Hi again Gwenaelle
Happy to do a multivariate analysis - what's the best way to approach this?
Rgds
Mark
-----Original Message-----
From: FSL - FMRIB's Software Library [mailto:[log in to unmask]] On Behalf Of Gwenaëlle DOUAUD
Sent: Wednesday, 14 April, 2010 12:25 AM
To: [log in to unmask]
Subject: Re: [FSL] FSLVBM GLM Setup
Hi Mark,
I'd be happy with just the three two-way comparisons :-).
However if you have the possibility, as univariate test between illness 1 and 2 does not yield significant results despite you having the feeling that illness2 is more severe, you might want to do a multivariate (type SVM) analysis on the processed GM images to maybe increase your sensitivity and obtain some way of distinguishing illness 1 from 2...
Cheers,
Gwenaelle
--- En date de : Mar 13.4.10, Mark Walterfang <[log in to unmask]> a écrit :
> De: Mark Walterfang <[log in to unmask]>
> Objet: Re: [FSL] FSLVBM GLM Setup
> À: [log in to unmask]
> Date: Mardi 13 avril 2010, 13h58
> Hi again Gwenaëlle
>
> Many thanks, that's most helpful. Just finally - if you were a
> reviewer and saw just the three two-way comparisons, would you ask for
> the three-group analysis and post-hoc t-tests, or would you be happy
> with the three two-way analyses?
>
> Rgds
>
> Mark
>
>
> On 13/4/10 9:01 PM, "Gwenaëlle DOUAUD" <[log in to unmask]>
> wrote:
>
> > Hi Mark,
>
>
> > Thanks for your response - I guess I am interested in
> the
> >
> > result of an
> > F-test across the three samples so it sounds like it
> will
> > be
> > worth a try.
>
> Sure, but bear in mind that it will only tell you where are the
> > significant changes across the 3 groups, so you will
> still need to run the
> > post-hoc t-tests on each pair of groups to determine
> what's "driving" these
> > results.
>
> > With regards to what my question is: it's partially answered by my
> > results from the three 2-way analyses. I see that illness1 vs
> > controls shows some key reductions; illness2 vs controls shows more
> widespread
> >
> > reductions and in
> > larger clusters; illness1 vs illness2 shows no differences in either
> > direction. I'm interested in whether illness1 vs
> > illness2
> > really differ, as
> > the separate comparisons against controls implies that they should,
> > but a direct comparison between them suggests that they don't.
>
> Yes, it can happen, this means that though illness2 seems
> > more severe, this is not a significant effect.
>
> > (My numbers are
> > 20-30 in
> > each group). Does method II assist in this?
>
> There is no way of knowing for
> > sure, and if it does when you'll do the post-hoc
> t-test between illness1 and
> > 2, this would be just because you have increased your
> DoFs, not because you
> > would have asked a different question...
>
> Cheers,
> Gwenaelle
>
>
>
>
> > Rgds
> >
> >
> > Mark
> >
> >
> > On 13/4/10 4:20 AM, "Gwenaëlle DOUAUD"
> > <[log in to unmask]>
> > wrote:
> >
> > > Hi Mark and Jay,
> >
> > there is no
> > good answer to this question I'm afraid.
> >
> > Say
> > > you've got two subjects
> > in group A, 3 in B and 4 in C,
> > then both approaches
> > > are valid:
> >
> >
> > Method I
> >
> > A B
> >
> > 1 0
> > 1 0
> > 0 1
> > 0 1
> > 0 1
> > for the design.mat of the
> > first
> > > A and B groups with
> >
> > 1 -1 (A-B)
> > -1 1 (B-A)
> > for the
> > design.con
> > and then repeat
> > > for groups B and C, then groups A and C
> > (which is what
> > you did Mark).
> >
> > Method
> > > II
> >
> > A B C
> >
> > 1 0 0
> > 1 0
> > 0
> > 0 1 0
> > 0 1 0
> > 0 1 0
> > 0 0 1
> > 0 0 1
> > 0 0 1
> > 0 0 1
> > for the
> > >
> > design.mat of the 3 groups with
> >
> > 1 -1 0 (A-B)
> > -1 1 0 (B-A)
> > 0 1 -1
> > (B-C)
> > 0 -1 1
> > > (C-B)
> > 1 0 -1 (A-C)
> > -1 0 1 (C-A)
> > for the design.con
> > (t-tests)
> > and
> >
> > 1 0 1 0 0
> > > 0
> > for the design.fts (F-test, as many
> > columns as there are
> > rows in your
> > > design.con, you just need to click in
> > "F-tests" in the
> > Glm gui and then click
> > > in front of the two relevant
> > "Contrasts" you have
> > already set up)
> >
> > So with
> > > Method II, you can
> > also ask the question of where are
> > the changes *across the
> > > 3 groups*
> > (F-test with the design.fts). You also get an increase in DoF but,
> > > as
> > Tom Nichols said, if it happens that group C for instance has wildly
> > >
> > smaller variance, you can get inflated significances (or reduced
> > power if it
> > > has wildly larger variance).
> >
> > So it depends on what your main
> > question is,
> > > really.
> >
> > Hope this helps,
> > Gwenaelle
> >
> >
> > --- En date
> > de : Lun 12.4.10, Mark
> > > Walterfang <[log in to unmask]>
> > a
> > écrit :
> >
> > > De: Mark Walterfang
> > > <[log in to unmask]>
> > > Objet:
> > Re: [FSL] FSLVBM GLM Setup
> > > À:
> > > [log in to unmask]
> > > Date: Lundi 12
> > avril 2010, 14h18
> > > Hi all
> > >
> > > I'm in the
> > > same situation as Jay. I
> > have three groups
> > > (illness1, illness2
> > > and
> > > controls), all matched
> > to each other. I've run three
> > > two-way analyses,
> > >
> > > which is pretty
> > laborious and I'm pretty sure it's
> > not
> > > statistically
> > > ideal.
> > > What
> > I can't work out is how to set up the design matrices
> > > &
> > > contrasts
> > in
> > > the way Jay describes, as the online manual for
> > Randomise
> > >
> > >
> > doesn't really
> > > provide guidance here. Gwenaëlle, is this
> something
> > you
> >
> > > can
> > > advise on?
> > >
> > > Thanks in advance,
> > >
> > > Mark Walterfang
> > >
> >
> > >
> > > On 10/4/10
> > > 12:29 PM, "Jay Ives" <[log in to unmask]>
> > > wrote:
> > >
> >
> > > > I have 70 subjects in
> > > 4 groups and would like to test
> > > between
> > individual
> > > > groups and
> > > combinations of the groups. Can someone
> > >
> > please advise me how to set
> > > > up
> > > the design.mat and design.con files
> > to do this?
> > > Thx
> > >
> > >
> > > WARNING: This
> > > message
> > > > originated
> > from outside the
> > Northern/Melbourne/Western
> > > Health
> > > e-mail network.
> >
> > > > The sender cannot be validated. Caution is
> > advised.
> > >
> > > Contact IT
> > Services (+61 3
> > > > ) 9342 8888 for more information.
> > >
> >
> >
> >
> > >
> >
> >
> >
> > WARNING: This message originated from outside the
> > >
> > Northern/Melbourne/Western Health e-mail network. The sender cannot
> > be
> > >
> > validated. Caution is advised. Contact IT Services
> > (+61 3 ) 9342 8888 for
> > more
> > > information.
> >
> >
> > Dr Mark Walterfang
> > Consultant
> > Neuropsychiatrist
> > Neuropsychiatry Unit
> > Level 2, John Cade Building
> > ROYAL
> > MELBOURNE HOSPITAL 3050 AUSTRALIA
> > T +61-3-93428750
> > F +61-3-93428483
> > E
> > [log in to unmask]
> > W www.neuropsychiatry.org.au
> >
> > Research
> > Fellow
> > Melbourne Neuropsychiatry Centre
> > University of Melbourne
> > Level 2
> > & 3, Allan Gilbert Building
> > 161 Barry St
> > CARLTON SOUTH 3023 AUSTRALIA
> > T
> > +61-3-83441800
> > F +61-3-93480469
> > E [log in to unmask]
> > W
> > www.psychiatry.unimelb.edu.au/mnc
> >
>
>
>
>
>
> WARNING: This message originated
> > from outside the Northern/Melbourne/Western Health
> e-mail network. The sender
> > cannot be validated. Caution is advised. Contact IT
> Services (+61 3 ) 9342
> > 8888 for more information.
>
>
> --
> Dr Mark Walterfang
> Consultant Neuropsychiatrist
> Neuropsychiatry Unit
> Level 2, John Cade Building
> ROYAL MELBOURNE HOSPITAL 3050 AUSTRALIA T +61-3-93428750 F
> +61-3-93428483 E [log in to unmask] W
> www.neuropsychiatry.org.au
> --
>
WARNING: This message originated from outside the Northern/Melbourne/Western Health e-mail network. The sender cannot be validated. Caution is advised. Contact IT Services (+61 3 ) 9342 8888 for more information.
|