Dear Torben,
thank you for your research into spm_spm.m.
Torben Ellegaard Lund wrote:
> Dear Baerbel
>
> You could manipulate the mask spm is using for estimating the AR(1)
> process around line 796 in spm_spmthe j vector simply contains the
> indices of voxels to be included in the mask.I pursume that these
> indices are the same as the ones you would get using [Y,XYZ] =
> spm_read_vols(V,mask);
As far as I see, this mask (ie, the variable VM and what later becomes
mask.img/hdr) is not used in ReML estimation. VM is produced only AFTER
the second pass, ie, after ReML estimation where the F-criterion had
already been applied. Wouldn't you agree instead that in order to get
the single-voxel-volume through the ReML process, the global
defaults.stats.fmri.ufp should be set to one which leads to UF=0 in line
485 and to a threshold of zero in line 743?
> You may also want to have a look at the paper below concerning the
> relation between AR(1) and physiological noise modelling.
> Unfortunately correction for pulse artefacts is not quite as simple as
> just resampling the pulse data at TR.
The plan was to use the SPM machinery to estimate a regression model on
single-channel physio data as the dependent variable (ie, the physio
data replace the MR voxel time courses).
In the pulse time courses, we manually set (or correct) the peak
markers, then compute the pulse rate at every peak time, then
interpolate the peaktime-rate curve in order to get the rates at the
volume times of the MR scans. Would you please provide little more
detail why you wouldn't recommend this strategy, and why such data
wouldn't serve as additional regressors in models with MR data as the
dependent variable?
Best wishes,
Baerbel
>
>
>
> Hope this helps
>
> 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 10G, 5th floor, room 31
> Noerrebrogade 44
> 8000 Aarhus C
> Denmark
> Phone: +4589494380
> Fax: +4589494400
> http://www.cfin.au.dk
> [log in to unmask]
>
>
>
>
>
>
>
> Lund et al. Non-white noise in fMRI: does modelling have an impact?.
> NeuroImage (2006) vol. 29 (1) pp. 54-66
>
>
>
>
>
> Den 29/10/2009 kl. 14.21 skrev Karl Friston:
>
>> Dear Baerbel,
>>
>>> what would happen if all voxels of a volume were passed to ReML
>>> estimation, ie, when the F-threshold criterion
>>> j = sum((Hsqr*beta).^2,1)/trMV > UF*ResSS/trRV;
>>> (line 743 in spm_spm '$Rev: 946 $') was not applied?
>>> Would the results be worse? Is the F-thresholding just a way to save
>>> processing time and can be omitted?
>>
>> Not really. The rational for thresholding is to select responsive
>> voxels (e.g., grey-matter
>> voxels that belong to the class of voxels that show evoked
>> responses). This ensures that
>> the temporal correlations estimated within this class are a better
>> approximation to the true
>> correlations (they would be biased estimates if we included white
>> matter or other voxel
>> classes that showed a different form of serial correlations)
>>
>>> The background to these questions is the following: I would like to
>>> pass physio (pulse) data through the SPM machinery and hence use
>>> volumes that contain only one voxel (the value of which is the pulse
>>> sample at one particular TR). It is unlikely that this voxel passes
>>> the above F-thresholding criterion and likely would run into an
>>> empty analysis error. With an AR(1) model in mind, I need to have
>>> ReML estimation done. This means I would have to pass this voxel
>>> through the ReML process even if it does not meet the criterion
>>> (which could be done by temporarily setting the global
>>> defaults.stats.fmri.ufp=1 which leafs to UF=0).
>>
>> I would not worry; If the voxel of interest shows the same sort of
>> serial correlations as others
>> that survive the F-criterion, then everything should be OK. Remember
>> that the F-threshold is not
>> for disallowing subsequent analyses (this is determined by the global
>> thresholding option). The
>> F-threshold is only used to specify a subset of similar voxels for
>> ReML estimation of non-sphericity.
>>
>> I hope this helps - Karl
>>
>>
>> PS: I do not really know what physio (pulse) data are though.
>>
|