Subject vs. Session Standardization and Functional Connectivity


I’m looking at differences in functional connectivity between two conditions. Each of 8 subjects has 10 fMRI sessions across months; each session has ~700 trials. Specifically, I’m analyzing denoised beta activations for each trial. In preprocessing the data before analyzing roi-to-roi correlations, I standardize the betas (for each averaged roi) per each subject. I do so because differences in activation between sessions are of importance, and conditions of interest are distributed differently across sessions (i.e. there is more of Condition A in earlier sessions vs. later sessions). Also, perhaps of note, many trials are neither A or B.

Are there strong reasons against this procedure, and instead standardizing by session? To infer significance, I simply combine all subject’s data use a non-parametric, label permutation procedure.



  • I think that if you keep the session effect in your data, the permutation scheme should respect this structure: permutations should be within-session and not across session. Otherwise, your null cistribution would actually ignore the session effect, which is inconsistent with your modeling assumption.

  • One question is: does the session effect matter in terms of neuroscience ? Is it meaningful ? I am not so sure, and for that reason, I would try and normalize session-wise (which is not a big deal, given the number of trials per session). That would be cleaner I think (and bzw, I would keep the permutation scheme session-wise).



Hi @bthirion ,

Keeping the session-structure in permutations cleaned my results some (and took a bit longer), but thank you!

I’m looking at long-term memory effects, memory activation patterns change quite a bit from session to session. But I think you are right in that a session-wise normalization is cleaner.

One follow-up question, when I’m normalizing my betas per session–should I first limit my trials to those conditions of interest (~40-80% depending on the session)? Or should I continue to include all my trials before normalizing, then extract betas of interest?

Many thanks,


Hm. Not fully sure what to do.

If there is a strong effect of the conditions of interest, given that this proportion varies from session to session, this creates indeed a confounding effect, and you should rather limit the trials to those of the conditions of interest. But that’s just an educated guess.