RSA Group Analysis
Zoltan Apa
Dear Community,
My following question is concerning RSA analysis, but first the facts:
I have 2 groups (older and younger) and 2 fMRI conditions (Encoding and Retrieval).Whole brain search light procedure was conducted to compute Encoding Retrieval Similarity (ERS) maps at Item and Set levels for each individual. Item level was computed for repeated presentation of the same stimulus. Set level was computed as an interaction between a given stimulus and all other presented stimuli for all other subjects. Thus, Item level is computed based on specific reactivation specific to a single stimulus, Set level is computed based on general reactivation of stimuli information. In other words, Set level serves as a baseline for the Item level.
Objective is to decipher possible differences between the groups.
After obtaining the ERS maps, the common practice until now is to compute (Item level)-(Set level) at group level i.e. (Yi-Ys)-(Oi-Os) to assess group differences.
Q. Since the ERS maps were computed by default by contrasting between Item and Set levels, is it necessary to contrast the two levels/groups (Yi-Ys)-(Oi-Os) to show group differences? Or a comparison: (Yi-Ys) compared with (Oi-Os), or (Yi) compared with (Oi) results would suffice evidence for finding/not finding group differences.
Any feedback is very much appreciated!
Zoltan
Nick Oosterhof
A (naive?) approach might be to compute, for a particular participant in a particular ROI (set of voxels) the similarity of the encoding similarity matrix (in vectored form, see cosmo_dissimilarity_matrix_measure) and the retrieval similarity matrix (also vectorized). This would then result in a single similarity value (correlation) for that ROI. Group differences can then be assessed through a two-sample t-test of older vs younger. Running a searchlight would result in a whole-brain group differences map. Or do you see reasons why this may not work?
Also: when computing dissimilarity matrices, it may be good to use center_data=true.