Dear Las (?),
>Hi,
>
>I'm a 'greenhorn' concerning SPM and have a few questions. I'd be
>very happy, if someone has some suggestions.
>
>The problem is the following: We have a PET study with two groups, each
>of them with two conditions (alertness and some kind of GonoGo-Test),
>6 Scans per subject and condition. Furthermore there is a
>neuropsychological score for the GonoGo-Test, which shouls be used as a
>covariate. Ok, after preprocessing the data I calculated the mean of the
>realigned, smoothed and normalized PET Scans. With these I used the
>multi-study, conditions and covariates design. Now the questions
>
>1.) I compared the results from the mean data with those obtained
>without avering and using multi-study with repititions design for
>a simple contrast. The results were significantly different.
>Is there an explanation ?
>What is the better design then ????
>
I am not sure I understand what you mean when you say you averaged the
data. Presumably you averaged within condition and subject such that you
get one image per subject and condition. I would certainly expect the
results to look quite different then. Effectively what you have is a fixed
effect (non averaged) v.s. a random effects (averaged) model.
In your case (looking for task-by-group interactions) the random effects
model woul be the better one. However, if you want to look at
group-by-covariate interactions you would be better off doing things
slightly differently (I assume now you get one score for each scan in the
GoNogo condition).
If you set up a fixed effects model (non-averaged) you may then use the
con*.img to do the second level analysis using "basic models".
You should then use a Multi-subj: condxsubj interaction and covariates
model.
>2.) One interesting thing is the interaction of the covariates
>with a simple contrast, let's say (GonoGo-alertness). Are the contrasts
>(0 0 -1 1 1) and (0 0 -1 1 -1) the right ones ? The third and fourth
>number are alertness and GonoGo for one of the two groups. And is
>(-1 1 -1 1 1) describing the overall effect ????
>
Again I am slightly confused. First of all: have you used a single column
to model the covariate? It is likely that a better model would be to model
it per group, or even per subject. With 6 scans per condition you can
afford it.
Secondly: I assume you mean a task-by-covariate interaction. If so, that is
not given by your contrast (0 0 -1 1 1), which will give you the effects of
task in group 2 + the effects of the covariate averaged across the two
groups. Note also that the scale of these effects may be very different.
As far as I understand you have scores only from the GoNogo scans, and not
from the alertness scans, right? If so you cannot assess any
task-by-covariate effects. However you can assess group-by-covariate
interactions, see below for procedure.
>3.) What is the form, in which the covariates have to be given ?
>I have read in the mailbase, that they should be positive and negative??
>Is that right???
>
Here is what I suggest, use a "Multi-subj: condxsubj interaction and
covariates model" with two conditions and 1 covariate. Specify that you
want a covariate-by-subject interaction, and specify "No centering".
Normally you should pick centering by subject mean, but because you have
scores only for one condition you cannot do that. Therefore it now becomes
important that you center your covariates yourself (i.e. get both positive
and negative values). So, when entering the covariates you should calculate
the mean score for each subject, and subtract that number from that
subjects scores. When prompted for the covariate vector you should enter 0
for each alertness scan, and the appropriate corrected (subtracted) value
for each GoNogo scan. Please correct me if anyone can think of a simpler
way.
This will give you one covariate column for each subject, and when you want
to look at the main effect (FxEff) you simply enter a contrast with ones
for all those columns.
A Random effects group-by-subject test is done by creating subject specific
contrasts at the first level thereby creating con*.img images, and then
enter those into a two-sample t-test in the "Basic models" module.
>Thanks
>
Good luck Jesper
>
>
Jesper Andersson
Wellcome Dept. of Cognitive Neurology
12 Queen Square
London WC1N 3BG
phone: 44 171 833 7484
fax: 44 171 813 1420
|