Dear SPmers,
I have questions about the most appropriate matrix design in SPM99b for multi-group multi-conditions, and random effect analyses. Basically, I have two PET studies with four different conditions each, and several of my cognitive components of interest (CCI) are shared by some of the 8 conditions used in these two study.
1. In SPM99, I run a first analysis with the following parameters in The Full Monty... option :
- 2 groups
* group 1 (ST1) : 12 subjects X 12 PET scans each, four conditions : a b c d
* group 2 (ST2) : 18 subjects X 12 PET scans each, four conditions : e f g h
- no covariates or nuisance variables
- AnCova by subj (within group)
- scaling of subj (within group) grand means (to 50)
- Centre global covariate (after grand mean scaling) around subj (within group) means (<= AnCova by subj (within group))
- Threshold masking proportional 0.8
- Implicit mask : yes
- explicit mask images : no
- Global calculation : mean voxel value (within per image fullmean/8 mask)
On this basis, I compute simple two-conditions subtractions (within each group) , interactions, conjunctions and so on ...
My first question is about the influence exerted by group2 on the contrasts computed in group1. If I run an analysis using only group1, the contrasts are the same, but the number of degrees of freedom is lower, because only the subjects involved in this comparison are accounted for. Given that one of my CCI is evidenced in a conjunction involving ONLY the conditions from group1 ((1-2)+(3-4)), does it matter that this result was obtained in the background of a multi-study design matrix (and therefore with more df)? And is this correct to consider that it does not matter given that I used the within group options for centering, scaling and Ancova ?
2. I was interested to similarly test these data using random effects model. Andrew Holmes wrote in spm_RandFX.man that more than two conditions may be included, but should only consider contrasting them in pairs. However, in a spm-list message (15 Dec 1998) about random effects and conjunctions, he state that it may be inaccurate to put more than two conditions into the 2nd level of an SPM random effects comparison because the independance criterion is usually not met in this situation, and the net result may be to have spurious additional df and therefore invalid inference. Hence, I wonder if the following procedure might be really valid:
**1st level : to create individual adjusted-mean images with the option MultiSubj: Condition means (AnCova by subject).
**2nd level : in PET models, using Multi-group: conditions & covariates:
- 2 groups with their 4 conditions each (1 adj-mean image / subject / condition)
- no covariates, nuisance, global normalisation, grand mean scaling
- Threshold masking proportional 0.8
- Global calculation : mean voxel value (within per image fullmean/8 mask)
... and them compute simple categorical paired comparisons, but also interactions or conjunctions between paired comparisons. Actually, the resulting SPM maps are quasi identical to the ones obtained using the fixed-effect analysis described above, except that if this procedure is valid, I can extend more validly my inference to the population level and increase the robustness of my results.
I will really appreciate any comments and/or suggestions.
Philippe
//// ,,,, ////
--[[. .]]------[* *]------[@ @]--------------
[~] [~] [#]
PEIGNEUX Philippe, Lic. Psych., Chercheur
E-mail: [log in to unmask]
Cyclotron Research Centre - Neurology Unit
Liege University
Batiment B30
Allee du 6 Aout, 8
B-4000 Liege
BELGIUM
phone: +32-4-3662316
fax: +32-4-3662946
Neuropsychology Department
Liege University
Batiment B33
Boulevard du Rectorat, 3
B-4000 Liege
BELGIUM
phone: +32-4-3662394
fax: +32-4-3662808
http://www.ulg.ac.be/neuropsy/
---------------------------------------------
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
