Hi Maxi,

First, just to make sure I am understanding your setup correctly, when you say "during preprocessing I partitioned the band-pass filtered time series..." I imagine you probably mean "during first-level analyses" (i.e. by setting the number of frequency bands for that particular ROI to 3)? In CONN, in addition to a general band-pass filter defined in Preprocessing which applies to all conditions and all ROIs equally, you can also define a condition-specific band-pass filter in Setup.Conditions (in the time-frequency decomposition field) and that would apply to all ROIs equally but only for that condition, or define a ROI-specific band-pass filter in First-level analyses (in the number-of-frequency-bands field) and that would apply to all conditions but only for that specific ROI. I am going to assume here that you used the latter approach, but please let me know otherwise. 

Regarding (1), yes, graph-measures are notoriously sensitive to the choice of nodes/ROIs and typically you want to avoid having redundant/overlapping ROIs, so I would probably also recommend removing the default left frontal pole ROI (either in your main analyses or at least as an additional analysis to check the possible influence of this factor in your results). That said, I do not see any reason to expect that the inclussion/removal of that default ROI would affect your graph measures in a group-specific manner (you can expect differences in the graph measures across all subjects when you include/remove that additional node, but not interactions with your subject groups), so I doubt this is acting as a confounding effect in your results. 

Regarding (2), the effect here is a bit more subtle. First (and assuming you used the 'First-level analysis' filtering mentioned above; this would not happen if using the alternative 'Setup.Conditions' approach) the correlations between your frequency-specific VBM ROI signals and the rest of ROIs are generally going to be lower than the correlations between the same ROIs if using the same frequencies across all ROIs (instead of a maximum value of 1 your correlations are only going to go as high as the percentage spectral power within each frequency band; this is the effect that Noah was mentioning). This means that, generally, the global efficiency measures for your frequency-specific VBM ROIs are going to be lower than the global efficiency measures for your "entire-frequency-band" VMB ROI.  Second, different frequency bands are by definition uncorrelated so the VBM ROI signal acoss the three bands that you are exploring are not going to show any between-band correlations (e.g. you should not have any edges going from VBMROI-Freq1 to VBMROI-Freq2). Even though the level of correlations between these frequency-specific signals and the rest BOLD signals from other ROIs can still be correlated across subjects, the lack of direct links between these nodes means that including/excluding the nodes corresponding to the other frequencies is very unlikely to affect the global-efficiency measures your VBM ROI node. All this said, the same as in case (1) above applies, even if there will clearly be main effects of including/excluding these additional frequencies across subjects in your graph measures I cannot see any obvious mechanism by which this could introduce artifactual group interaction effects. Yet there are many mechanisms by which some other existing difference between your groups could lead to your observed results (for example perhaps the difference between your subject groups is due to the VBM ROI showing generally high connectivity with other key areas -so that it tends to survive the graph-forming threshold for all subjects- but that connectivity is higher in one subject group compared to the other. That would not show a global-efficiency effect when using the entire frequency band but it may show that effect when using a frequency-specific band by virtue of that lowering all the connectivity values between this ROI and other regions hence making those measures more sensitive to differences between subjects)

So summarizing, I do not see any obvious issues with these analyses but the interpretation of these results is going to be complicated/subtle. In order to convince yourself and the reviewers that this effect is real and actually reflecting something specific to the lowest-frequency band in the VBM ROI BOLD signal timeseries I would suggest to:

a) first make sure your results survive a bonferroni or FDR correction across your 3 frequency bands (you allegedly looked at the results for each frequency band separately unless you had an a priori reason to focus on the lowest-frequency band; this additional correction accounts for that);

and b) run some additional analyses trying several similarly-interpretable analyses (e.g. use the Setup.Conditions frequency decomposition to analyze the graph measures when restricting all ROIs and not just the VBM ROI to the same frequency band; try including/removing the additional left frontal pole ROI and/or including all three frequency-specific nodes into your graph analyses vs. using a single frequency-specific node; try using a single same-frequency-band for all ROIs but using a more restrictive graph-forming threshold -that will tend to decrease the global efficiency measures across all of your ROIs instead of just your VBM ROI) and see to what extent your results are robust to these variations and what interpretation best fits those results. 

Hope this helps
Alfonso