Hi Serge,

the order of the EVs only becomes important when it comes to
orthogonalisation in Feat: EVs can be made orthogonal to other EVs with
lower numbers. When you have partially correlated EVs there is a part
of the overall explained variance which can either be explained by the
one or the other EV. Orthogonalisation is a way of constraining the
model such that the bit of variance that can be explained by both EVs
actually gets associated with one (the one that isn't orthogonalise)
and not the other  EV. That's why the order becomes important.
In your experiment the one EV (EV11) that gets orthogonalised to the
first 10 EVs will surely loose some of the associated explained
variance (i.e. it will loose significance as soon as the noise EVs are
partially corellated with the EV of interest) but the idea is this: in
some areas where EV11 on its own used to be significant, some of that
significance can actually be explained by 'noise' EVs. Wherever EV11
after including EV1-EV10 is still significant, this significance is
there despite the fact that you've modeled some of the artefacts.

This approach aims at reducing the false-positive rate and as such will
make your analysis more conservative. It can, however, also increase
you significance: when EVs 1-10 explain a lot of variance in the the
space orthogonal to EV11 (i.e. variance that EV11 would not have
explained anyways) it will actually result in a decrease of residual
variance which means a potential increase in Z-score. The question
whether Z-score increases or decreases depends on the amount of
variance that EV1-EV10 explain that cannot be explained by EV11 vs what
can be explained by EV11.


hope this all makes sense
ta
christian



On Thursday, May 8, 2003, at 17:08 Europe/London, Rombouts, S.A.R.B.
wrote:

> Hi Christian,
>
> I have tried to use MELODIC to get the noise components, and then use
> these
> noise components as EVs, as you suggested in the mail archive.
> I'm afraid some things are not completely clear to me:
>
> - You suggested to use first the noise regressors in the model, and
> then the
> regressors of interest. I do not understand why the order of EVs would
> make
> a difference?
>
> - I have tried it with 10 noise EVs, and 1 EV of interest (EV11, which
> is a
> simple square wave). EV11 was orthogonalized to the first 10 EVs.
> Further,
> for the 10 noise EVs I turned off convolution, temporal filtering and
> temporal derivative. Then I got my design matrix with a very odd
> looking
> EV11, caused by the orthogonalization (I think): the square wave could
> no
> longer be recognized. The contrast I was interested in was (0 0 0 0 0
> 0 0 0
> 0 0 1) So far so good I hope.
> The results I got however were less significant than when analyzing
> with
> only the square wave (no noise regressors): z-scores dropped (for
> example
> the highest z-score dropped from 14 to 12). Do you think my approach is
> correct? If so, isn't it surprising to see effects of interest become
> less
> significant with 11 Evs, of which 10 explain noise?
>
> Thanks a lot,
> Serge.
>
>
>
> -----Original Message-----
> From: FSL - FMRIB's Software Library [mailto:[log in to unmask]] On
> Behalf
> Of Stephen Smith
> Sent: Monday, May 05, 2003 6:47 PM
> To: [log in to unmask]
> Subject: Re: [FSL] preprocessing in feat and melodic
>
>
> Hi Jack,
>
> On Mon, 5 May 2003, Jack Grinband wrote:
>
>>> ..."intensity normalisation", which also
>>> includes a simple thresholding step as well. you won't be getting
>>> this step in 3a and 3b, so that probably explains things.
>>
>> Is this a thresholding for extreme values?
>
> no, it zeros values below a lowish threshold to remove background
> voxels
> from further calculation.
>
>> I often get ICs that are clearly motion artifacts.  Since melodic
>> performs motion correction before doing ICA, I assume that these
>> components are due to the non- linear effects of  motion.  Presumably,
>> any linear effects would be removed by mcflirt. Is that right?
>
> that's pretty much it, yes. by "linear" you mean here "rigid body".
> residual
> effects could be also be slight inaccuracies in the motion estimation /
> interpolation artefacts (similar thing), or physics effects such as
> "spin-history".
>
>> I am interested in removing some of these components from my FEAT
>> analysis.  If I create regressors of no interest, it seems to me that
>> I am reducing the power of my regressors of interest.  Christian had
>> mentioned in a previous message that it's possible to make a 4D
>> representation of the noise and subtract it out.  How can I do that?
>
> you can do that with the -f option - if it's not already in the email
> archive, Christian can mayb expand slightly on that.
>
> Thanks, Steve.
>
>
>  Stephen M. Smith  MA DPhil CEng MIEE
>  Associate Director, FMRIB and Analysis Research Coordinator
>
>  Oxford University Centre for Functional MRI of the Brain
>  John Radcliffe Hospital, Headington, Oxford OX3 9DU, UK
>  +44 (0) 1865 222726  (fax 222717)
>
>  [log in to unmask]  http://www.fmrib.ox.ac.uk/~steve
>