Dear Jerry,
Jerry Allison wrote:
> Search as I might, I can't find a succint answer to this rather simple question.
>
> I have two identical fMRI sessions in a single subject.
>
> The paradigm is a simple AB for each session.
>
> I prescribe the number of sessions as two in the SPM99 design matrix.
>
> This results in a design matrix having 4 columns. The first two columns contain the condition effect matrix for the two sessions, offset in time. The last two columns contain white bars representing session effects for each session, again offset in time.
>
> In oder to compute a t contrast for the specified condition, do I specify the contast as [1 1 0 0]?
>
> I have processed this set of data two ways. When I analyze the data as two sessions, I get what appears to be reasonable activation for the n-back task that we are using. When I process the data collectively as one session, at the same level of significance, there seems to be global activation. The activation patterns are probably similar, but when I lump the data into one session, I have to dramatically lower the p value to see the pattern. What would explain this behavior.
If I understand you correctly you see more activations when you treat the two sessions as one. I completely agree with Eric that this is quite counterintuitive.
I think there might be one tentative explanation though.
The session effect, which is no longer modelled in you second analysis, can be seen as a kind of step function when going from one session to the next. I.e. lets assume we have a "true" seesion effect that manifest itself as a step down when going from session 1 to session 2. Let us also assume this effect is large compared "true" activations, which is often the case. Lets further assume you have, for each session, an ABAB... design, such that both sessions start with A and
finish with B.
The stepfunction will now be modelled by the high-pass filter at the best of its ability, but due to its limited frequency content it wont be able to deal properly with the step at the junction between the sessions. Hence, the residuals after the HPF will have positive values at the end of the first session, and negative at the start of the next. In the next modelling step these residuals will be modelled by anything that has a "step" between the sessions, the obvious candidate
being the condition regressor which will look like (schematically)
[0 1 0 1 0 1 (end of first session) 0 1 0 1 0 1].
Bottom line, you may be modelling session effects with your condition regressor. Without the HPF this shouldn't happen, since then the session effect should be orthogonal to the condition effect. After the HPF it might not be, and the residual session effects could potentially be picked up as condition effects. This effect would be expected to manifest itself (spatially) like the session effects, and should therefore look like the "global" activation you describe.
Its a bit of a long shot, but it could happen.
By the way, why would you want t model them as a single session?
Good luck Jesper
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
|