Roland,
> I am aware that the assumption of exchangeability doesn't hold in a 1st level
> fMRI analysis because of the autocorrelation of the scans. On 2nd level the
> assumption surely holds for intersubject paradigms. But what happens to
> intra-subject paradigms on 2nd level? The documentation I found left out this
> part.
The short answer is that, yes, exchangeability holds at the second
level as long as you only analyze one scan per subject. This follows
from independence of different subjects.
> As we compare conimages which for the T-test equivalent most often are
> differences of beta weights I would see no problem for
> a) Block designs in which the blocks of interest are reasonable far apart (in
> time)
> b) sparse imaging designs with high TR
>
> as in both cases the beta weights themselves should not be
> correlated. Is this assumption correct? What about "normal"
> event-related designs in which the event types are
> separated in time by about the length of an hrf?
>
> And what about sparse imaging designs on a first level? Is there some rule of
> thumb how long TR must be for autocorr. being low enough for using SnPM here?
>
At the second level, with only one measurement per subject in the
model, it doesn't matter if the beta weights (within a subject) are
correlated. You are boiling it down to one measure per subject, and
hence the only possible correlations are between subjects (which isn't
a concern).
Does this clarify things?
-Tom
|