Dear Paola,
My guess is that one or more of the patient groups are very different
from the other patient groups. In the two-level model, this results in
the single beta for the pooled patient groups not fitting the data
very well, and hence large ResMS, and relatively reduced significance.
In the five-level analysis, the overall fit of the model is improved,
the patient betas can differ, but on average, they differ from
controls in a similar way to before, but with greater significance due
to smaller ResMS.
You can check whether my guess is correct by:
- comparing the four patient beta images in the five-level model
(should show some differences)
- comparing ResMS images (should be lower in five- vs two-level)
- comparing con images (5 and 2 should be similar, from my understanding)
- comparing spmT images (presumably stronger in 5, since you found
more significance)
If comparing images seems difficult, you could just compare
mean/median of in-mask voxels, or pick a voxel of (ideally a priori)
interest, and just look at that voxel.
Best,
Ged
P.S. For the AND question, I think you could look at conjunctions of
the four control vs patient contrasts:
[-1 1], [-1 0 1], [-1 0 0 1], and [-1 0 0 0 1]
with the conjunction null hypothesis. I don't think you could use the
global null to answer the OR question, since the contrasts will
probably be dependent; the conjunction null is valid regardless of
dependence, but the intermediate or global nulls require independent
contrasts (see help spm_getSPM (which is what I just had to do to
remind myself of this!))
On 6 May 2010 10:12, Paola Valsasina <[log in to unmask]> wrote:
> Dear SPM List,
>
> I have some doubts about the results of a full-factorial analysis and want to be sure that I specified the correct design.
>
> I want to compare healthy controls activations with multiple sclerosis patients. These patients can be divided into four sub-group, according to disease phenotype (with progressive gravity).
>
> First, I did a statistical design letting all patients together (full-factorial design with one factor: group, and two levels:controls and patients). I compared activity in patients vs controls by creating the contrast -1 1 and obtained some clusters of differences.
> Then, I did again a full factorial design with one factor (group), but with five levels, keeping the patients in four separate groups according to the disease phenotype. In this latter model, I did again a comparison of all patients vs controls by creating the contrast -4 1 1 1 1. In this way, I obtained much more differences than those of previous two-level model.
> Why does this happen? Maybe because in the first model I see only differences present in all subjects (a sort of "AND" condition) whereas in the second model I see also differences driven even by one single subgroup (a sort of "OR" condition)? I am not sure if this is the explanation or if there is any other issue that I am ignoring..
> Thank you for any advice
> Kind regards,
> Paola
>
|