Dear all,
Interesting discussion, I actually do recall a discussion on the list
quite some time ago, but this issue might have come up during
continuation off list, I do not remember anymore.
A reason for doing normalization > GLM 1st level > GLM 2nd level could
be that spatial interpolation errors at the reslicing level (eg after
normalization) have a very limited influence as they are probably
averaged out by the many time points that are estimated in a 1st level
GLM. When you do normalizing+reslicing after 1st level GLM on model
parameters, you have far less points at the 2nd level to fit your GLM
model on, and hence interpolation errors have a bigger influence or
might cause biases especially at high intensity contrasts in your image
(eg gray matter <> CSF boundaries).
This effect might be so small that it can be ignored, but it would
indeed be good to have an (empirical) grasp on it.
I do not see many other reasons for doing normalization before 1st level
GLM, which does not mean they do not exist, of course.
Cheers,
Bas
--------------------------------------------------
Dr. S.F.W. Neggers
Division of Brain Research
Rudolf Magnus Institute for Neuroscience
Utrecht University Medical Center
Visiting : Heidelberglaan 100, 3584 CX Utrecht
Room B.01.1.03
Mail : Huispost B.01.206, P.O. Box 85500
3508 GA Utrecht, the Netherlands
Tel : +31 (0)88 7559609
Fax : +31 (0)88 7555443
E-mail : [log in to unmask]
--------------------------------------------------
-----Oorspronkelijk bericht-----
Van: SPM (Statistical Parametric Mapping) [mailto:[log in to unmask]]
Namens Mohamed Seghier
Verzonden: woensdag 11 maart 2009 13:41
Aan: [log in to unmask]
Onderwerp: Re: [SPM] normalization of stats - your opinion
Hi Michael and Torben,
In my opinion, this is a perfectly valid option and some studies have
used the first option (e.g. in patient studies or case reports).
However, option 2 is more popular for some "practical" reasons; for
instance:
1- display purposes for single subjects: results/activations correctly
projected on the glass brain with meaningful coordinates;
2- timeseries extraction at particular foci of interest: it is not
unusual to extract VOIs at specific coordinates based on previous
findings;
3- region-based analysis: where the regions of interest are defined from
functional maps of different sessions (e.g. longitudinal studies);
Best,
Mohamed
Torben Ellegaard Lund wrote:
> Dear Gazzaley
>
> This is a really good question, and I don't think it has been
> discussed at the list. I think the FSL people usually does it the
> other way around, but I hvae not seen any paper comparing the benefits
> of the two approaches. I guess the reason for the current procedure
> could in part be historical, spatial normalisation of raw data was
> needed for a fixed effect group analysis which was once very popular.
> But with the current summary statistics approach it is indeed possible
> to do the stats first and normalise the contrast images (the
> normalisation parameters should of course not be estimated from the
> maps of parameter estimates). From a computational point it would make
> sense to only normalise the maps of parameter estimates in stead off
> all the raw images. The default template in SPM has a resolution of
> 2x2x2mm but most people use at least 3x3x3mm voxels so the
> normalisation involves a great deal of interpolation. For interleaved
> sequences I think it makes just as much sense to interpolate the maps
> of parameter estimates as the raw images. Similarly one could argue
> for smoothing of beta or contrast images instead of the raw images,
> but if AR(1) modelling is turned on the two do not commute. But which
> is the more correct?
>
> I think there are goods reasons for normalising the maps of parameter
> estimates but I also have the habit of running stats on normalised
data.
>
>
> Best
> Torben
>
>
> Torben Ellegaard Lund
> Assistant Professor, PhD
> The Danish National Research Foundation's Center of Functionally
> Integrative Neuroscience (CFIN) Aarhus University Aarhus University
> Hospital Building 30 Noerrebrogade 8000 Aarhus C Denmark
> Phone: +4589494380
> Fax: +4589494400
> http://www.cfin.au.dk
> [log in to unmask]
>
>
>
>
>
>
>
>
>
> Den 07/03/2009 kl. 08.17 skrev Michael T Rubens:
>
>> Just out of curiosity, and to help guide my analysis, what makes more
>> sense?
>>
>> 1) Run stats (i.e., GLM) in native space, then normalizing the
>> resulting statistical image or
>> 2) Running the statistics on normalized data?
>>
>> is the answer different for different analyses? what do you do? what
>> is your justification?
>>
>> Thanks for any advice,
>> Michael
>>
>> p.s. I have been in the habit of typically running stats on
>> normalized data
>>
>> --
>> Research Associate
>> Gazzaley Lab
>> Department of Neurology
>> University of California, San Francisco
>
>
|