Hi Anna,

That is an excellent suggestion, and I see how that could be very helpful to investigate the role of anticorrelations. I am attaching a patch (this patch is for release 15e, simply copy this file to the conn distribution folder overwriting the file with the same name there) that allows you do that. Simply select "one-sided (negative)" in the adjacency-matrix threshold options to use "negative-sign" thresholds (e.g. when using .15 cost threshold units it will select the strongest 15% negatively-correlated connections; when using 0.25 correlation coefficient threshold it will select those connections with r<-0.25, etc.) 

Note that there are some very interesting subtleties. For example, when focusing on positive or two-sided correlations, typically brain connectivity networks show "small-world" properties (high global and local efficiency; behaving somewhere in between a lattice and a random network topology). When focusing on negative correlations, though, the global efficiency seems very high (almost as high as random networks) but local efficiency seems very low (lower than both lattices and random networks, which was somewhat unexpected, at least to me; I guess it probably relates to anticorrelated regions being mostly long-range connections?). 

Hope this helps
Alfonso
Originally posted by Anna Zilverstand: