On Wed, Nov 11, 2009 at 12:06 AM, Guise, Kevin G. <[log in to unmask]> wrote:
> I tried pre-filtering the data prior to reconstruction, and this seemed to
> improve results in the sense that localization of sources of high-frequency
> power changes seems more realistic. However, I still get the "fly dropping"
> appearance in the t-maps.
As I think I said before it's not something I'd expect with more than
8 subjects. Perhaps worth looking into.
> I have three additional questions related to what we've previously
> discussed. First, since I'm doing my stats at the sensor level and then
> choosing the time-frequency window for reconstruction based on this, would
> it be more appropriate to create figures showing differences in means
> between conditions rather than t scores?
What you are right about is that there is no reason to do the stats
for the second time at the source level (i.e. threshold the t-maps).
Showing t-maps can still be justified, however, because they emphasize
effects that are both strong and consistent (although without
separating well between the two things). On the other hand showing
mean differences is also probably OK if that's the only way not to
have those fly droppings.
> Secondly, are there any general
> guidelines that I can follow when setting the filter cutoffs when filtering
> the data prior to reconstruction? I'd just like to note something so that my
> choices don't seem completely arbitrary when it comes time to send out the
I can suggest one of two things. You can base those bands on your
sensor-level TF statistics. Perhaps for selecting the bands you should
do an uncorrected t-test or cluster-level correction because with FWE
correction you'll only get 'the tip of the iceberg' and the band will
be too narrow. Alternatively you can set the bands based on the
literature, either something specific to your task or just classical
definitions of different bands.
> Finally, when transforming the data to the time-frequency domain for
> the sensor-space analysis, I set the frequency bins linearly between 2 and
> 100 Hz; given the relationship between frequency of the Morlet filter and
> its frequency resolution, would it be advisable to instead set the frequency
> bins log-linearly? I figure this would take advantage of the resolution at
> low frequencies with minimal loss of information at high frequencies, and
> make the plots easier to read.
This is an interesting line of thought, although perhaps wavelets are
not the best way to implement it. The problem is that with wavelets
you have very high frequency resolution for high frequencies so if you
increase spacing between the bins in that range you might completely
miss something. Also SPM will have trouble with exporting this kind of
thing to images because it expects linear frequency spacing. A better
strategy (suitable for looking at high-gamma) is decreasing the
frequency resolution with increasing frequency. This can be done with
multitaper spectral analysis. I have a tool for that in MEEGTools
toolbox (I think it should be there for the most recent SPM public
update). It's not completely generic but you can also look at the code
and adjust it if necessary. In the future I'm planning to provide more
flexible spectral analysis tool in the main SPM but this might take