Dear Thomas,
does your point below regarding settings previous to the FDR to the settings
in the results display, or to the default settings in the estimation process
(where F = 1 leads to saving data for all voxels but at the cost of time and
memory)? I tend to apply an automatic SVC by using ROI masks (using the WFU
pickatlas), followed by FDR thresholding...does this violate assumptions
underlying the use of FDR?
many thanks,
Tobias Egner.
-----Original Message-----
From: Thomas E. Nichols
To: [log in to unmask]
Sent: 10/22/2003 4:41 PM
Subject: Re: using FDR with SVC
Andrew,
[log in to unmask] wrote:
> I wondered if I could ask you about FDR and small volume corrections
> more generally. In particular, in the case where, in order to 'see'
> an activation, one has to first threshold quite liberally ( P < 0.05
> uncorrected say), so that a lot of the brain is active, and then
> apply a small volume correction. Depending on the initial
> thresholding, I've noticed that the P values associated with the
> SVC/FDR will differ - presumably due to the large number of voxels
> in the small ROI?
Yes, this is an excellent point. The problem is that some voxels
within the VOI may be excluded because they fail to meet the original
(whole volume) significance critereon. This is very bad as far as FDR
is concerned, because the FDR calculation assumes that the statistic
values are unconstrained (i.e. not pre-filtered based on intensity).
> I wondered if there is some principled way that one can initially
> threshold the SPMs using an uncorrected P value, before applying SPM
> and FDR?
Yes. Use an uncorrected threshold of "1". The uncorrected threshold
is interpeted as a P-value if less than *or* equal to 1. A P-value
threshold of 1 says "Show me everything". Then apply the SVC mask.
This should give you consistent results.
-Tom
PS: A "Show me everything" threshold might cause the cluster-counting
routine, spm_cluster, to crash. If this is the case, instead try
using lower and lower uncorrected thresholds until the SVC cluster
size k equals the SVC search volume S in voxels (shown in the footer).
When the cluster size (or sum of cluster sizes) equals S you're not
excluding anything based on intensity.
-- Thomas Nichols -------------------- Department of Biostatistics
http://www.sph.umich.edu/~nichols University of Michigan
[log in to unmask] 1420 Washington Heights
-------------------------------------- Ann Arbor, MI 48109-2029
>
> Many thanks for your help,
>
> Andrew
>
>
>
> At 17:25 21/10/2003 -0400, you wrote:
> >Andrew,
> >
> >Sorry for the delay. I'm afraid I don't know what could be
happening.
> >I take it that without FDR you do get *some* voxels?
> >
> >I would be very grateful if you could send me the relevant files
> >so I could debug this. I would only need
> >
> > 1. Statistic image of interest, .hdr, .img & .mat
> > 2. SPM.mat file
> > 3. The image itself, or a URL to the ROI-defining image.
> >
> >Thanks!
> >
> >-Tom
> >
> > > Dear Dr Nichols,
> > >
> > > Sorry to bother you, I had a question about the use of FDR with
the small
> > > volume correction in SPM 99. I've been analysing some PET and fMRI
data
> > > with FDR, and SVC (we are interested in the amygdala and
hippocampus in
> > > particular). I have been the option of creating a sphere centred
around a
> > > specific coordinate in the SVC correction, and this works fine.
However,
> > > when I use an image for the small volume (based on an analyze
image of the
> > > amygdala from the MarsBar ROI library), I get the following error
message:
> > >
> > > " Matrix is out of range for deletion
> > >
> > > Error in ..... /spm99_fdr/spm_P_FDR.m on line 107 =>
Ts(spm_read_vds(Vm)
> > > ==0 = [];
> > >
> > > Error in ..... /spm99_fdr/spm_list.m on line 276 => [Qu QPs] =
spm_P_FDR
> > > (u,df,STAT,n,Vspm,Msk);
> > >
> > > Error in spm99_fdr/spm_VOI.m on line 170 => TabDat =
> > > spm_list('List',SPM,VOL,Dis,Num,Str,hreg); "
> > >
> > >
> > > Would you have any idea as to what the problem might be?
> > >
> > > Many thanks for your help,
> > >
> > > Sincerely,
> > >
> > > Andrew Lawrence
> > >
> > >
> > >
> > > Andrew Lawrence,
> > > MRC Cognition & Brain Sciences Unit,
> > > 15 Chaucer Rd,
> > > Cambridge CB2 2EF,
> > > UK
> > >
> > > Tel: (44) 1223 355294 ext. 790
> > > Fax: (44) 1223 359062
> > > email: [log in to unmask]
> > > web: http://www.mrc-cbu.cam.ac.uk/~andrew.lawrence/
> > >
>
>
> Andrew Lawrence,
> MRC Cognition & Brain Sciences Unit,
> 15 Chaucer Rd,
> Cambridge CB2 2EF,
> UK
>
> Tel: (44) 1223 355294 ext. 790
> Fax: (44) 1223 359062
> email:
> web: http://www.mrc-cbu.cam.ac.uk/~andrew.lawrence/
-Tom
