Dear Joaquim, dear SDM experts,

I read some new papers describing a method to analyze the similarity between two meta-analytic result maps with image correlation (pearson r between images transformed to 1D feature vecture, see e.g. Riedl et al., 2015, NIMG, "Meta-analytic connectivity and behavioral parcellation of the human cerebellum"). Based on the image correlation, it is then possible to carry out things like hierarchical cluster analysis over multiple maps.

I thought about trying out this method for maps from sdm. In principle one could calculate image correlation for either thresholded or untresholded result maps, which are named something like "MyMean_z_p_0.00500_1.000_10.nii.gz" and "MyMean_z.nii.gz" I think. I would prefer to work with untresholded maps, because that avoids having to select a particular threshold and keeps all the information of an image. Image correlation analysis between untresholded maps (e.g. contrast images from GLM analysis) has been repeatedly done for original fmri data. Do you think it is – from a  statistics point of view – in principle okay also to use untresholded meta-analytic effect size maps for image correlation, just like people are using untresholded fMRI maps for image correlation? And would I have to take the "MyMean_z.nii.gz" then?

Many thanks for your help, and also thank you so much for hosting this nice forum. I really appreciate the support!

Best,
Matthias Schurz

Dear Matthias,

This sounds very interesting.

I would also prefer to work with unthresholded maps, as they have all information, and in this case I would use the effect size or the z-value map. I would maybe use Spearman correlations.

The use of thresholded maps, or even binary maps, would be also an alternative, though then I would maybe use something such as the Dice coefficient rather than a correlation coefficient.

Best,

Joaquim