Hi @anege,
It’s entirely technically valid both to produce a fixel mask in subject space and compute mean FD within it, and to produce a fixel mask in template space and compute mean FD within it. The reason the latter is more conventionally employed is because the delineation of the pathway of interest in derivation of that mask only needs to be performed once, and can then be applied across all subjects. The former could be done if you could specify an automated & robust mechanism for consistently segmenting the same pathway across subjects within the space of each individual subject, but even doing so there would still likely be residual doubts regarding whether any effect observed was due to a difference in FD across subjects, or a difference in the behaviour of the segmentation method across subjects.
In this specific case the difference is arising through use of the fixelcorrespondence command, which is really a quite basic process in its current form, and may or may not behave as one may subjectively expect it to in any regions where the subject FOD and template FOD are not highly similar.