Hi Rich,
only some ideas to that topic..
this regressor of no interest should be highly correlated to the paradigm, mostly the errors will occur in the 2back-blocks. So for that it seems to be problematic to model this extra regressor of no interest, like I would not use extra regressors of no interest for movement parameter, if there is task-related movement, because in that case you can not differ between nuisance variance and good variance you want to explain. But if you do so, a single regressor will be enough I would suggest. The only reason to model two accuracy regressors of no interest should be, that you will include one of them into your contrast in a different way than the other.
Depending on your intention other suggestion should be a parametric modulation of the amplitude of your n back regressors (in that case with trial onsets) in order to to model out this accuracy related variance.
May be you could use only the session data in the analysis, where the accuracy was over a specific theshold..
Kind regards,
Christoph
|