Hi Jennifer,

That's a very interesting question. Both approaches are perfectly valid, and their difference is generally rather subtle. In the first approach CONN is computing the average of the Fisher-transformed correlation values characterizing the correlation between the seed A and each of the voxels within the cluster B, while in the second approach CONN is computing the (single) Fisher-transformed correlation value characterizing the correlation between the seed A and the cluster B. When computing correlations between seeds that contain multiple voxels, CONN averages the BOLD timeseries across all those individual voxels and uses the resulting average BOLD timeseries for computing connectivity (e.g. correlation values), so effectively the difference between the two approaches arises from where exactly are you effectively averaging across all voxels within a seed: in the first approach you first average the BOLD timeseries across all voxels in A, then compute the correlation coefficient between the resulting (average) timeseries and the BOLD timeseries at each individual voxel within B, and last you average the resulting (Fisher-transformed) correlation coefficients; while in the second approach you first average the BOLD timeseries across all voxels in A, then average the BOLD timeseries across all voxels in B, and only then you compute the correlation coefficient between the resulting two (average) timeseries. One could even imagine a third approach, similarly characterizing the connectivity between A and B, where one first computes all correlation coefficients between each voxel within A and each voxel within B, and only then average the resulting Fisher-transformed correlation coefficients (e.g. in CONN you can compute those average connectivity values using the function conn_vv2rr). In any case, coming back to the interpretation of these different approaches, if the regions are relatively small and homogeneous, the differences between these two approaches should be extremely minor, as individual voxels will show very similar BOLD timeseries and averaging will produce the same effect whether at the timeseries or at the connectivity level. If you are finding relatively large differences, that probably indicates that the voxels within cluster B are non-homogeneous (as both approaches are averaging timeseries across all voxels in A), so that would suggest that perhaps you would be better served by dividing that cluster into smaller homogeneous sub-regions in order to better characterize the potentially different connectivity between A and each of those subregions in B. 

(if, on the other hand, the results are very similar and you just want to choose one of these two approaches I would typically suggest the first approach as that is much simpler to compute -and perhaps also more closely directed to the nature of your original SBC analyses, as these are based on connectivity between a seed area and individual target voxels-)

Hope this helps

Alfonso

Originally posted by Jennifer Siegel :

I am running a connectivity analysis looking at the relationship between a behavioral change value and connectivity change from a seed (pre vs. post). 

Subject effects: 

ALL SUBJECTS

BEHAVIORAL CHANGE

[0 1]

Conditions: 

Pre 

Post

[-1, 1]

The data that I get out when a) run second-level analysis and then import the data to get the Fisher's transformed values from this analysis is DIFFERENT to b) when I output mask of significant cluster, rerun a second level between seed and significant cluster (in this case a mask) ROI-to-ROI analysis and import data to get Fisher's transformed values  

1) How do I interpret these differences? 

2) Which one gives me the pre and post connectivity between the target and significant cluster so that I can then create a scatter graph of these relationships?  I assume it would be b above?  

I should note that when I subtract out the pre and post values from a and b - I get exactly the same output.  Why are the fisher's transformed values slightly different?