Dear Jason,
> I will first start with what we have: Within an fmri study,
> One group
> Five subjects
> Two conditions
> Auditory Monitoring versus its own baseline
> Working Memory versus its own baseline
> Two nuisance variables
> anxiety score (one score per subject)
> Depressive mood score (one score per subject)
>
> One covariate of interest
> error score on the working memory task
> This is what we did
> Design Description
> Desgin: Full Monty
> Global calculation: mean voxel value (within per image fullmean/8 mask)
> Grand Mean scalingL (implicit in PropSca global normalization)
> Global normailzation: proportional scaling to 50
> Parameters: 2 condition, +1 covariate, +5 block, +2 nuisance
> 10 total, having 7 degrees of freedom
> leaving 3 degrees of freedom from 10 images
>
> Is this a valid way of looking at this? We are concerned with the
> large degrees of freedom that we are using up. Also how would we
> accurately interpret such a model? Does the statistical map only
> represent activations that are associated with the covariate of
> interest after controlling for anxiety and depression scores?
Firstly I assume this is a second level analysis where you have taken
'monitoring' and 'memory' contrasts from the first level. If this is
the case you should analyse each contrast separately. Secondly do not
model the subject effect: At the seond level this is a subject by
contrast interaction and is the error variance used for inference.
Thirdly a significant effect due to any of the covariates represents a
condition x covariate interaction (i.e. how that covariate affects the
activation).
I would use a covariates only single subject design in PET models (for
each of the two contrasts from the first level). A second-level
contrast testing for the effect of the constant term will tell you
about average activation effects. The remaining covariate-specific
contrasts will indicate whether or not there is an interaction.
I hope this helps - Karl
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
|