FBA case-study

1

Dear Mrtrix experts,

I have more of a methodological question:

As part of my PhD thesis I wan’t to do a case-study using a FBA to compare a single subject to a group of healthy controls (30-40). Subjects match in age, education-level and gender. All DWI scans were done with a b-value of 1000.

I plan to follow the steps outlined in the current MRtrix 3.0 documentation
(http://mrtrix.readthedocs.io/en/tag_0.3.16/fixel_based_analysis/ss_fibre_density_cross-section.html).

Are there any steps that need to be adjusted? Is it possible and valid to use the FBA on a single-subject level?

Any advice, help or suggestion is greatly appreciated,

Thank you very much
Tom

2

I’m not sure we’re going to be able to help a great deal here, beyond what’s already in the documentation. There are issues in general with doing single-subject v.s. group comparisons, but I’m pretty sure I’m not I’m the best person to discuss them… This is something that I’d advise you to discuss with statisticians familiar with neuroimaging, see whether they raise any concerns. But otherwise, I can’t think of anything that would differ in terms of the actual steps required - everything would be the same in terms of preprocessing all the way up to the final statistical analysis.

We’d also need to check whether the fixelcfestats is even capable of running such an analysis - it definitely wasn’t designed for it, so there is a chance it might fail completely with a group size of 1…

3

Hi Tom,
It won’t fail, but you will only have 30-40 possible permutations/relabelling of the ‘groups’. Say you have 30 controls, then your minimum achievable p-value would be 0.033.

Unfortunately the way fixelcfestats is implemented it doesn’t check for duplicate permutation labelling while taking into account your groups (it assumes that when your subject numbers are large enough, the chance of duplicate group assignment is low). So this means it will actually run 5000 permutations (default) and generate a very poor permutation distribution (with many repeated values).

There might be other ways to perform 1-vs-many permutation testing other than just having a very large number of controls. However, I’m also not really up with the literature on this.
Cheers,
Dave