Hi Markus,
> like to test the difference between two groups according the correlation of
> painexperience with the BOLD-signalchanges in fmri.
My understanding of this question is that you want a model:
fmri = a1 + b1*painexp + a2 + b2*painexp
and that you are interested in how the slopes/correlations with
painexp differ between the groups (e.g. null-hypothesis b2=b1)
One way to view the above model is that you have a single factor -
group (which has two "levels"), and a single covariate - painexp, and
you wish to model the main effect of group along with the interaction
of group and factor.
Hopefully the terminology I have used in the previous paragraph both
makes sense to you and fits reasonably well with the flexible
factorial setup in SPM5. Note that painexp is a covariate, not a
factor, but that you are interested in the covariate-by-factor
interaction.
Create a factor for group, "specify-all" scans (I assume all group1
first then all group2), and enter the factor matrix as(*1):
[ones(N1+N2,1) [ones(N1,1);2*ones(N2,1)] ones(N1+N2,2)]
where N1 and N2 are the numbers of subjects in your two groups. Now
add a new Main effect, for factor 1. Don't add any interactions here,
as you only have one *factor*.
Now, under covariates add a new covariate, enter painexp, and choose
interactions with factor 1, and centering as factor 1 mean.(*2)
When you run this model, you should hopefully see a design matrix with
two column of zeros and ones which pick out your two groups, and two
columns which have the covariate values for each group (and zeros
elsewhere). These four columns are a1 a2 b1 b2 from my equation above.
The contrast you are then interested in is likely to be one of:
b2 > b1 : t-contrast [0 0 -1 1]
b2 < b1 : t-contrast [0 0 1 -1]
b2 ~= b1 : f-contrast [0 0 -1 1] (actually either of above)
Incidentally, I don't think you're being stupid -- I think
covariate-by-factor interactions are a confusing issue, and I also
think the factorial design specification in SPM5 is very powerful (and
hence very complex) and not yet well-documented enough. If I find
time, I might try to put something onto the Wiki(*3) but for the
meanwhile, I hope this helps.
Best,
Ged.
P.S. Footnotes:
*1(Volkmar, please correct me if wrong -- I don't really understand
what the first column "for replication" is doing in these factorial
designs; can I just put group as the first factor, and then
ones(N1+N2,3) for the unused factors?)
*2(I may be wrong here, I'm still learning about cov-by-factor
interactions and centering. My suggestion here comes straight from the
SPM5 help -- that "centering" should match "interaction with".)
*3(or maybe Volkmar/Will Penny/someone else already has plans to write
more docs for this?)
> unfortunately I have no experience in creating such a factorial model in
> spm. therefore your explanation is a bit to difficult for me. if i could
> ask you to have a quick look about what i did?
> i tried the option “flexible factorial” at the specify the 2nd level.
> I created 2 factors (the first “subject” with all the defaults, the
> second “painscore” as my covariate I want to test between the two groups).
> At “specifiy subjects or all scans & factors” I gave in “specify all”,
> specified the con-files of each of my participants (patients and controls)
> and specified the factor matrix as “1 1 1 … 2 2 2 ….” (which is 1 for
> patients and 2 for controls in the way I entered the scans, which I think
> is totally wrong as a factor matrix).
> now at this step I am absolutely to stupid to go ahead. I dont know what to
> specify at “main effects & interactions”. I also wonder where I should give
> information about the values of my second factor “painscore”. probably I
> was wrong at this step too?
>
> I am sorry to bother you and the spm list with such dumb questions. but I
> am not familiar with matrices (even after reading “the joy of matrix maths”
> which was no joy) and only rarely familiar with the mathematical background
> of factorial designs.
> if my question is to stupid, is there an “understandable” paper about
> factorial designs in spm?
>
> thanks in advance
>
> markus
>
|