Fig. 2
From: The ‘un-shrunk’ partial correlation in Gaussian graphical models
Partial correlations obtained with different shrinkage values. This Bland–Altman plot compares partial correlations obtained from two datasets. a, b The axes represent the difference versus the average of the estimated partial correlation. c, d The axes represent the difference versus the average of the p-values. Datasets were simulated to encode the same associations differing only in their sample sizes (and thus their optimal shrinkage values). We proceed as follows: (i) simulate a network structure, (ii) simulate a dataset from this network, (ii) create a copy of the first dataset and concatenate them. The new concatenated dataset encodes the same association structure as the first, however it has double the sample size. The optimal shrinkages are 0.42 and 0.54. In grey: The ‘shrunk’ estimates. In red: The new ‘un-shrunk’ estimates. Unbiased/comparable estimates must be around zero. The new ‘un-shrunk’ method provides coefficients that are directly comparable with differences in the order of 10–07