Dear Benjamin,
> Can anyone tell me how to use SPM96 to do a group analysis with
> fixed-effect model together with conjunction analysis as suggested by
> Friston, Holmes, and Worsley (1999)? I have an fMRI study with a group
> of only six subjects. Each subject performed four experimental
> conditions/tasks (A, B, C, D) in a box-car design with D (control
> condition) alternating with one of the other three conditions in each
> run as follows:
>
> Run 1: D A D A D A D A
> Run 2: D B D B D B D B
> Run 3: D C D C D C D C
>
> Each run was repeated three times for every subject. Now, I want to see
> if the cortical area(s) activated in a condition is statistically
> reliable across subjects. Because there are too few subjects in the
> group, random-effect model is not appropriate for the analysis.
> According to Friston, Holmes, and Worsley (1999), the fixed-effect
> model together with conjunction analysis is a good way to determine
> reliable activations when there is a small group of subjects. How do I
> use SPM96 to perform the analysis?
There are two levels to your question:
1) How does one do a conjunction analysis to test for conjoint
activation over several subjects?
2) How does one extend the inference to the population from which the
subjects came? (as described in Friston et al 1999 and in greater depth
in Friston et al (1999) NeuroImage 10:385-396)?
The implementation of conjunctions in SPM has changed from SPM96 to
SPM99 because of advances in Gaussian field theory due to Keith
Worsley. Now conjunctions are obtained simply by a voxel being
significant in all the contrasts (in SPM06 a conjunction was tested
with the average over the contrasts in, and only in, voxels that did
not show significant differences among the contrasts. Cathy Price will
be writing an update paper on conjunctions in the near future).
You cannot do conjunction analyses in SPM96 of the sort that allows for
the generalization to the population in terms of corrected P values.
However by mutual masking of all the subject-specific contrasts the
minimum T value (or less) can be specified and this value can enter
into the equations for uncorrected p values in Friston et al (1999)
NeuroImage 10:385-396. I would recommend this approach or using
SPM99.
I hope this helps - Karl
|