[color=#000000]Hi Jeff,[/color]
[color=#000000]Not exactly, GlobalCorrelation is still a weighted metric, in particular it is simply computing the row-average of the voxel-to-voxel correlation matrix (ie. for each voxel, compute the connectvitiy/correlation between this voxel and every other voxel in the brain, the average of all those r values is the GlobalCorrelation value for this voxel). The ICC metric is just the same as the GlobalCorrelation metric, but it instead averages the r^2 values (instead of the actual -signed- r values)[/color]
[color=#000000]All of these voxel-to-voxel measures are simple metrics that describe [i]some aspect [/i]of the connectivity pattern between a voxel and the rest of the brain (and each metric focuses on slightly different aspects of these patterns). ICC focuses on the "strength" of those patterns (how large are those individual r-values, irrespective of sign), while GlobalCorrelation focuses on the "height" of those patterns (what is the average r-value in those patterns). If, for example, you find positive task-related differences in ICC at some region (e.g. higher ICC during task compared to rest), that means that the patterns of connectivity between that region and the rest of the brain are "stronger" (higher absolute values) during the task condition compared to the rest condition (this may indicate stronger positive correlations, stronger anticorrelations, or a combination of the two). You can then look at those actual patterns simply by using the resulting cluster/blob as a seed in standard seed-to-voxel analyses to further interpret what might be driving these "stronger" connectivity patterns during the task condition.[/color]
[color=#000000]Regarding the actual sign of the individual ICC values, if you are [i]not [/i]using normalized ICC measures (ie. unchecking the 'normalization' box in the first-level anaysis tab) then all of the ICC values will be positive (since those represent average r^2 values, which are always positive). When using normalized measures instead, then those average r^2 values are normalized to z-scores (with mean zero, variance 1 across the entire brain, separately for each subject), so in that case you may find positive as well as negative ICC values (and they simply represent above-average or below-average, respectively, original ICC values). [/color]
[color=#000000]Hope this helps[/color]
[color=#000000]Alfonso[/color]
[i]Originally posted by Jeff Browndyke:[/i][quote]Hi, Alfonso.
I noticed that CONN 16.b includes a few new voxel-wise metrics, one of which is Global Correlation. Is this an example of an unweighted ICC metric as mentioned in your prior posts? If it is, maybe it could address the question I have below.
I'm trying to get my head around how to interpret directionality of connectivity (increase/decrease) using the ICC metric results. I'm obtaining results in default-related networks that should be anti-correlated with task condition, but when I look at the eye (X) analysis bars for group x time (i.e., ICC effect sizes for each condition and group at each time point) some of our significant ICC blobs are positive and others are negative. If ICC is a weighted metric using absolute values, in which both positive and negative connectivity are incorporated, why am I getting negative ICC effect size bars? How does one drill down to see if the ICC blobs reflect task-positive or task-negative connectivity?
Thanks,
Jeff[/quote]