − p)/(1 − r²)) follows tn−p when there is no edge. Panel (a) shows the correlations (below the diagonal) and partial correlations (above); panel (b) the graph of the significant ones. Switch to marginal correlations to see how much denser the "correlation graph" is, and use simulated data to see how often the true graph is found.
Partial: r from the inverse of the sample covariance, df = n − p (exact, Theorem 18.L). Marginal: the ordinary correlation, df = n − 2. Bonferroni: reject when p < α/m, m = p(p − 1)/2 pairs; Benjamini-Hochberg: step-up adjusted p-values. A coris sample smaller than 462 is a random subsample (seed in the readout). Simulated data: N(0, K−1) with K = I − ρA, A the adjacency matrix of the true graph (dashed grey in panel b).
code/ch18_undirected_graphs.py (partial_correlations, pcor_tests, adjust, edge_exclusion, coris_continuous, sim_precision).