Hi, part of your problem is that you are modelling the means in two separate places; if you look at
the paired t-test (which is where you want to start with this design) the pre-post difference is
achieved with a 1, -1 etc modelling, as the means are taken out with subject-specific EVs.
So - what you want here is a slight addition to the paired t-test, allowing separate modelling of
the pre-post difference, averaged across subjects, but separately for the two groups.
Hence start with
http://www.fmrib.ox.ac.uk/fsl/feat5/detail.html#PairedTwoGroupDifference
And you need to duplicate EV1 to a separate EV. Each should be full of 1 and -1 values for one of
the groups, and zeros for the other group, and then the other way round for the other group.
Then set a contrast between these two EVs to get the group difference.
Cheers, Steve.
On Sat, 20 Jan 2007 00:03:51 +0000, Jenny Robinson <[log in to unmask]> wrote:
>Dear FSL Gurus,
>I have run into a bit of a pickel and am hoping that you can help! I am
>currently trying to run a higher level analysis on a pre-post treatment
>study, in which we have two groups (a placebo group, and a treatment
>group). All subjects performed the same task prior to intervention (time
>1), and then post intervention (time 2). What I would like to do is look
>at the differences in BOLD signal change between the groups (so, looking
>for significant changes in the [task time 2 - task time 1] difference
>i.e., a group*task interaction). I have consulted the webpage
>(http://www.fmrib.ox.ac.uk/fsl/feat5/index.html) and these archives as to
>how to set up the analysis, but it doesn't seem clear that any of the
>examples given would accomplish what I'm looking to do. How would you
>recommend setting up the design matrix?
>
>I have tried numerous things, but I'm either getting warnings. I'm
>clearly doing something wrong.
>For example...
>SUBJ-TIME GROUP EV(GROUP1) EV(GROUP2) EV(TIME1) EV(TIME2)
>Subjx-1 1 1 0 1 0
>Subjx-2 1 1 0 0 1
>Subjy-1 2 0 1 1 0
>Subjy-2 2 0 1 0 1
>Example Contrasts:
>time2-time1 1 -1 -1 1
>Something like this is clearly flawed, and won't run at all. The warning
>that at least one EV is (close to) a linear combination of the others
>comes up, and it tells me I should alter my design.
>
>Alternatively, I ran something like this as well:
>SUBJ-TIME GROUP EV(GROUP1T1) EV(GROUP1T2) EV(GROUP2T1) EV(GROUP2T2)
>Subjx-1 1 1 0 0 0
>Subjx-2 1 0 1 0 0
>Subjy-1 2 0 0 1 0
>Subjy-2 2 0 0 0 1
>Example Contrasts:
>Group1(Time2-Time1) -1 1 0 0
>Group2(Time2-Time1) 0 0 -1 1
>And then ran an F-test on them.
>There has got to be a more statistically sound way.
>
>Your explanation of a two-sample paired t-test only uses one group that
>performed the task (or condition, if you will) twice. Here, we have 2
>groups, and I'm not sure how to account for that without FSL yelling at me.
>
>Any insight?
>
>Thanks in advance for your time and genius :)
>Kind regards,
>Jenny
>===========================================================
=============
|