Dear AFNI team,

I am running an analysis on a data set that has been acquired using a mixed block/event design stimulation protocol and I have some doubts about how to model it correctly, given various collinearity issues popping out from 3dDeconvolve. I'll briefly outline the design, for clarity:

- there are 6 EPI runs
- on runs 1, 3, 5 subjects are instructed to approach the performance of the task with a specific "mental attitude" A
- on runs 2, 4, 6 subjects are instructed to perform the identical task but with a different "mental attitude" B (note that the external simulation does not change between odd- and even-numbered runs)
- within each run there are 4 active task blocks (approx. 30 sec each), alternating with baseline blocks of passive fixation (approx. 20 sec each)
- within each block, there is a rapid presentation (every 2.5 sec) of two classes of stimuli (congruent and incongruent, Stroop-type stimuli), in random order

I am modelling each event type (congruent, incongruent) separately for each run, because I am interested in effects arising from adopting different "mental attitudes" (A and B) in performing the task, which is a between-runs effect.

Now, if I apply 3dDeconvolve to the data set (including motion parameters, and a few other regressors for error trials and visual cues for the beginning and ending of task blocks), I see that there is often a "medium" strength collinearity (r ~ 0.5-0.6) between the regressors for congruent and incongruent trial types belonging to the same run (I can send you an example plot as an attachment, along with the full design matrix if that helps). This collinearity seems to arise from the block structure embedding the events, that is, it seems to me that the large part of shared variance between the two regressors is represented by the steep climb and descent phases from and to the baseline that mark the beginning and the end of a block, respectively. I wonder if this collinearity is a problem and, if it is, what would be a viable strategy to get around it.

To complicate things further, I was also contemplating including in the model, in addition to the event-related regressors, six gamma-convolved box-car regressors representing just the block structure for each run, in order to try and dissociate the effect of a putative cognitive component related to task performance that is *sustained* during the full length of a block (a kind of "mental set"), from the transient effect of performing the single trials within the block (see e.g., the recent review in Neuroimage on mixed block/event-related designs by Petersen and Dubis). Now, I would imagine that collinearity issues would get even worse with this new model and I would greatly value your thoughts on this (e.g., should I orthogonalize the regressors modelling the events to the regressors modeling the blocks?).

Thanks in advance for any suggestion,


giuseppe