On Thu, 4 Sep 2008 21:58:47 +0100, Bob Spunt <[log in to unmask]>
wrote:
>Dear SPM Experts,
>
>Let's say you have a study that you've scanned in three functional runs.
>When are you EVER justified in modeling this as a single session? For
>example, in past inquiries on this I've seen folks who have done this
>because of a priori interest in session effects; others have done this
>because the frequency of specific types of events is confounded with
>sessions (e.g., condition A has 1 instance in Session #1, but 8 instances in
>Session #2; this typically comes up a lot when conditions are defined by
>subject response). But despite the fact that I know that it HAS been done,
>I'm not sure if it SHOULD be done, so I'm writing to see if anyone has
>thought about this and has a relatively straightforward answer.
>
>If there are cases where this is an acceptable analytical option, the second
>question is how to do this appropriately. I've attached a picture of two
>models of the same single-subject data, one modeled as two sessions and the
>other as one. In order to model as a single session, I added in a regressor
>for each of the sessions (last two columns of the single-session model). Is
>this an acceptable method for accounting for session differences in a
>single-session model? If not, what is?
>
>Any advice would be much appreciated...thanks in advance for any help you
>can offer!
>
>Sincerely,
>Bob Spunt
It's an empirical question, not one you can answer _a priori_.
By "add[ing] in a regressor for each of the sessions (last two columns of the
single-session model)," you've modeled what one might call the "main effect of
session".
By your model _not_ having one column for each condition _and_ each
session, you're not modeling a condition-by-session interaction.
In statistics more generally, there are presumably many real-world situations
where interactions which might occur don't alter the model's fit much.
In fMRI, one can do what you did---examine it as an empirical question---or
try to come up with reasons why it may or may not matter.
Aside: in the first session in your experiment, it appears the first
condition "lives" in the first half of the session, whereas the other three live in
the second half. This might be problematic, insofar as it introduces some
confounding with drift.
Cheers,
S
|