Fisher's z transformation on correlation values

Hi @John_Mosher @pantazis !

I have a question regarding applying a Fisher transform on correlation values. One of my colleague suggested applying such transform in MIA (cf. PR in MIA).

For the context, in MIA, to assess the consistency of the activity recorded within a region, two correlation measures are computed

  • the inter-channel correlation (across the time courses of all relevant channels within the region)
  • the inter-patient correlation (across the time-courses averaged per patient within the region).

Both correlation metrics are defined as the mean of all pair-wise Pearson correlations within a region, either among contacts or among patient average time-courses.

I have two questions :

  1. What do you think about the approach?
  2. Do you think it is still valid even on a small number of values?

Thank you for your insights!


Hi AnneSo,

I have not conducted similar analysis in any of my work, but I would also suggest applying the Fisher transform on correlation values before averaging. An advantage is that you can directly average the z-values whereas with correlation you would need to account for the different sample sizes, an issue that may arise when averaging across different subjects. You provided a nice citation that is quite informative of the issue.

Regarding the idea of inter-channel correlation within a region, my feeling is that this would probably be more relevant for EEG (where patterns are more spatially coarse) than MEG.


Thank you for your insight Dimitrios!
Warm regards,