>On Friday 04/02 Dr Crivello wrote:
>We are trying to understand some results we obtain between the comparison of
>two contrast and the conjunction of the same two.
>Our design is
>Contrast 1 = activation 1 - control 1
>Contrast 2 = activation 2 - control 2
>we first tested the following design
>contrast 1 - contrast 2 and we observed an activation in the occipital
>This means that there is a significant difference between these two
>Then we tested the conjunction of contrast 1 and contrast 2, and we find
>also an activation of the same voxels.
>how can we interpret these results ?
>Richard Perry answered that the statistical threshold of the interaction SPM
>(activation 1 - control 1) -(activation 2 - control 2) was different from
>the threshold of the conjunction.
>This is not the case : the threshold was p<0.001 uncorrected for both.
>It is still not clear how we may interpret these results? the contrasts
>comparison evidenced that the rCBF increase was significantly greater in
>contrast 1 than in contrast 2 while the conjunction analysis revealed that
>for these voxels the activation was not significantly different....
Dear Drs Etard and Crivello
I think the confusion might relate to the different thresholds that are
being referred to.
A) the threshold of the specified interaction
B) the threshold for the conjunction; and
C) the threshold for eliminating interactions from the conjunction.
Option C is only available in SPM97. If you have a previous version of SPM,
it is probably not appropriate to use the conjunction facility.
If you are using SPM97 and you used a threshold of p<0.001 for C then this
probably accounts for why you have an overlap in your conjunction (B) and
interaction (A). The statistical procedure for calculating interactions
is different for A and C. To ensure that you have excluded all
interactions from the conjunction (C) you need to exclude all interactions
at p = 0.05. ( p<0.05 is stricter than p<0.001
because it eliminates more interactions).
As I wrote in a previous email today, the other option is
that your region has voxels that respond to both contrasts and voxels that
are differentially affected by one or the other BUT you should not get
both effects in the same voxels.
If you look at the Z scores in a specified voxel for contrast, 1, contrast
2 and the
interaction, this will indicate where the mistake has arisen.