Dear Peyman,

On Mon, Nov 16, 2009 at 3:22 PM, Peyman Adjamian <[log in to unmask]> wrote:
> What I would
> like to do is the following:
> I have two trials of 10s length, silence and masker, each repeated 80 times.
> The masker is a simple white noise played to both ears. I have used the
> beamformer analysis in the ctf software to localise activity to both
> auditory cortices as increase in power (masker vs silence). But what I also
> need to do is some sensor space analysis in the frequency domain. I need to
> know whether there are significant differences between masker and silence in
> specific frequency bands. I thought the best way to do this would be to take
> a handful of channels above one auditory cortex and do TF analysis in a
> specific time window to get a measure of induced brain activity with the
> masker onset. For example, I'd like a measure of change in the 15-25Hz band
> between silence and masker periods. This would require that the fft analysis
> is performed for each trial and then averaged (separately for masker and
> silence).
> Can SPM do what I described?

Yes, definitely. The only thing I'm slightly concerned about is the
length of your trials. The TF analysis in SPM is slightly inflexible
at the moment (which will change in the near future). What it does is
computes TF estimate putting a wavelet around each sample. So if you
are interested in looking just up to 25 Hz for long time windows of
several sec perhaps you should downsample your data to 50Hz, otherwise
you'll get huge TF datasets. The next thing you should decide is
whether your statistics will be within subject (like shown in the
tutorial chapter) or between subjects. If you want to do an analysis
within subjects then you should export each trial as an image. So you
compute TF on an epoched file and then you run 'Convert to images' on
the resulting TF dataset.

Now your data is 4-dimensional - 2 scalp dimensions, time and
frequency. Unfortunately, SPM can only do statistics on 2D or 3D data.
So either you should get rid of the spatial dimension by averaging
over channels when you export to images and then you'll get 2D
time-frequency images that you can run stats on. Or you can get rid of
the frequency dimension by averaging over frequency end then you'll
get 3D scalp x time images where the power will be averaged over the
frequency band you specify.

If you want to do statistics across subjects the above still applies
but then you should first average your TF dataset (and possibly run
'Rescale TF') and then export the average TF dataset to images. Then
you'll get 2 images per subject and will need to run a paired t-test
across subjects.

> Also, is it possible to do TF analysis with
> some stats (say bootstrapping or Man-Whitney) on a beamformer virtual
> sensor?
>

SPM is famous for its stats but doing something other than parametric
stats with it is not very straightforward. If you'd like to use
parametric stats it's the same as above. You just need to get your
data in SPM format. I'm not very familiar with what CTF software does
with those virtual electrodes but my guess is it just adds them as
additional channels to your dataset. If that is the case you can
convert the dataset, specifically selecting those channels (don't
press 'just read' but configure the conversion and one of the options
you'll get will be selecting the channels). Also in your case you
should choose 'data' when asked how to define trials. Then you proceed
as usual and you'll end up with 2D TF images for those beamformer
channels.

How to do the stats is something that you should learn from the manual
(http://www.fil.ion.ucl.ac.uk/spm/doc/spm8_manual.pdf). In the
multimodal faces tutorial chapter in the MEG part there is an example
of statistical analysis of TF data. Also in the fMRI chapter you can
find a more detailed description of the how to do statistics. I'd
recommend you to do the whole multimodal tutorial, EEG and MEG and
then you'll see both the scalp x time example and TF example.

Best,

Vladimir