I am trying to run an MVPA analysis and after I run the first level, I am looking at the summary tab at the second level. I am not getting coverage of more than 30% of the voxels according to the graph even with all 10 factors. See attached screenshot I am wondering if I might be doing something wrong in the setup? Any advice would be appreciated .
Thank you!
Amy

Hi Amy,

That just indicates that the between-subjects variability in seed-to-voxel connectivity maps is relatively large. That could be caused by a particularly heterogeneous sample (e.g. when studying a combination of very different groups of subjects), or by either acquisition (e.g. short resting-state sessions) or noise factors (e.g. relatively large residual motion or physiological noise) which may impact the reliability of the individual-subject connectivity estimates (indirectly making the observed between-subjects variability larger). If the former, then you simply need more factors than usual in MVPA (and that will typically require relatively larger sample sizes in order to accurately measure them), while if the latter, then you could try to find ways to increase the robustness of your individual-subject connectivity estimates by a combination of more aggressive denoising, larger spatial smoothing, aggregating across multiple conditions if possible, etc.

Hope this helps
Alfonso
 
Originally posted by Amy Roy:

Dear Alfonso or any CONN expert,

I have been using CONN to investigate  rsfMRI differences between controls and patients with mental disorders. 
 
I found interesting results with MVPA but although I read your very good 2020 handbook, it's reference on MVPA (the review Norman et al.,2006), and also browsed through CONN's help search, I still haven't found answers for crucial questions. 

Would you be so kind to clarify or point to bibliography that answers the following?

1- Which is the classifier performance from the results I get in CONN's MVPA?

2- In section 'MVPA Methods' of Norman et al.,2006 the authors divide MVPA in 4 steps (feature selection, pattern assembly, classifier training and generalizing testing). For an analysis that only inserts a Patients > Controls contrast, which are CONN's default settings regarding those 4 steps?  

3- For this Patient > Control contrast I get negative connectivity at many cerebellum's sub-areas. How does CONN compute this difference? (e.g. separate MVPA results are computed for each group and then the t-tests to find differences at each voxel's time series?)


Thank you very much for your attention.
All best
Renzo