Hello,
The answer depends on the type of data/variable you want to input to MVPA and LibSVM. If you want to input the effect size (i.e. the betas) for each stimuli, you can compute them with 3dDeconvolve using --stim_times_IM. It might be a good idea to run 3dREMLfit to account for serial correlations. However, if the trials were highly correlated (e.g. in a fast event-related design), the program 3dLSS might be a better option as shown in the paper that described this approach.
[
dx.doi.org]
You can also have a look at 3dSVM for the implementation of the support vector machines in AFNI.
Best,
Cesar
