On Wed, 15 Oct 2008 16:09:32 +0100, Ronan Joseph Kelly <[log in to unmask]>
wrote:
>Dear FSL-users
Caveat: I typically use SPM, though I'm starting to play with some of the FSL
tools right now. So I can only make general comments. In particular, I've
never set up a statistical analysis (regression, ANOVA, etc) in FSL.
>I am conducting a VBM analysis on a longitudinal study involving the effect
>of different diets on brain volume. Participants were scanned at 3 different
>time-points; T1, T2, T3. There ended up being 3 subgroups at each
>time-point; one group were control subjects with a regular diet throughout
>the entire scanning period T1 - T3, the second group was on a regular diet
>between T1 and T2 and on the special diet between T2 and T3
(labelled 'Short
>term'), while the third group was on the special diet between T1 and T3 (ie
>'Long term'). So in all, I have 9 (3 time-points by 3 diet levels) different
>groups of interest.
From an ANOVA standpoint: I'm not an expert, but it seems to me that you
have two factors: group (3 levels) and timepoints (3 levels). The important
point being that group is a between-subject factor, and timepoint is a within-
subjects factor. Which makes this a repeated-measures design.
>My questions involve using FSL-VBM on a longitudinal study of this nature.
>
>My initial idea was to simply compare across diet level (ie Controls VS
>'Short term' special diet VS 'Long term' special diet etc). However, I began
>to consider whether it was correct to include multiple scans of participants
>at different time points in one individual EV. If I was to pursue this
>approach I would have 3 different scans for each participant in each group.
>Is that allowed? How should I carry this out?
You'd have to set things up in the context of a 2-factor repeated measures
ANOVA.
Of course, you also need to think about what you want to "get out" of the
timepoint factor. If you're not interested in time (unlikely), you could just
average the three scans for each subject and put that into a "normal" (non-
repeated measures) ANOVA. My guess is that you'd be interested in things
like group X time interactions.
>My other idea was to divide all my data into the 9 different groups, and to
>set up the contrasts in my Glm model using linear combinations of these
>subgroups. For example, a contrast exploring the effect of diet between
>controls and 'long term' particpants would have to compare all 3 subgroups
>of the controls, (minus) all 3 subgroups of the 'long term' subjects. Again,
>is this the correct method for carrying out such an analysis?
The problem with that is that you're ignoring subject effects, which can result
in the statistical test losing power. The canonical example of this is when you
have data which should be analyzed using a paired-t test, but you ignore the
pairing.
>Splitting up
>the subjects into subgroups like this would enable me to carry out other
>interesting contrasts using this linear combination method, but how
>practical is it? And once I reach the stage of setting up my contrasts, what
>coefficients would I use when for a simple contrast I would simply use +1
>and -1?
>
>I know that was long-winded but I'd appreciate any suggestions people had
to
>offer!
If you don't get answers here, I would post the question to one of the stats
USENET groups, since your real questions are not specific to neuoroimaging.
You could try sci.stat.math (google groups interface at
http://groups.google.com/group/sci.stat.math/topics) or sci.stat.consult
(http://groups.google.com/group/sci.stat.consult/topics).
Best,
S
>
>Cheers,
>Ronan
>
>--
>Ronan Kelly
>Neuroinflammatory Research Group
>Trinity College Institute of Neuroscience
>University of Dublin, Trinity College
>Dublin 2
>Ireland
>
|