Hi Alfonso and CONN team,

I have run MVPA analyses and observed significant results. However, the cluster sizes are small and I wonder whether this is an issue? Details of my study are as follows:

4 groups: 

• participant with neurodevelopmental disorder (n=38)


• control group (n=37)


• participant with neurodevelopmental disorder + condition (n=33)


• control group + condition (n=17)

Analyses:

• MVPA analyses (10 factors, 64 components)


• results are more fragmented plus very small clusters (k<10) for 9 and 10 factors


• Analyses run to look at:


• 1. Effect of group on connectivity controlling for covariates + condition. 7 MVPA factors were deemed suitable


• 2. Effect of condition on connectivity controlling for covariates + group. 6 MVPA factors were deemed suitable


• 3. Interaction effect of group x condition on connectivity controlling for covariates. 6 MVPA factors were deemed suitable

Results:

For each set of analyses run, significant clusters surviving FWE were observed. Cluster sizes ranged between 16-34. When non-parametric analyses were run, no significant results were observed. Can you advise whether small cluster sizes are an issue? Is there an expected cluster size? Are such results publishable? The MVPA factors differ for the various tests. Is this allowed, or should they be kept consistent across all tests? What can I do to improve? 

Thanks.

Your study Ns per group are likely too small for the MVPA procedure.  I generally limit to 6 MVPA eigenpatterns when the N is in the ~80-100 range.

Did you look at the eigenpattern distributions?

Warm regards,

Jeff

Hi Jeff,

Thanks for responding. 6 MVPA eigenpatterns were deemed suitable for my study as well.

Kind regards,

Ranila

Hi,

I am encountering the same problem in my analysis with a small sample size of only 10 subjects. I also observed significant clusters which survived FWE, but no cluster was observed when TFCE and permutation test were run.

Following some general AI-based suggestions, 2 or 3 factors were recommended. This seems to align somewhat with Alfonso’s previous advice regarding small samples, where a ratio of 5:1 or 10:1 (subjects to factors).

In my testing, reducing the number of factors indeed led to an increase in cluster size. However, I am still concerned about whether 10 samples are fundamentally too small for a robust fc-MVPA.

Additionally, I noticed you mentioned that 6 factors were deemed suitable for your study. I would be very interested to know how you determined that 6 was the optimal number for your specific study? Did you base this on the cumulative variance explained?

Thanks.