Hi Yifei,

Some thoughts on your questions below
Best
Alfonso
Originally posted by Yifei Zhang:
Option 2.a is the correct one if you want to compare between the two groups the association between A and the small-world measures (the [0 0 1 -1 0 0] contrast compares the regression coefficients between the patient group and the control group). Positive values will indicate greater association in patients compared to controls, and negative values will indicate greater association in controls.


p-FDR (seed-level correction) controls for false discovery rates at the level of each seed ROI (separately for each seed ROI, controlling only for multiple 'target' ROIs), while p-FDR (analysis-level correction) controls for false discovery rates at the level of the entire analysis (controlling for multiple 'source/seed' and 'target' ROIs). If you want to make inferences regarding individual connections (between pairs of ROIs) and you do not have specific source/seed a priori hypotheses, then you would need to use the p-FDR (analysis-level correction). Of course this sort of inferences are the hardest to obtain since you do not have any a priori way to limit the search space (e.g. select one or a few source ROIs of interest only). For this sort of unconstrained hypothesis (looking at the entire ROI-to-ROI connectome) it is usually more reasonable to try to obtain first less spatially-specific inferences. For example, you may make inferences regarding individual ROIs by selecting any seed-level corrected threshold (either F-test or NBS seed-level methods) combined with and connection-level threshold (not necessarily corrected). In this case you may conclude for example that some specific ROI(s) show altered connectivity, even if you may not assert exactly the specific ROI-to-ROI connections that show this effect. Last, you may also make inferences regarding individual (sub)networks of ROIs, by selecting any network-level corrected threshold (and again any connection-level threshold, not necessarily corrected). In this case  you may conclude for example that some subnetwork of ROIs show altered connectivity, even if you may not assert exactly which or how many ROIs are exactly in this network nor the specific ROI-to-ROI connections that show this effect. Even if these sort of ROI-level or network-level inferences are less spatially-specific than the original connection-level inferences, they are also considerably more sensitive (easier to find with reasonable/limited number of subjects), so they can still give you important information regarding your research questions and serve as a good starting point for reducing your search space and generate more spatially-specific hypotheses for future analyses (e.g. about individual connections).

Let me know if this clarifies.