Thank you for your response.
I am not sure that I know what is meant by "demean". We did use the -D
option, but it seems you are suggesting we do something additional with the
covariate in the design matrix. Can you please describe what we should do?
Thank you very much,
Bhargav
Gwenaëlle DOUAUD <[log in to unmask]> wrote:
Hi Matt,
it seems that you have not demeaned your ev in your
stats.mat. You should do so and run exactly the same
command you've used without forgetting the -D option.
Cheers,
Gwenaelle
--- Matt Glasser a écrit :
> Tom and others,
>
>
>
> I am helping Bhargav with this project. We tried
> using the -D option,
> however now all of our t statistics are very small
> (between -1 and 1, with
> most very close to zero). This doesn't make sense,
> however, as we would
> expect at least some voxels to show t statistics
> greater than 1, just by
> chance. Our randomize commandline, our design
> matrix and our contrast file
> follow this message. We have verified that our 4D
> FA skeleton file and mean
> skeleton mask files are okay. Does anyone know why
> we are getting such low
> t values? Would it be possible to get r values
> instead using something
> other than randomise?
>
>
>
> Thanks,
>
>
>
> Matt.
>
>
>
> randomise commandline:
>
> randomise -i all_FA_skeletonised -o tbss -m
> mean_FA_skeleton_mask -d
> stats.mat -t stats.con -n 5000 -c 3 -D
>
>
>
> stats.mat:
>
>
>
> /NumWaves 1
>
> /NumPoints 10
>
> /PPheights 7.800000e+02
>
>
>
> /Matrix
>
> 6.700000e+02
>
> 5.400000e+02
>
> 5.300000e+02
>
> 6.000000e+02
>
> 5.800000e+02
>
> 7.800000e+02
>
> 7.500000e+02
>
> 6.100000e+02
>
> 7.000000e+02
>
> 7.200000e+02
>
>
>
> stats.con:
>
>
>
> /ContrastName1 Positive
>
> /ContrastName2 Negative
>
> /NumWaves 1
>
> /NumContrasts 2
>
> /PPheights 7.800000e+02
> 7.800000e+02
>
> /RequiredEffect 3.726 3.726
>
>
>
> /Matrix
>
> 1.000000e+00
>
> -1.000000e+00
>
>
>
> _____
>
> From: FSL - FMRIB's Software Library
> [mailto:[log in to unmask]] On Behalf
> Of Thomas Nichols
> Sent: Wednesday, October 10, 2007 5:10 AM
> To: [log in to unmask]
> Subject: Re: [FSL] TBSS-voxelwise correlations
> between FA and Test scores
>
>
>
> Bhargav,
>
> Did you include an intercept in the model? Or
> specify the -D option and use
> a centered covariate? That might well explain the
> large t values (without
> an intercept you may be testing if the FA values are
> non-zero, not a very
> interesting question).
>
> randomise uses the general linear model and
> contrasts to test effects of
> interest with t-tests and so doesn't produce
> correlation coefficients.
>
> You can threshold on the basis of corrected or
> uncorrected P-values with
> fslmaths/avwmaths, the -thr option, and the
> corresponding One-Minus-P-value
> images; see the randomise help page
> http://www.fmrib.ox.ac.uk/fsl/randomise/ for the
> exact filename extensions
> for each of these types of images.
>
> Hope this helps!
>
> -Tom
>
> On 10/9/07, Bhargav Kumar Errangi
> wrote:
>
> Hello,
>
> We are using TBSS to look for voxel-wise
> correlations between FA and test
> scores. Could you please answer the following
> questions for us?
>
> 1) Using the GLM tool, we specificy 1 covariate and
> provide values for each
> of 10 subjects in the EV tab. Then in the contrast
> tab, we specify 2
> contrasts (1 and -1)
> 2) The output is a t statistic image. We were
> exepcting an r statistic map.
> Does FSL calculate the t from the r?
> 3) Will the "1" contrast yield positive correlations
> and the "-1" contrast
> yield negative correlations?
> 4) The t statistic values are very, very large. We
> only get reasonbable maps
> when we threshold at t>20. Do you know why this
> might be?
> 5) Is there a way to threshold the map based on
> p-values rather than t
> statistics?
>
>
>
> ____________________________________________
> Thomas Nichols, PhD
> Director, Modelling & Genetics
> GlaxoSmithKline Clinical Imaging Centre
>
> Senior Research Fellow
> Oxford University FMRIB Centre
|