mtpc edge-level data
kcampbell
Hi Chris,
First, thank you for your excellent work on the braingraph R package — it’s been an invaluable tool for my research on neonatal structural connectomes.
Apologies if this topic has been addressed before, but I’m wondering whether there are any strong reasons against applying MTPC to evaluate edge-level GLMs?
To provide some context: I’m working with 64×64 connectivity matrices derived from probabilistic tractography of diffusion MRI data. These matrices are symmetrized (with diagonal zero), yielding 2016 unique edges. I’ve applied a set of 39 thresholds to the data and have used MTPC to test group differences in various global- and vertex-level network metrics, as per the braingraph R user manual.
I’m now considering applying MTPC at the edge-level to maintain a consistent threshold-based framework throughout my analysis. My rationale is to avoid the threshold-dependency required by more 'conventional' approaches like the NBS, which necessitate the a priori selection of a single threshold — something I’m trying to steer away from to avoid such bias.
I recognize that applying MTPC at the edge-level introduces substantial multiple comparisons, of course. That said, my understanding is that MTPC incorporates FWER correction across both thresholds and spatial units (e.g., vertices/nodes), and this would extend to edges as well.
Would the main limitation here be reduced sensitivity or increased conservatism (compared to NBS, for example)? Or are there conceptual/statistical reasons why edge-wise MTPC might be discouraged?
For what it’s worth, I’ve modified the source code to allow MTPC to be applied to edge-level metrics – and it worked well (aside from the greater computational demand). Before fully adopting this, though, I’d really appreciate your feedback / hear your thoughts!
Thanks again for your time and for developing such a great package!
Warm regards,
Kayleigh
