I would like to do trial-wise decoding. Since the example you provided is based on run-wise beta estimates, I have a few questions about the GLM specification and cfg.design creation in this case:
My task is a 200-trial (5 runs) ER design, with 3 events in each trial: fixation, cue, action. I am interested in the cue phase and would like to classify the reward magnitude (high vs. low). Can I use 200 regressors to represent 200 cues respectively (and another regressor of no interest to model all action phases), and use the generated 200 beta images for the decoding analysis?
I can either group these 200 beta images based on 5 runs (40 trials in each) or treat them separately. Should I do a leave-one-run-out or leave-one-trial-out CV? which do you recommend?
Since the reward magnitude is estimated from a model based on subjects’ actions, the number of high- and low-reward trials may be not equal in each run. Is it necessary to do bootstrap samples to maintain the balance of the training data?