Hi Max,

The peaks are an artefact of whitening (due to the particular form of
the covariance components used to model serial correlations) and I don't
think they will appear if you adjust your times series for the constant
term.
The F-contrast allows you to specify a subspace (its null space, ie the
subspace not spanned by the corresponding partition of the design
matrix) that is removed from the (whitened and filtered) data.
So you should define an "effects of interest" F-contrast where each row
contains a 1 for a regressor of interest, e.g. eye(8). And make sure it
doesn't include the constant term, the movement parameters and any other
covariates of no interest. Others might be able to give you more precise
advice in the context of DCM.

Best regards,
Guillaume.

Maximilien Chaumon wrote:
> Beginning to get the answer.
>
> Just as a reminder, I have first and last points of each of my sessions greatly offset from the rest of the data.
> I found the post below.
> But still have a question in response to the question asked by Guillaume in that post.
> I am using an F contrast that encompasses all the regressors I use to build my contrast of interest.
> i.e. if I'm interested in a T contrast
> [0 0 0 1 -.5 -.5 0 0]
> I adjust the data with a F contrast
> [0 0 0 1 1 1 0 0]
> Does the null space of that contrast contain the offset term? Is this like mean adjusting my data?
> or should it be
> [0 0 0 1 .5 .5 0 0]
> or does that not matter?
>
> Thanks a lot,
> Max
>
> https://www.jiscmail.ac.uk/cgi-bin/webadmin?A2=SPM;VQ5Uzg;20100215184029%2B0000
>
> Dear Jason,
>
> see the help text in spm_regions.m:
> what is displayed is the first eigenvariate of the filtered and adjusted
> response in all suprathreshold voxels within a specified VOI. In
> particular, the filtering consists of a temporal high pass filter and
> whitening (non-sphericity correction), see l. 134-135, while the
> adjustment is made by specifying a bi-partition of the design matrix
> through an F-contrast, see l. 151.
> Which F-contrast did you use to adjust the data? Does the null space of
> that contrast contain the offset term (effectively mean-centering the
> data)? I currently manage to reproduce what you observe only if I don't
> adjust the data with respect to its mean.
>
> Best regards,
> Guillaume.
>
>
> Jason Stretton wrote:
>> can anyone explain what the graph represents when clicking the
>> eigenvariate button? I'm particularly confused by the first and last
>> response being substantially higher than the rest of the time-series?
>>
>> Thanks
>>
>> Jason


--
Guillaume Flandin, PhD
Wellcome Trust Centre for Neuroimaging
University College London
12 Queen Square
London WC1N 3BG