Hi Steven,
I would first consider the ultimate purpose of such a regression. In the FA example, what this process is doing is saying “if FA is globally higher / lower throughout the entire WM in this subject, regress that effect out, and then see if there’s a residual difference in this specific region”. How does this logic then extend to fixels and the AFD metric? Personally I would ask:
“If the estimated WM-like density is globally higher / lower throughout the WM in this subject (even after intensity normalisation), regress that effect out, and then see if there’s a residual difference in a specific region.”
If it were me operating under this premise, I’d be capturing the mean AFDtotal throughout the WM / analysis mask and using that as the global regressor, since it’s the most faithful interpretation of the purpose of such a regression. You could use the mean AFD across all fixels within the template fixel analysis mask; as you say it’d effectively give greater “weight” to regions of the template with more fixels compared to those with fewer fixels, but that bias would be equivalent across all subjects, so I’m not convinced that it’d be a huge issue. But it would make estimation of that regression downstream of the effects of FOD segmentation and fixel correspondence, which I think are just confounds to the derivation of that regression that are not beneficial to its purpose.
Cheers
Rob