Hi,
> >> I have questions about the high-level analysis, particular in whether
> >> or not including group mean. For the simplest case, if we have two
> >> condition A B in one scan session. we can either define low level
> >> contrast, A>0, and B>0 (suppose 0 is the implicit baseline), and feed
> >> them up to the pair-T test (the paired-T example), or define contrast
> >> A-B and feed them up to the simple T-test (group mean). In the first
> >> case, the model will ask to include subject-mean, while for the later
> >> case, it's not necessary. Am I right to say that both methods will
> >> give the same results?
> >
> >It is correct that doing a paired t-test after feeding A and B up to
> >second level (according to the paired example in the manual) is exactly
> >the same as doing the A-B contrast at first level and feeding that into a
> >second-level group mean test. With the latter approach, the second-level
> >model is a single EV of all 1s - this "group mean" gives you the output
> >you want; you have already thrown away the true within-subject mean effect
> >at first level by taking the contrast at first level.
>
> Do you mean we should get exactly the same results whether we throw away
> the within-subject mean effects, or include them? Or, we can only throw
> it at the first-level analysis, but not at the second-level analysis.
Whether you "throw away" the mean by subtracting the first-level results
before feeding up, or model them out at second-level (and ignore those
EVs) is the same, yes.
> >> The third way, though a little bit unusual, is to define A, -A, B,
> >> and -B at the first level, then do second level group average like (A+
> >> (-B)) and (B+(-A)). We will get a cope for each subject and then feed
> >> them up to the even higher model using simple T-test group mean (the
> >> multiple-subject and multiple-session example). will this still give
> >> the same result?
> >>
> >> if the third way is correct, I think we can do something different in
> >> the pair-T model, like:
> >>
> >> input EV1 EV2
> >> sub1 cope1 1 0
> >> sub1 cope2 -1 0
> >> sub2 cope1 0 1
> >> sub2 cope2 0 -1
> >
> >You don't need this, no - the first two approaches in the first paragraph
> >are both equivalent and both fine.
>
> I know this is not necessary in this case. But theoretically it is
> correct, right? So this goes back to the question as to under which
> condition should we include individual mean?
Well, no the example you're giving above looks like you're including the
first-level conditions which include the mean effects, but not modelling
them at second level, so that's not the same as the paired t-test - the
second-level variance will be inflated and your results less sensitive.
> >> Now move to the more complicate case (the true multiple-session
> >> case), in which I have to define contrast across sessions (e.g.,
> >> training effect). Say if I have 2 EVs from each session, and I want
> >> to test the several learning effect by using appropriate combination
> >> of these contrast. To make things easier, can I do sometime like this:
> >>
> >> input EV1 EV2 EV3 EV4
> >> sub1 cope1 1 1 0 0
> >> sub1 cope2 1 -1 0 0
> >> sub1 cope3 -1 1 0 0
> >> sub1 cope4 -1 -1 0 0
> >> sub2 cope1 0 0 1 1
> >> sub2 cope2 0 0 1 -1
> >> sub2 cope3 0 0 -1 1
> >> sub2 cope4 0 0 -1 -1
> >> ........
> >>
> >> I input the EV1, EV3 .... for the one group analysis, and EV2, EV4
> >> for other. Is this correct? So the key point here is that whether
> >> this type of model only accept simple averaging contrast, like (1 1 1
> >> 1), but not comparative contrast, like 1 1 -1 -1 in the above case?
> >> I would assume that this is not the case since we can simply cheat
> >> the program by define negative contrast at the first level, like the
> >> third way aforementioned.
> >>
> >> another issue in this case is the rank efficiency. seems I am not
> >> able to include all the following contrast, like 1 -1 -1 1, 1 1 -1
> >> -1, 1 -1 0 0, 0 0 1 -1? but the later two is important for the
> >> simple effect. does that mean I have to setup another model?
> >
> >I'm not sure this is right....for example you don't seem to be modelling
> >out the within-subject mean correctly. Anyway - maybe the answers above
> >may be enough to help work out this section too - if not feel free to get
> >back to us.
>
> The within-subject mean confused me a lot :(. Also, there is a practical
> issue here. as mentioned above, in my study, I have 20 subjects each has
> 20 conditions (copes). it seems rather complex to include them in one
> model. In SPM, one can define cross-session contrasts at the first level
> analysis, and then feed them up for simple group mean. In FSL, can we do
> something similar, say at the second level, we do the cross-session
> contrasts for each subjects separately, then feed them up for
> third-level analysis. if so, how can we do this?
Sure - you could do a separate cross-session contrast at second-level for
each subject using fixed-effects option and then feed the results up to
third level.
Cheers.
>
> Thanks a again for your help.
> Gui
>
>
>
>
>
>
> ==========================
> 263电子邮件－信赖邮自专业
--
Stephen M. Smith, Professor of Biomedical Engineering
Associate Director, Oxford University FMRIB Centre
FMRIB, John Radcliffe Hospital, Headington, Oxford OX3 9DU, UK
+44 (0) 1865 222726 (fax 222717)
[log in to unmask] http://www.fmrib.ox.ac.uk/~steve
|
