Hi Alex,

This is a broad topic but generally speaking ROI-based analyses aggregate the BOLD timeseries/responses across all of the voxels within each region, so one would use ROI-based analyses when dealing with relatively homogeneous regions where we can safely make the assumption that different voxels or areas within this region are showing similar BOLD responses, and use voxel-based otherwise (when we do not want to make that assumption). Because connectivity analyses typically deal with two regions at a time (e.g. a "source" or seed region and a "target" region), we have to make that decision about these regions twice, first decide whether we want to aggregate across all voxels within our seed area/region, and then decide whether we want to aggregate across all voxels within our target area/region. If the answer is "yes" for both, we would use some form of ROI-to-ROI analysis (e.g. RRC), if the answer is "yes" for one and "no" for the other we would use some form of ROI-to-voxel analysis (e.g. SBC), and if the answer is no" for both we would use some form of voxel-to-voxel analysis (e.g. fc-MVPA). For more details about all these methods see https://web.conn-toolbox.org/fmri-method...

Now, as you mention, even if you choose SBC, meaning that you are confident that the seed area/ROI is relatively homogeneous but you are not sure or prefer not to commit regarding the target areas, you can still limit the analyses to only a subset of target voxels (e.g. using masking) and/or compute statistics that consider only results within those voxels only (e.g. using cluster-level statistics with small volume correction). Compared to RRC, where the connectivity between two regions is computed as the fisher-transformed correlation coefficient between the average BOLD timeseries in these two regions, an SBC+SVC approach would compute first the connectivity between a seed ROI and every voxel within the target ROI (e.g. as the fisher-transformed correlation coefficient between the average BOLD timeseries in the seed, and the BOLD timeseries in each target voxel separately), then compute voxel-level statistics in group/second- level analyses, and last compute cluster-level statistics characterizing properties of potential clusters of supra-threshold voxels or areas within the target ROI (with these cluster-level statistics corrected for multiple comparisons to take into account the size of the target ROI, as only voxels within that target area are considered). In general one would expect that, compared to an ROI-to-ROI approach, an SBC+SVC approach to give you additional flexibility to detect effects that may not be homogeneous across the target region, perhaps at the cost of reduced sensitivity to detect effects that are homogeneous across the target ROI, but both approaches are perfectly fine/correct (when used in the context of a single seed and target region; RRC approaches have some additional machinery built to deal with cases where you want to repeat these analyses across multiple pairs of seed and target ROIs and how to deal with additional multiple comparison corrections necessary in those cases). For more details about the methods used for cluster-level statistics see https://web.conn-toolbox.org/fmri-method...

Hope this helps

Alfonso

Originally posted by Alex G:

Hi Alfonso and CONN users, Can anyone explain what the difference in calculation of ROI-ROI analysis versus SBC analysis with SVC (via the Import Data) option is? My understanding is that ROI-ROI analysis calculates the average time-series in each ROI and generates a R-to-Z transformed correlation between those two averages. Is there a difference in calculation for SBC analysis restricted to an ROI, and is there a preference for what approach is considered optimal? Finally, do the standard corrections/thresholding for analyses differ? Thanks in advance, Alex