Dear Anette,

On Wed, Apr 21, 2010 at 9:02 AM, Anette Giani
<[log in to unmask]> wrote:
> Filtering:
>
>
>
> First, I would like to understand the difference between filtering during
> preprocessing (spm_eeg_filter) and filtering during inversion (inverse.lpf &
> inverse.hpf). Does it make any difference at all whether I filter during
> preprocessing or during inversion?

In principle these are both filtering operations, but the one in
preprocessing is implementing with butterworth filter whereas the one
in inversion is implementing by projection to the null space of
discrete cosine set (same as is done for fMRI). The precision in
frequency of these discrete cosines can be quite low depending on your
epoch length. So the most precise way is filtering continuous data
before epoching. Filtering in inversion is only there to try out
different options without having to redo preprocessing every time.

>
>
>
> Second, what is the difference between filtering and windowing my data?
> Similarly, what would you suggest for steady state responses: Filter
> frequencies of interest prior to inversion or window them after inversion?

> Interestingly, windowing my data around 40 Hz (which corresponds to the
> auditory stimulation frequency) reveals auditory activation. However, if I
> apply a more broadbent window of 3:45 Hz all auditory activation seems to be
> gone. In this case, all activation seems to be localized in visual cortex
> (Importantly, visual stimulation occurred at a frequency of 6 Hz.) Is it
> possible that stronger visual activation outperforms auditory activation?
>

Yes, it's possible and the results you get empirically answer your
question. The inversion aims at producing a solution which explains
maximal variance in the data. Therefore if the signal at low
frequencies is much stronger than at high frequencies the algorithm
will focus at modelling the low frequencies. Windowing after inversion
will not solve this because at this stage the solution is already
computed. So if this is indeed a problem for you, you should filter
your data to the range of interest before the inversion.

> Source Priors:
>
>
>
> I would like to insert auditory and visual source priors for localization.
> However, in different conditions I expect either the auditory, the visual
> cortex or both cortices to be active. So far I  invert all conditions
> simultaneously (as suggested in the manual) and defined a general ROI (which
> comprises visual and auditory cortex). However, I was wondering if it makes
> more sense to invert each condition by itself and to define specific ROIs
> for each condition (e.g. auditory cortex for auditory conditions, visual
> cortices for visual conditions and auditory and visual cortices for the
> multimodal conditions).
>

The question is, what conditions you are planning to compare. If you
want to put images from some conditions in the same design later I
suggest that you invert them together.  It's not quite clear to me
whether you introduce priors via the new fMRI prior mechanism or by
restricting solutions, but in either case if you compare conditions
later you should use the same parameters for inverting all of them
because what you want to show is that there are differences in the
data rather than differences in the parameters of the inversion. To
take an extreme case if you in the visual condition restrict the
solution to V1 and in the auditory to A1, you'll definitely get big
differences in the stats, but these will not be physiologically
meaningful.

> Inversion:
>
>
>
> Is there a way to determine phase locking of the steady state responses
> after inversion?
>

I'm not sure what you mean. Phase locking to what? If your responses
are phase-locked to the stimulus then you should get an ERP.

> Lastly, do you have any idea why I get reasonable results with GS, while ARD
> does not work at all?
>
>

Hard to say but GS and ARD are just two different optimization
schemes. If you get different results it means that one of the gets
stuck in a worse local maximum. So you can always use the one that
yields higher model evidence.


Best,

Vladimir
>
>
>
> Looking forward to your answers!
>
>
>
> Thanks a lot for your help,
>
> Anette