[FieldTrip] LCMV time window

[Sarang S. Dalal]

Dear Jim,

Indeed, 3 samples is much too few for constructing the beamformer filter —as you guessed, this results in a poor covariance estimate that is furthermore low-rank and therefore difficult to reliably invert.

But, it is no matter, as you can construct the filter with your complete trial (perhaps 100ms as a minimum length), but examine the resulting time series (source.avg.mom) for whichever shorter intervals you are interested in.

If you are not contrasting conditions, you may consider applying:
cfg.lcmv.weightnorm   = 'nai’;

For reference, our group typically uses the following additional options for such data:
cfg.lcmv.reducerank   = 'no';
cfg.lcmv.fixedori     = 'yes’;
cfg.lcmv.projectnoise = 'yes';

It is so far undocumented, but our plotting tools in fieldtrip/contrib/nutmegtrip could help navigating the resulting source time series (in your path you’ll need: a full copy of SPM8, fieldtrip/contrib/nutmegtrip , and fieldtrip/template/atlas/aal )
cfg=[];
cfg.mripath = ‘yourmri.nii’; % or template MRI if voxel grid is in MNI space
cfg.funparameter = 'avg.mom';
cfg.atlas = ft_read_atlas('ROI_MNI_V4.nii'); % AAL atlas (optional)
nmt_sourceplot_spm8(cfg,ft_convert_units(source,'mm'));


Cheers,
Sarang


On 01 Jun 2016, at 23:43, Jim Cavanagh <jim.f.cav at gmail.com<mailto:jim.f.cav at gmail.com>> wrote:

I get radically different outputs depending on the time window used for LCMV (primarily MEG, but also a bit with EEG).

My data are sampled at 250 Hz, filtered .5 to 20 Hz, calculated as the percent change (sourcediff = (source A - source B)  / source A).

With 3 samples (12 ms), LCMV outcomes are uninterpretable (i.e. values of 100,000 % change).  Increasing samples yield increasingly realistic values, especially at about 25 samples (100 ms). I presume this has to do with the covariance, especially since there doesn't appear to be much noise in the data.  This is unfortunate since some of the ERF events that I'd like to capture are short-lived.

1) Does anyone have an explanation for this issue so I can account for it?
2) Or, barring that, feedback on an optimal way to determine temporal window size for LCMV?


--
Jim Cavanagh
Assistant Professor
University of New Mexico
https://sites.google.com/site/jcavanaghunm/home
_______________________________________________
fieldtrip mailing list
fieldtrip at donders.ru.nl<mailto:fieldtrip at donders.ru.nl>
http://mailman.science.ru.nl/mailman/listinfo/fieldtrip

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mailman.science.ru.nl/pipermail/fieldtrip/attachments/20160602/fc12b2c2/attachment.html>