Re: Question on RSA with CoSMoMVPA

[Nick Oosterhof]

Forwarding a message from Doris who recently joined the CoSMoMVPA google group:

On 7 August 2018 at 13:00, Doris Pischedda <doris.p...@gmail.com> wrote:

I would like to run ROI-based RSA using CoSMoMVPA. I had a look at the exercise in the LABMAN2017 course and have a question about the input to construct the dissimilarity matrix. It seems you calculate correlations using the glm_T_stats_perrun.nii file. I guess the file contains (contrast?) values for each condition and for each run, right? How do you construct that image exactly (e.g, with betas from the GLM, stimulus vs. baseline contrasts, etc.)? If you have some code for the procedure, that could also be helpful. Thanks for your help!

the gym_T_stats_perrun.nii files were generated using AFNI. First typical preprocessing was done (I think slice time correction, motion correction, and signal normalization by computing percent signal change). Then the general linear model (3dDeconvolve) was used with  boxcars of the onsets convolved with a canonical HRF. The values in these files contain t-statistic values for the contrast against baseline (not modelled, but these were periods when no stimulus was presented to the participants). 

Note that for RSA one has typically one predictor for each stimulus across all runs the experiment, not per run, so you could do a simpler analysis than we did in preparing the dataset. (An exception is cross-validated mahalanobis distance).  

Does that answer your question?

 

[Doris Pischedda]

Ok, thanks. I guess I should average beta maps for each condition across runs and create a 4D .nii file with imCalc. Then, it would probably work with the script in the demo.