Dear Irina,

I am afraid this cannot be solved from a distance. I would make sure I have 
got a 'clean install' with all the paths set correctly and try again.

Good luck,

Helmut


----- Original Message ----- 
From: "Brote, Irina" <[log in to unmask]>
To: "'Helmut Laufs'" <[log in to unmask]>
Sent: Wednesday, July 06, 2005 11:02 AM
Subject: AW: PCA and SPM2


Dear Helmut,

thanks for the answer. I did such an experiment with the finger tapping and
tried to do a PCA. As Ferath explained on his homepage I took a contrast
that is for the mean, and when I do so I get an eigenimage where the first
already explains 100 % of the variance. In my way of understanding this
should be impossible.

I did tried some other contrasts as you did and I got the same results.
Everytime the first eigenimage explains 100 % of the variance. I also asked
Ferath about this results, but he doesn't understand them also.

Many thanks, Irina

-----Ursprüngliche Nachricht-----
Von: Helmut Laufs [mailto:[log in to unmask]]
Gesendet: Dienstag, 5. Juli 2005 21:32
An: Brote, Irina
Cc: SPM (Statistical Parametric Mapping)
Betreff: Re: PCA and SPM2


Dear Irina,

Ferath Kherif is probably the one who should answer you.

But let me try:

I have not directly compared the 'same' analyses between the two toolboxes.
But the results I got with MM for SPM2 were not completely none-sense. As
you know, this is hard to judge with data driven analyses unless you use
simulated/known data.

As you are a physicist, I am sure you know a lot more about PCA and fMRI
than I do. Nevertheless, a few things you may want to consider:

fMRI data is very noisy, in fact, it is very well possible that the first
eigenimages calculated with the toolbox are not at all paradigm related (or
related to what you hope to see) but pick up things like scanner drift
(unless filtered), motion (!!) etc.

What I usually do is to enter the realignment parameters (and other
regressors modelling 'noise', or in general: all kinds of effects of no
interest) in my SPM design matrix and create an F contrast for the "effects
of NO interest". In the toolbox, you then select this contrast and chose the

option "project into residual space", which means that the data is adjusted
for this contrast in the sense that the variance explained by this contrast
is removed from the data (I use 'adjust' in a complementary way to SPM,
where 'adjusted data' means that you get the data which the contrast you
select fits).  Briefly, you remove some 'noise'  from the data before PCA is

run. You can thus achieve 'more meaningful' eigenimages with a higher order.

If you have a finger tapping (or similar) data set, you could try and play
with that.

When looking at the results, chosing the color scale 'hot' sometimes helps.
Also, using an spm mip routine to look at the images might help (John
Ashburner once posted a script on the list which you may want to modify; I
attach one version which you may use at your own risk).

I am afraid this is about all I have to say!

With best wishes,

Helmut



----- Original Message ----- 
From: "Brote, Irina" <[log in to unmask]>
To: <[log in to unmask]>
Sent: Tuesday, July 05, 2005 2:07 PM
Subject: PCA and SPM2



Dear Helmut,

I saw at the SPM Mail list that you already have used the toolbox of Ferath,
but with SPM99. We are only using SPM2 and I also have downloaded the new
toolbox for SPM2. I also tried to do a PCA but I got some strange results.
Do you have experience about using the toolbox, and is there a chance to
help me? Perhaps I get this strange results only with the new toolbox. But
you also can compare both versions of the toolbox.
Many thanks in advance.

Best,

Irina

----------------------------------------------------------------------------
------------
Dipl. Phys.-Ing. Irina Brote
Klinik für Psychiatrie
Justus-Liebig-Universität Giessen
Am Steg 28
D-35385 Giessen, Germany

phone +49 641 99-45782
FAX +49 641 99-45789
email: [log in to unmask]
----------------------------------------------------------------------------
------------