Dear Hamed,
You can still define session-specific variance components (and high pass
filter) even if you concatenated your data into a single session, see eg
spm_fmri_concatenate.m.
Serial correlation in fMRI has been looked at for many years, see eg the
"Analysis of Functional MRI Time-Series - *" series published in the
mid-90's:
http://www.fil.ion.ucl.ac.uk/spm/doc/biblio/Keyword/NONSPHERICITY.html
See also Keith Worsley's papers:
http://www.math.mcgill.ca/keith/fmristat_paper/fmristat_abstract.html
and many others since then.
Best regards,
Guillaume.
On 31/03/17 14:39, hamed nili wrote:
> hey,
>
> i just saw this email!
> thanks,
> two follow-up questions:
> 1. so am i right that this approach doesn't make much sense when
> multiple runs are /concatenated/?(by run concatenation i mean having the
> same regressor for conditions that are shared across runs, setting the
> hpf to Inf and using Fourier basis functions to filter the data for each
> run).
>
> 2. i understand that ignoring temporal dependencies and assuming that
> the residuals are i.i.d is theoretically wrong (hence doing GLM as
> oposed to GLS), but have you or anyone else looked at this in a rather
> systematic way?
> so i am asking whether in a typical fMRI dataset, GLM and ReML
> implementation can give very different results or not? and if yes, under
> which scenarios are they mostly different?
>
> thanks again
> hamed
>
>
>
>
> On Wed, Nov 16, 2016 at 11:37 AM, Guillaume Flandin <[log in to unmask]
> <mailto:[log in to unmask]>> wrote:
>
> Dear Hamed,
>
> The ReML estimation takes place session by session, see
> spm_est_non_sphericity.m.
> The choice of covariance components, AR(1) or FAST, is unrelated to
> having multiple sessions or not.
>
> Best regards,
> Guillaume.
>
>
> On 11/11/16 14:44, hamed nili wrote:
> > Dear SPM experts,
> >
> > I am writing to ask about the GLS implementation in SPM. So as far
> as I
> > know SPM uses restricted ML to obtain an estimate of the covariance
> > matrix (required for the GLS solution).
> > So when entering multiple runs, how does SPM compute the
> covariance matrix?
> >
> > Intuitively, I think it is wrong to do the same as it does for one run
> > via concatenating the raw time series and a modified design matrix.
> > I am also curious whether FAST would be a better choice than AR(1) for
> > the serial correlations in this specific case.
> >
> > thanks,
> > Hamed
>
> --
> Guillaume Flandin, PhD
> Wellcome Trust Centre for Neuroimaging
> University College London
> 12 Queen Square
> London WC1N 3BG
>
>
--
Guillaume Flandin, PhD
Wellcome Trust Centre for Neuroimaging
University College London
12 Queen Square
London WC1N 3BG
|