On Tue, Aug 7, 2012 at 2:06 PM, Maren Strenziok
<[log in to unmask]> wrote:
> Hi Donald,
>
>
> thanks. Just to confirm, you are saying that I need to do the following
> for...
>
> effect of time: subtract time 1 from time 2 and run a two samples
> independent t-test on the difference maps
>

If you use the contrast of 1/2 1/2 you will get the effect of time, if
you use 1 -1, you will get the group*time interaction since the input
files are the difference in time. The effect of time could also be
estimated with a one-sample t-test.



> effect of group: average time 1 and time 2 and run a two samples independent
> t-test on the difference maps

Correct. The contrast would be 1 -1. This tells you where the task is
different in the two groups.

>
>
> I don't understand what you suggest for the interaction effect? Can you
> please specify?

See above.

>
>
> Thanks, Maren
>
>
>
>
>
>
> On Tue, Aug 7, 2012 at 11:32 AM, MCLAREN, Donald <[log in to unmask]>
> wrote:
>>
>> On Tue, Aug 7, 2012 at 7:27 AM, Maren Strenziok
>> <[log in to unmask]> wrote:
>> > Hi Donald,
>> >
>> > thanks for you reply. What you describe gets me to the effect of time.
>> > What
>> > do you suggest to get to the effect of group (and interaction effect) if
>> > I
>> > have unequal numbers of subjects in each group?
>>
>> For the interaction, use a two-sample t-test.
>>
>> For the group effect, you need to average time1 and time2 instead of
>> subtracting them.
>>
>> >
>> > Maren
>> >
>> >
>> > On Fri, May 25, 2012 at 11:54 AM, MCLAREN, Donald
>> > <[log in to unmask]>
>> > wrote:
>> >>
>> >> On Fri, May 25, 2012 at 10:24 AM, Maren Strenziok
>> >> <[log in to unmask]> wrote:
>> >>>
>> >>> Hi Tom,
>> >>>
>> >>> after you laid out the 2x2 ANOVA design in a post a few weeks ago, I
>> >>> tried it out and wonder what your opinion is on the following
>> >>> situation. I
>> >>> have scans from two groups (trained, non-trained) and two time points
>> >>> (pre-training, post-training). Although each subject has a pre- and a
>> >>> post-training scan, I have unequal numbers of subjects in my groups
>> >>> (10,
>> >>> 12). So when I prepare the difference map required to look at the
>> >>> effect of
>> >>> group, following your logic, I would first add the map of group A at
>> >>> time
>> >>> point 1 and group A at time point 2, then add map B at time 1 and map
>> >>> B at
>> >>> time 2. If I do that I get two files with different volume numbers as
>> >>> group
>> >>> A has only 10 subjects and group B has 12. When I now subtract these
>> >>> maps
>> >>> from each other, the result is a difference map that consists of 10
>> >>> volumes.
>> >>
>> >>
>> >> I'm not quite sure I follow what you did, but what you should end up
>> >> with
>> >> after subtracting time1 from time2 is 22 volumes (1 for each subject).
>> >> Essentially, for each subject, you subtract the two time points and get
>> >> a
>> >> difference image. Then you put those difference images into a
>> >> two-sample
>> >> t-test (10 in one group, 12 in the other group).
>> >>
>> >>
>> >>>
>> >>> Which part of my data was omitted here? The last 2 subjects of group
>> >>> two?
>> >>> Is this good practice or should I try something else? Also, to
>> >>> complicate
>> >>> things, I have DTI data that I submitted to the tbss processing stream
>> >>> in
>> >>> FSL. So in addition to a map with data, I also need a mask file which
>> >>> I
>> >>> would make from the pooled group I guess. Can you let me know whether
>> >>> the
>> >>> FEAT GUI to set up the design and matrix files is the better option in
>> >>> my
>> >>> case?
>> >>>
>> >>> Any comments are highly appreciated.
>> >>>
>> >>> Maren
>> >>>
>> >>> On Tue, May 15, 2012 at 6:00 PM, Thomas Nichols
>> >>> <[log in to unmask]> wrote:
>> >>>>
>> >>>> Dear Ciara,
>> >>>>
>> >>>> Repeated measures, if you read the list, is quite tricky.  Really the
>> >>>> best thing is to eliminate any repeated measures and run separate
>> >>>> analyses
>> >>>> for each contrast of interest.  I.e. If the design is
>> >>>>
>> >>>>   A1 A2 B1 B2
>> >>>>
>> >>>> then compute the effects for each subject with fslmaths:
>> >>>>
>> >>>>   Main effect of A vs B:  (A1+A2)-(B1+B2)
>> >>>>   Main effect of 1 vs 2:  (A1+B1)-(A2+B2)
>> >>>>   Interaction:  (A1-A2)-(B1-B2)
>> >>>>
>> >>>> Then you'll have reduced your data to a set of 24 images which you
>> >>>> can
>> >>>> submit to a one-sample t-test.
>> >>>>
>> >>>> Does this help?
>> >>>>
>> >>>> -Tom
>> >>>>
>> >>>>
>> >>>> On Tue, May 15, 2012 at 7:16 PM, Ciara Greene
>> >>>> <[log in to unmask]>
>> >>>> wrote:
>> >>>>>
>> >>>>> Hi FSL users,
>> >>>>>
>> >>>>> I'm a bit confused as to how to go about implementing a 2x2 repeated
>> >>>>> measures design in randomise.
>> >>>>>
>> >>>>> I have resting state data from 24 subjects under 4 conditions - 2
>> >>>>> factors, each with 2 levels - each of which was acquired in a
>> >>>>> separate run.
>> >>>>> I have conducted concat ICA on this data and now want to compare ICs
>> >>>>> across
>> >>>>> the different scanning conditions using dual regression. I
>> >>>>> understand that
>> >>>>> implementing repeated measures designs in randomise can be tricky;
>> >>>>> is there
>> >>>>> a way to extend the 1 factor/4 levels design described at
>> >>>>> http://www.fmrib.ox.ac.uk/fsl/randomise/index.html#Ex:RepeatMeas to
>> >>>>> examine
>> >>>>> main effects of factor 1 and factor 2 and their interaction?
>> >>>>>
>> >>>>> Thanks for your help,
>> >>>>> Ciara
>> >>>>
>> >>>>
>> >>>>
>> >>>>
>> >>>> --
>> >>>> __________________________________________________________
>> >>>> Thomas Nichols, PhD
>> >>>> Principal Research Fellow, Head of Neuroimaging Statistics
>> >>>> Department of Statistics & Warwick Manufacturing Group
>> >>>> University of Warwick, Coventry  CV4 7AL, United Kingdom
>> >>>>
>> >>>> Web: http://go.warwick.ac.uk/tenichols
>> >>>> Email: [log in to unmask]
>> >>>> Phone, Stats: +44 24761 51086, WMG: +44 24761 50752
>> >>>> Fax:  +44 24 7652 4532
>> >>>>
>> >>>>
>> >>>
>> >>
>> >
>
>