Figure 2

Type I errors and power for simulated networks. The left panel shows the average number of false positives per replicate simulation (out of twenty simulated networks as in Figure 1) for different strategies for setting the significance threshold for both the variational Bayes method and the lasso methods. The right panel shows the corresponding power for all of the methods. The methods compared are as follows, vb^a and vb^b: variational Bayes methods without and with averaging in both directions of regression, and a posterior probability threshold of ${\widehat{p}}_j>0.99$, vb^c and vb^d: variational Bayes methods without and with averaging, and a posterior probability threshold of ${\widehat{p}}_j>0.5$, ${\ell }_1^{\mathsf{\text{a}}}$ and ${\ell }_1^{\mathsf{\text{b}}}$: randomized lasso with stability selection with the number of false positives bounded below 1 for the entire network, without and with averaging, ${\ell }_1^{\mathsf{\text{c}}}$ and ${\ell }_1^{\mathsf{\text{d}}}$: randomized lasso with stability selection with the number of false positives bounded below 1000 for the entire network, without and with averaging, and finally ${\ell }_1^{\mathsf{\text{e}}}$ and ${\ell }_1^{\mathsf{\text{f}}}$: regular lasso with the number of false positive bounded below 1 and 1000 for the entire network.