Dear Juan Chen, I can imagine a case in which this can happen ...For example, if a region comprises different functionally defined subregions – take the cingulate cortex, which can be divided into more anterior portions (ACC) and more posterior portions (PCC). While the ACC is believed to be more engaged in "cognitive" tasks, the PCC is believed to be more engaged in "emotional" tasks (cognitive-affective division of the ACC, Bush et al., 2000). Similar subdivisions have been proposed for the mPFC as well (Etkin et al., 2011) – transferring this rationale to your task, it could also be that two subregions are more strongly connected during one condition than during the other. Moreover, I presume that – even if it is the same region – the peaks of activation should be different.  

Dear Juan, 

To my knowledge, if you are doing a group analysis or between group analysis, you should be consistent across your study. I would suggest you (even you are not interested in the idea) to use the same seed and repeat your analysis to see what results you get. When we don't use the same seed / ROI even they are similar, it's hard to interpret the results across a group of between group. 

Hope it helps,

Saman

In my humble opinion, defining ROIs individually is also acceptable (in fact, there are several possibilities of dealing with this issue):

"If you are defining ROIs individually (as in the second functional-ROI case, or some of the anatomical cases) you analysis will likely work better (have higher signal-to-noise) if you keep the ROI small. This is because you are only taking one measurement from the whole ROI – so by enlarging the ROI to include voxels with a weaker effect you are actually ‘watering down’ the signal."

source: http://fsl.fmrib.ox.ac.uk/fsl/fslwiki/PPIHowToRun

I agree with Ramzi. Smaller ROIs give higher SNR. Moreover,  there are several way of doing the ROI definition. Sticking to native space ROIs has the advantage that you definitely get the brain response out of the correct anatomically region of that person without any warping. But different approaches (e.g. MNI ROI) should give comparable results.

Best,

Daniel

I am doing similar analysis subject-wise for the same pair of seeds and performing regression henceforth. However, I am in a fix on not being able to quantify correlation coefficient to some standardized score to check for relevance. If anyone could throw some light on this matter? Thank you.