Thank you so much for your thorough advice.  The methods are much clearer now.

Take care,
Laura


________________________________
From: Michael Harms <[log in to unmask]>
Reply-To: FSL - FMRIB's Software Library <[log in to unmask]>
Date: Tue, 10 Nov 2009 17:36:52 -0500
To: <[log in to unmask]>
Conversation: [FSL] simple correlation
Subject: Re: [FSL] simple correlation

Hi Laura,
A correction to what I just wrote:
While you would get the same estimates for the betas (EVs) themselves
with or without the controls in your "separate slopes" model, I believe
I was wrong in saying that the residual error variance would not be
affected.  All the subjects that you include will contribute to the
residual variance estimate, and thus to the estimated standard error of
the EV, and thus to the t-value associated with your contrasts.  So, if
you really want to run two separate analyses (one for each group), then
you should create a separate 4D input file for each group.

If I (still) don't have that right, someone on the list please correct
me.

cheers,
Mike H.


On Tue, 2009-11-10 at 15:55 -0600, Michael Harms wrote:
> In principle, you could leave the controls in, since you are not doing
> any statistics that involve the residual error variance. (Or you could
> include values for the controls in the design matrix, but just don't
> compute any contrasts that involve the controls' EVs).  Basically, for
> what you are trying to do, you should get identical results running
> completely separate analyses for each group (provided you set up
> exchangability groups for 'randomise', see below).  If I were running
> the analyses, I would actually do it both ways, and confirm that was
> indeed the case.
>
> The contrast:
> 1 0 1 0
> tests whether the linear combination of the intercept and the estimated
> slope is significantly different from zero, which is not (by my
> understanding) what you are interested in testing.
>
> If you want to specifically test whether the correlation (slope) is
> significantly different from zero, then you just want a single 1 or -1
> in the position of the EV for the behavioral measure.
>
> If you do indeed want to run two completely separate analyses, then you
> are (in my opinion) best off splitting your input into two separate
> groups and proceeding as such.  You could in principle conduct such an
> analyses as you have outlined, but in that case you need to set up
> exchangability groups, otherwise the permutations will include swapping
> across groups, which would no longer be equivalent to running two
> completely separate analyses.
>
> Best,
> Mike H.
>
>
> On Tue, 2009-11-10 at 14:58 -0500, Danielian, Laura (NIH/NINDS) [E]
> wrote:
> > Hi Mike,
> >
> > Thank you for the quick reply.  You have suggested 2 options and I want to make sure I understand how each would affect the output of the analysis.
> >
> > >I would exclude the controls entirely from the input data set and the
> > >design matrix, since you are not interested in their correlations.
> >
> > Is it necessary to remove the controls from the .mat file in order to obtain correct results, or will it work as I set it up with the controls padded with zeros?  If they need to be removed, I assume that I would need to extract those images from the 4D all_FA_skeletonised image?
> >
> > >Also, the contrast matrix for testing the correlation should be:
> > >0 0 1 0  (positive correlation for dg1)
> > >0 0 -1 0 (negative correlation for dg1)
> > >0 0 0 1  (positive correlation for dg2)
> > >0 0 0 -1 (negative correlation for dg2)
> >
> > Just to check, this would perform independent correlations for each disease group and not group them together in a single correlation with the behavioral measure.  If so, then I'm not sure I understand how adding the "exchangeability groups" would change the analysis....
> >
> > >Lastly, you might want to consider setting up "exchangability groups",
> > >such that the permutations are limited to within the dg1 and dg2 groups,
> > >Respectively.  (Or, alternatively, since you don't seem to be interested
> > >in any contrasts involving both dg1 and dg2, just run two completely
> > >separate analyses -- one involving the dg1 group, and the other
> > >involving the dg2 group).
> >
> > This is exactly what I want to do (run two completely separate analyses -- one involving the dg1 group, and the other involving the dg2 group).  Based on previous posts on 2 group correlations to a single EV, I thought that the first method would cover that.  How would using an exchangeability group change the analysis?  Is one a comparison of the means and the other a within-subject correlation?  Would you be able to discuss how the statistics are performed in each case?
> >
> > Thanks,
> > Laura
> >
> >
> > ---------------------------------
> >
> > Laura Danielian
> > Biomedical Engineer
> > National Institutes of Health
> > Building 10 CRC Room 7-5753
> > Bethesda, MD 20892
> > 301-496-2168
> >
> >
> > ________________________________
> > From: Michael Harms <[log in to unmask]>
> > Reply-To: FSL - FMRIB's Software Library <[log in to unmask]>
> > Date: Tue, 10 Nov 2009 12:31:49 -0500
> > To: <[log in to unmask]>
> > Conversation: [FSL] simple correlation
> > Subject: Re: [FSL] simple correlation
> >
> >
> >
> > On Tue, 2009-11-10 at 16:59 +0000, Laura Danielian wrote:
> > > Hi,
> > >
> > > I appologize for the simple question, but statistics are not my strength.  I
> > > have run the TBSS analysis and I want to investigate potential correlations
> > > between FA values within each voxel of the FA skeleton and a single
> > > behavioural measure.  I have 3 groups (2 diseases, 1 control) and want to
> > > correlate values within each of the disease groups separately.  I am not
> > > interested in the correlations for the control group.   I have set up the
> > > following design.mat and design.con files (simplified to show only 2 subjects in
> > > each group).  Are they appropriate for what I am trying to do?  Rows 3-4
> > > represent the controls.  The values in EV3 and EV4 were demeaned within the
> > > specific diagnosis.
> > >
> > > EV1(disease group 1)  EV2(disease group 2)  EV3(beh. meas. dg1)  EV4(beh.
> > > meas. dg2)
> > > 1  0  0.26  0
> > > 1  0  -3.4  0
> > > 0  0  0  0
> > > 0  0  0  0
> > > 0  1  0  -4.3
> > > 0  1  0  4.68
> > >
> > > I have set up 4 contrasts to test both positive and negative correlations
> > > within each disease group:
> > >
> > > 1 0 1 0
> > > 1 0 -1 0
> > > 0 1 0 1
> > > 0 1 0 -1
> > >
> > > And will use the following randomise command:
> > >
> > > randomise -i all_FA_skeletonised -o tbss -m mean_FA_skeleton_mask -d
> > > design.mat -t design.con -n 5000 -c 3 -T2 -V -D
> > >
> > > Does this seem appropriate?  Thank you so much for your help!
> > >
> > > Take care,
> > > Laura