I'm sure Allison knows this, but just to add a point, it's a gross error
to do _both_ 3) (what I've been calling "voxelwise intensity
normalization," for want of a better term) and 2) (what SPM calls
"global scaling").  But it's really only a gross error if you do
voxelwise intensity normalization _followed by_ global scaling.  There
might be arguments against doing global scaling followed by voxelwise
intensity normalization, but they'd be subtler than the argument against
the opposite order.

Stephen J. Fromm, PhD
Contractor, NIMH/MAP
(301) 451-9265
 
 

-----Original Message-----
From: Nugent, Allison C. (NIH/NIMH) [E] 
Sent: Wednesday, May 21, 2008 2:55 PM
To: Fromm, Stephen (NIH/NIMH) [C]; SPM
Subject: Re: comparability of data from different scanning sessions

I'd add that in SPM the choice is between global mean scaling (choosing
"none" under global normalization), as you describe or the "scaling"
option under global normalization, that divides every volume (time
point) by the mean of all voxels in that volume.  

So there's three choices:
1) dividing all voxels in a session by the mean over all time points and
all voxels in that session
2) dividing all voxels in a time point in a session by the mean over all
voxels in that volume
3) dividing each voxel by the mean of that voxel over all time points

Method 2) would have the added benefit of possibly removing spike-type
noise or a wandering baseline that affects all voxels at a time point
equally.  However, it is altering your time series, which may affect
your results. 

Allison Nugent
MRI Physicist
SNMAD/MIB/NIMH/NIH
Office: (301)451-8863
Mobile: (301)408-8560
[log in to unmask]


-----Original Message-----
From: SPM (Statistical Parametric Mapping) [mailto:[log in to unmask]]
On Behalf Of Fromm, Stephen (NIH/NIMH) [C]
Sent: Monday, May 19, 2008 1:55 PM
Subject: Re: comparability of data from different scanning sessions

Some people like "voxel-wise" better, because it nominally provides true
percent signal change, whereas grand mean scaling gives units in
fractional or percent signal change, but it's not truly local percent
change.

On the other hand you could argue that due to partial volume effects,
it's not clear to what extent you're really isolating true percent
signal change in the case of voxelwise intensity normalization.

The main disadvantage of voxelwise scaling that I see is that it changes
the spatial structure of the data.  Meaning, when doing things like the
familywise error correction, SPM assumes that you're dealing with a
Gaussian random field, but that's no longer going to be true if you
scale by each voxel's own session mean.  A related issue is, in the case
of voxelwise, whether to do it before or after smoothing.  Myself, I'd
lean towards after, because my guess (but it's only a guess) is that you
get more noise cancellation by smoothing if you smooth the data before
dividing by each voxel's average over time.

The most important issue is which method is more "valid." One of the SPM
gurus told me (probably on this list) that the best way to measure that
would be to see which method leads to less variation---say, between
subjects.  You'd think that voxelwise would do that, but I don't know of
any study that has looked at actual numbers.

As for which situations would merit one method or the other, I'm not
sure.

One final note:  you shouldn't do global scaling if you do voxelwise
intensity normalization.

Best,

Stephen J. Fromm, PhD
Contractor, NIMH/MAP
(301) 451-9265
 
 

-----Original Message-----
From: Jason Steffener [mailto:[log in to unmask]] 
Sent: Monday, May 19, 2008 12:23 PM
To: Fromm, Stephen (NIH/NIMH) [C]; [log in to unmask]
Subject: Re: [SPM] comparability of data from different scanning
sessions

Dear Stephen,

Hello. You say that the voxel-wise vs whole brain normalization is
debatable. I am currently 
in a position of addressing which to perform. Would you care to comment
on the pros/cons 
of the two approaches and the potential situations in which one should
be performed over the other?

I hope you don't mind and thank you.
Jason.


> A "stronger" form of intensity normalization would be to divide each
> voxel by the average over that single voxel over the entire run.  (So
> "space" = "that particular voxel only", and "time" = "that run"; time
is
> the same as in GMS.)
> It's debatable whether you should do GMS or the strong "voxel-wise"
> normalization, but you should certainly do GMS.  (GMS is provided for
in
> SPM; voxelwise is not hard to do but isn't directly an option in SPM.)
>  
----- Original Message ----
From: "Fromm, Stephen (NIH/NIMH) [C]" <[log in to unmask]>
To: [log in to unmask]
Sent: Monday, May 19, 2008 10:51:44 AM
Subject: Re: [SPM] comparability of data from different scanning
sessions

I haven't used SPM5 much at the single subject level.

I can't find anything about grand mean scaling in the interface.  The
code (in spm_fmri_spm_ui.m) makes it look like it's done automatically
if you don't do global normalization (meaning "Global normalization"
equals "none", not "scaling").

Stephen J. Fromm, PhD
Contractor, NIMH/MAP
(301) 451-9265
 
 

-----Original Message-----
From: Julia Weiler [mailto:[log in to unmask]] 
Sent: Monday, May 19, 2008 10:40 AM
To: Fromm, Stephen (NIH/NIMH) [C]; [log in to unmask]
Subject: Re: [SPM] comparability of data from different scanning
sessions

Thanks for the fast response.
I am using SPM5.

Julia



Fromm, Stephen (NIH/NIMH) [C] schrieb:
>
> It's done at the first level (that is, subject level).
>
> What version of SPM are you using?
>
> Stephen J. Fromm, PhD
> Contractor, NIMH/MAP
> (301) 451-9265
>
>
------------------------------------------------------------------------
>
> *From:* Julia Weiler [mailto:[log in to unmask]]
> *Sent:* Monday, May 19, 2008 10:35 AM
> *To:* [log in to unmask]
> *Cc:* Fromm, Stephen (NIH/NIMH) [C]
> *Subject:* Re: [SPM] comparability of data from different scanning 
> sessions
>
> Dear SPMers,
>
> I previously posted a message regarding the analysis of fMRI data from

> different scanning days.
> I received the advice to perform grand mean scaling in order to 
> normalize my data.
> Now, I would like to ask for some help regarding how to perform this 
> analysis:
>
> Do I have to specifiy grand mean scaling on the 1st or 2nd level of 
> the analysis?
> And, if I have to do it on the first level, do I have to use a PET 
> model for this operation?
> I could not find an option for grand mean scaling on the first level 
> of the fMRI analysis
> (or is it the same as global normalization)?
>
> Thank you very much in advance
>
> Julia
>
>
>
> Fromm, Stephen (NIH/NIMH) [C] schrieb:
>
> There are various forms of intensity normalization.
>  
> Grand mean scaling is the "weakest" form:  you divide each voxel by
the
> average over "space" and "time", where "space" is "all intracerebral
> voxels" (based on SPM's criteria for what's in vs out of brain) and
> "time" is "all timepoints (volumes) in a given run/session".  (I like
to
> call it a "run"; SPM usually calls it a session.)
>  
> Since grand mean scaling is "weak," it would be difficult to argue it
> does any harm.  And it can do a lot of good:  if one run is
arbitrarily
> 7% (say) higher in signal, GMS will take that into account.
>  
> A "stronger" form of intensity normalization would be to divide each
> voxel by the average over that single voxel over the entire run.  (So
> "space" = "that particular voxel only", and "time" = "that run"; time
is
> the same as in GMS.)
>  
> It's debatable whether you should do GMS or the strong "voxel-wise"
> normalization, but you should certainly do GMS.  (GMS is provided for
in
> SPM; voxelwise is not hard to do but isn't directly an option in SPM.)
>  
> Best,
>  
> Stephen J. Fromm, PhD
> Contractor, NIMH/MAP
> (301) 451-9265
>  
>  
>  
> -----Original Message-----
> From: Julia Weiler [mailto:[log in to unmask]] 
> Sent: Wednesday, May 14, 2008 8:39 AM
> To: Fromm, Stephen (NIH/NIMH) [C]; [log in to unmask]
<mailto:[log in to unmask]>
> Subject: Re: [SPM] comparability of data from different scanning
> sessions
>  
> We did not perform grand mean scaling so far.
> Is it common to do this, if one has data from different scanning days?
>  
> Julia
>  
>  
> Stephen J. Fromm schrieb:
>   
>> On Tue, 13 May 2008 09:12:55 +0200, Julia Weiler
<[log in to unmask]> <mailto:[log in to unmask]>
>>     
>  
>   
>> wrote:
>>  
>>   
>>     
>>> Dear SPM experts,
>>>  
>>> I am a new user of fMRI and are experiencing some unclarities
>>>       
> concerning
>   
>>> my data interpretation.
>>> Our task required scanning the same subjects twice on two different
>>>       
> days
>   
>>> using the same task.
>>> We conducted three conditions (A,B,C) on the first day and one
>>>       
> condition
>   
>>> (A2) on the second day.
>>> A and A2 were basically the same. The scanner and scanning
parameters
>>> were the same for the two days.
>>>  
>>> When extracting the time courses using MarsBaR, we experience the
>>> following strange phenomenon:
>>>  
>>> Conditions A,B,and C of the first day usually show similar time
>>>       
> courses.
>   
>>> However, compared to these
>>> conditions, condition A2 stays always close to zero.
>>>  
>>> Hence, significant activations in a contrast A2 > B for instance,
>>>       
> seem
>   
>>> to be due to deactivation in B
>>> rather than activation in A2. For the reverse contrast, B>A2,
>>>       
> however,
>   
>>> there seems to be activation
>>> in B and no signal changes in A2.
>>>  
>>> This is the same for a large number of ROIs. The time course for the
>>> condition of the second day is
>>> always close to zero with respect to the other conditions (although
>>> condition A2 of the second day
>>> was identical to condition A of the first day).
>>>  
>>> Is it necessary to somehow normalize the data if I want to compare
>>> sessions acquired on different days?
>>> Or could I be making any other mistake?
>>>     
>>>       
>> Did you perform grand mean scaling?
>>  
>>   
>>     
>>> Any suggestions would be very much appreciated.
>>>  
>>> Regards
>>> Julia
>>>  
>>> --
>>> Julia Weiler
>>>  
>>> Ruhr-University Bochum
>>> Inst. of Cognitive Neuroscience
>>> Department of Neuropsychology
>>> GAFO 05/606
>>> Phone:     +49-234-3223574
>>> Fax:       +49-234-3214622
>>>  
>>>       
>
========================================================================
> =
>   
>>>     
>>>       
>  
>   
>
>
>
> -- 
> Julia Weiler
>  
> Ruhr-University Bochum
> Inst. of Cognitive Neuroscience
> Department of Neuropsychology
> GAFO 05/606
> Phone:  +49-234-3223574
> Fax:    +49-234-3214622

-- 
Julia Weiler

Ruhr-University Bochum
Inst. of Cognitive Neuroscience
Department of Neuropsychology
GAFO 05/606
Phone:    +49-234-3223574
Fax:    +49-234-3214622