Figure 4. . Comparison of our coefficient-based method with Granger causality analysis. (A) Comparison of the parametric tests for GRu (blue curve) and GRc (cyan) with the nonparametric methods for GRc (green for GRcF and blue-green for GRcT; see the text for details) and MVAR (red for local test and gray for global). The x-axis indicates the strength of input correlations (i.e., pink noise) in the simulated network. The desired false-alarm rate is set to 2% as in Figure 3F and the number of observed time samples is T = 3,000. Error bars indicate one standard deviation over the 500 random networks as in Figure 3E. (B) Comparison of the miss rate improvement (decrease) with respect to parametric GRc for the 500 networks in A as a function of the number S of surrogates (x-axis). Red indicates the local test, gray the global test, green the full-network nonparametric GRc, and blue-green the target-only nonparametric GRc. (C) Details of the performance of the five methods in B as a function of the mean incoming weight per node (left) and the network density (right). The plots for the miss rate are similar to those for the ROC-based prediction power in Figure 2C. (D) Comparison of the computational cost for the surrogate-based method and parametric tests as a function of the number T of observed samples (left) and network size (right). Only GRcF is shown, as GRcT takes a much longer time in the unoptimized version that we use.