Hi Michael,
> De: Michael Scheel <[log in to unmask]>
> Objet: [FSL] questions about covariate and demeaning in randomise
> À: [log in to unmask]
> Date: Lundi 14 mars 2011, 11h53
> dear fsl experts,
>
> I'm conducting a tbss analysis starting with a simple FA
> comparison between two groups (controls and patients). I
> want to include age as a covariate.
> I've perused the online documentation and the mailing
> archives but I'm now unfortunately more confused than
> before.
I have to say I am not surprised...
> Attached are my design files. I would be very happy if
> someone could take a look at it and tell me if they are ok.
You need to demean the age within group (maybe this is what you've done but the rounding is indeed not quite good enough)...
> Contrast 1 and 2 are the group comparisons for FA
> differences
> In Contrast 3 and 4 I want to ask where is the slope of
> Age-vs-FA different between groups
> In Contrast 5 and 6 I want to ask for where FA negatively
> correlates with age for each group seperately.
Yep.
> I have some additional questions:
> 1) As far as I understood one should add for each group an
> extra EV into the design matrix (age controls, age patients)
> to allow for
> group specific correlations of age and FA, for example.
> If I would only add one extra EV age for all subjects I
> would assume that the age-FA-dependence is linear and
> the same across groups.
> However what might be the advantage of such an approach -
> would it be more sensitive to differences?
It completely depends on whether there is an interaction indeed between group and age, in which case you are modelling your data better by doing it this way... You can always check if it's the case by precisely looking at contrasts 3 and 4.
> 2) Just to clarify. The -D option in randomise does not
> replace the demeaning of the covariate in the design matrix.
> it only takes care of the data itself. Correct?
Correct.
> 3) The randomise error "Warning: You have demeaned your
> data, but at least one design column has non-zero mean"
> appears in my case although age was properly demeaned.
> After demeaning there will be often the situation due to
> rounding that the mean of the demeaned data will not exactly
> be 0.
Correct.
> Am I correct that randomise will give this warning even if
> the mean of the demeaned data is something small like
> 2.368672e-16?
Not sure how low it needs to be.
> In my case the demeaned ages have a mean of -0.2 - is that
> good enough?
Maybe not, especially this is up to -0.6 in the control group... One more decimal should do.
> 4) If I want to manually create my design files I
> understand /NumWaves, /NumPoints and /Matrix.
> However what are /PPheights for. What do I need to look
> after or be careful with here.
Nothing, this is not necessary for randomise.
Cheers,
Gwenaelle
--------------------------------------------------------------------
Gwenaëlle Douaud, PhD
FMRIB Centre, University of Oxford
John Radcliffe Hospital, Headington OX3 9DU Oxford UK
Tel: +44 (0) 1865 222 523 Fax: +44 (0) 1865 222 717
www.fmrib.ox.ac.uk/~douaud
--------------------------------------------------------------------
|