Network analysis: edges and statistical significance
Hello,
I'm new to network analysis, and I would like to ask a doubt about edges and what they mean. From what I have understood reading articles and tutorials related to network analysis, the presence of an edge between two nodes means that a covariance exists between the two variables represented by those nodes. When we apply some regularization processes (e.g., EBICglasso), we select parameters to set the "level" of edge shrinkage, and edges are calculated like partial correlations.
Does it mean that these associations are statistically significant, as if they were considered as partial correlations with p<0.05? Or maybe the concept of statistical significance does not fully apply here?
Thanks,
Francesco
Comments
Hi Francesco,
I would guess that if you use something like EBICglasso, the cutoffs are based on EBIC, not on a p-value. But I'll ask our experts for clarification.
Cheers,
E.J.
Hi Francesco,
Indeed, the presence/absence of edges is then based on the model selection routine. The interpretation is a bit similar to significance, as the idea is the same (edges are only included if there is enough evidence to reject the null), but the routine is also very different. Mainly, you don't have a fixed false positive rate which you do have with significance testing (namely, alpha). See also https://bpspsychub.onlinelibrary.wiley.com/doi/abs/10.1111/bmsp.12173 for a description of this.
Best, Sacha