Dear Henrik,
> We have conducted a Working Memory study in H2O-PET with the following
> design:
>
> 8 subject
> 4 Conditions:
> A: WM 1 (high load)
> B: WM 1 (low load)
> C: WM 2 (high load)
> D: WM 2 (low load)
> 3 Scans/per condition/subject
>
> Thus we have a total of 96 scans. To look for the main effect of
> Working memory and for the domain specific effect we have chosen the
> Multi-subject: cond x Subj interaction & covariates design. We find
> nice WM main effects and also interesting domain specific effects.
>
> Experimental question: We are interested if there is a correlation
> between performance (as measured by RT) and WM-specific activation.
>
> What kind of design do we have to choose? As we work mainly with
> fMRI-studies we first thought of a second level analysis. That is:
> Formulate subjects specific contrast, e.g. WM 1 high load minus WM 1
> low load on the first level, then feed the eight resulting con-images
> into basic models (simple regression) and chosse the median of the RT
> for the three WM 1 high load scans as covariates into this modell. Then
> our analysis should yield regions in which there is a correlation of
> one WM domain with performance.
>
> Question 1: Is this analysis correct?
Yes it is but it may not be the most sensitive analysis because you are
proceeding to a second-level analysis whereas you have scan-specific
performance measures.
> Question 2: There are several options at the first level in which it is
> possible to specify scan specific covariates. Is it intelligible to
> choose one of these models and feed in scan specific RT in order to
> answer our experimental question? If so, what is the appropriate model?
> We have tried several but if we enter scans and covariates we use up all
> our degrees of freedom, e.g. if we use Mulit-subj: covariates only. So
> something must be wrong.
I would simplify your model and omit subject x condition interactions.
You could then enter the condition x performance interaction as a
covariate of interest. This is simply the mean corrected performance
data multiplied by 1 for high load and -1 for low load. You could do
this in a condition-specific fashion for WM 1 and WM 2 using two
covariates but ensure the behavioural data are centered within
condition before constructing the interaction (here use 0 for the
'other' condition).
I hope this helps - Karl
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
|