Dear Torben, dear Guillaume,

Thanks for your much appreciated feedback and apologies for my late reply.

My problem was solved shortly after my previous email and I failed to
reproduce the problem at first. Now I know that it was unrelated to SPM.

I agree with Torben that the beta estimates do not change although I was
suspecting something along the lines of what Guillaume suggested to be a
promising suspect. Perhaps the estimation of the autocorrelation would
only be a problem when doing inference on the estimates (which I'm not
interested in for now).

On a completely unrelated note, here is a random piece of advice to
everyone: always remember that the "glob" command from Python's "glob"
module returns sorted file lists on Mac, but unsorted lists on Linux
(probably depending on the file system), possibly shuffling your NIfTI
file sequence ;)

Sorry for blaming SPM!

Take care,
Michael


On 24/04/17 12:12, Guillaume Flandin wrote:
> Dear Michael,
> 
> The analysis mask is formed as the intersection of the implicit,
> threshold and explicit masks. If you want to get rid of the threshold
> mask to impose your explicit mask only, specify your model as usual with
> the explicit mask, then run this:
> 
> load SPM.mat
> SPM.xM.TH = -Inf(size(SPM.xM.TH));
> save SPM.mat SPM -v6
> 
> followed by model estimation.
> 
> In your previous model estimations, things went a bit funny because your
> analysis mask consisted of all voxels within the image (ie including
> background) and this affected the estimation of the temporal
> autocorrelation, as it is pooled over voxels.
> 
> Best regards,
> Guillaume.
> 
> 
> On 23/04/17 07:58, Michael Bannert wrote:
>> Dear SPM community,
>>
>> I would like to estimate beta parameters for every subject in my sample
>> using exactly the same explicit (whole brain) mask. I would like to have
>> beta estimates for exactly the same voxels in MNI space. Can SPM do this?
>>
>> There is the option to specify an explicit mask when setting up a GLM.
>> However, SPM automatically still applies some implicit masking. From
>> what I understand the critical parameter for implicit masking is
>> mask.tresh in the spm_defaults (usually = .8, meaning that voxels with
>> less than 80 % of the global mean signal are discarded).
>>
>> The strange thing is that when I set this to 0 in order to disable
>> implicit masking, the beta estimates become extremely noisy and my
>> effects are gone. This is very odd behavior.
>>
>> When I set it to .01, the betas estimates seem to be fine again but it
>> is not exactly what I want. The "effective masks" that SPM comes up with
>> for every subject still differ a little here and there, so this is no help.
>>
>> How can I achieve what I want with fMRI?
>>
>> Thanks & best,
>> Michael
>>
>