Hi Gregg,
What I think you'd like to do is include EVs-of-no-interest to account for the variance associated with these timepoints, when participants/patients fall asleep.
For those subjects that do not have these microsleep events, you include an EV as well - named the same and in the same order - so to allow for averaging. In this case, the EV is just dummy coded - in principle all 0s will do, but i believe FSL doesn't like that - so, we use a .txt file with: 0 0 1 - make sense?
For real sleep events, but in the start and duration, as usual.
You will need to fix your real condition EVs accordingly, and if sleep events happen in the middle of a block(s), it may be tricky to salvage the condition periods - although it should be fine. you just have to worry about overlapping independent EVs - as the model may becomes rank deficient - essentially meaning your predictors correlate too strongly; overlap - so that the variance can't be appropriately parsed.
In this case, i would declare the whole block of no-interest - make the error ev the whole duration, to soak it up and salvage the rest of your run.
Oh, and you want to convolve these sleep EVs of no-interest, in the usual way - double gamma function.
Cheers,
Ken
|