Dear Dr. Alfonso,

Hope you are doing well. I have a naive question: consider the case of task-based fMRI where we have multiple conditions. After denoising and concatenating condition specific TS, it is quite possible that some condition's time series will be shorter than the number of ROIs being looked at. In such a case, how is the semi-partial correlation calculated in such cases?


Thank you for your time and help


Regards
Pravesh

Dear Pravesh,

Sorry, computing semipartial correlation measures with insufficient degrees of freedom is simply not possible, in those cases the first-level General Linear Model required to compute those semipartial correlations will be over-determined and the resulting measures will be zero (or just machine-precision noise).

That said, this only applies to semipartial correlation or multivariate regression measures, where you are trying to compute the "unique" contribution of one ROI while simultaneously controlling for the effect of all others. In contrast, all standard bivariate correlation and bivariate regression measures would work just fine, since for those analyses the total number of ROIs is of no importance (a separate model is fitted for each seed ROI and target voxel or ROI)

Best
Alfonso


Originally posted by Pravesh Parekh:

Dear Dr. Alfonso,

Thank you for your reply and for clearing this up. I had checked one of such cases and became really puzzled when Conn was able to report semipartial correlation. Although the values were indeed close to zero, a lot of them turned out to be significantly greater than zero (one sample T-test), further adding to my confusion. Thank you for clearing it up. Perhaps Conn should generate a warning/error in such a case (I am still using 18b)?


Regards
Pravesh