FIGURE 2.

Weighted Average Bootstrap Stability Index and Network Subdividing. Shown are a sequence of independent component analysis (ICA) models with model orders K–1, K, and K+1. A component in ICA_K (central purple circle, marked by star) is highly correlated with a spatial map in ICA_K–1 and subsequently subdivides into two spatial maps in ICA_K+1. Due to network subdividing, the bootstrap stability index I_q for this node is low. However, the nearly identical component in ICA_K–1 is not affected by network subdividing and is consequently highly stable. Edge weights are correlations between spatial maps r_K–1, r_K+1.1, and r_K+1.2. The Weighted Average Bootstrap Stability Index ($\overline{I_q}$) for the marked node is then an average of the bootstrap stability indices for this network, with weights proportional to the correlations (r_K–1, r_K, r_K+1.1, and r_K+1.2). $\overline{I_q}$ is thus better able to estimate the true bootstrap stability of this component, independent of the effects of network subdividing.