Dear Alfonso,
During an ICC analysis, I stumbled on negative values, and at first I was puzzled, because as I read the conn manual, I expected ICC to be the "root mean square of the correlation coefficient values", so in other words I here got a root value of a necessarily positive value that was negative!
By digging on the original article by (Martuzzi et al. 2011), it's specified that "For statistical purposes, the values in the map are normalized to fit a Gaussian distribution with zero mean and unitary variance by subtracting the ICC-dth obtained at each voxel by the average value across all the voxels and dividing this by the standard deviation of the whole-brain map". I guess you implemented ICC in CONN in a similar fashion.
If that is true, the meaning of ICC is not simply "how much any voxel is central", but rather "which voxels are more central than the average voxel centrality".
If this is correct, I would suggest to rewrite the description of ICC in the documentation, here is my suggestion:
"ICC is a measure of network centrality at each voxel compared to the average. It characterizes the strength of the connectivity pattern between each voxel and the rest of the brain, then standardize to 0 mean by subtracting the average ICC and to a Z-score of variability 1 by dividing by standard deviation of whole brain (standardized root mean square of the correlation coefficient values)."
Thank you very much for confirming if this is the correct explanation for the negative values :-)
Best regards!
Stephen
During an ICC analysis, I stumbled on negative values, and at first I was puzzled, because as I read the conn manual, I expected ICC to be the "root mean square of the correlation coefficient values", so in other words I here got a root value of a necessarily positive value that was negative!
By digging on the original article by (Martuzzi et al. 2011), it's specified that "For statistical purposes, the values in the map are normalized to fit a Gaussian distribution with zero mean and unitary variance by subtracting the ICC-dth obtained at each voxel by the average value across all the voxels and dividing this by the standard deviation of the whole-brain map". I guess you implemented ICC in CONN in a similar fashion.
If that is true, the meaning of ICC is not simply "how much any voxel is central", but rather "which voxels are more central than the average voxel centrality".
If this is correct, I would suggest to rewrite the description of ICC in the documentation, here is my suggestion:
"ICC is a measure of network centrality at each voxel compared to the average. It characterizes the strength of the connectivity pattern between each voxel and the rest of the brain, then standardize to 0 mean by subtracting the average ICC and to a Z-score of variability 1 by dividing by standard deviation of whole brain (standardized root mean square of the correlation coefficient values)."
Thank you very much for confirming if this is the correct explanation for the negative values :-)
Best regards!
Stephen