[Eeglablist] Difference Waves and IC Clusters

[Matthew Stief]

Greetings Everyone,

Is it possible (or sensible) to combine the difference wave and the ICA
approach to isolating the component of interest?  That is, say I have an
attention capturing stimulus and a probe, and I want to measure the P1 to
the probe.  On some trials the probe doesn't appear, so of course one
approach is to stay in ERPland and just subtract the activity of the trials
where the probe did not appear from the trials when it did appear and
measure the mean amplitude at one electrode.  But is it possible to do
something similar on clusters of ICs instead of channels?  Is it possible
to run ICA on datasets that each contain all relevant conditions, then
cluster those ICs, choose a cluster that accounts for as much of the
component of interest as possible, and then within that cluster subtract
out the information coming from the trials where the probe does not
appear?  Does that even make sense?  Would it be better to just toss out
the no probe trials for ICA purposes so that the P1 elicited by the probe
has a better chance of being isolated into a component?

Perhaps some more detail and an appeal for more general advice is in
order.  My experiment consists of a dot-probe task where the cues are a
pair of photographs, one a nude male and one a nude female.  The
photographs are rather large and appear on either side of the fixation
cross at a moderate degree of eccentricity.  The cue appears for 100ms,
followed by a 50ms gap, then a probe appears on either the left or the
right side and persists until the participant presses a button indicating
which side of the screen the probe appeared on.  On 1/3 of trials the probe
appears on the left, on 1/3 it appears on the right, and on 1/3 it does not
appear at all.  My challenge is to try to isolate out the P1 elicited by
the isolateral probe from the ongoing response to the bilateral cue.  I of
course would like to compare the condition where the probe is concordant
with a male cue to the condition where the probe is concordant with a
female cue.

If I run ICA on a dataset that consists of all five conditions (Male
Concordant L/R, Female Concordant L/R, No Probe), then in general I do not
get a separate component that accounts for the left and right P1s elicited
by the left and right probes.  If I try to separate the data into the left
condition (Male Concordant Left, Female Concordant Left, No Probe) and
right condition (Male Concordant Right, Female Concordant Right, No Probe)
and run ICA on them separately I unfortunately fare little better.  Should
I try perhaps try separating out the No Probe condition into its own
dataset, in the hope that the probe P1 being more well represented in the
data will improve its chances of being adequately represented by a
component?  If I do that, does it make sense to run ICA on the No Probe
dataset as well and then feed it into the clustering in order to pursue the
difference wave approach I mentioned above?

I am also unsure of what the optimal approach for selecting epoch limits
and baselining is.  I have gathered that while more data is better for an
ICA, for the purposes of isolating a small component like the P1 it may
actually be advantageous to opt for less data so that the P1 is not swamped
by the addition of irrelevant neural sources.  So I initially tried for a
-250 to 1000 epoch, then a -250 to 500 epoch, and am now contemplating a
-250 to 250 epoch.  I am simply removing the mean of the whole epoch before
running ICA to eliminate drift and then removing a -250 to -150 baseline
(i.e. to the onset of the cue) after ICA has been run.  Does that make
sense?  Is it correct for me to limit my baseline to the onset of the cue
rather than to the onset of the probe?  Since the cue is timelocked to the
probe it seemed that I should.

I am worried that if I both restrict the length of my epoch AND separate
rather than concatenate my conditions then I will be rapidly approaching
not having an adequate amount of data for a good ICA.  Luckily I have quite
a bit of data for each subject, with about 350 trials for each condition.
In terms of k*n^2, with the longest epoch length I'm considering and
concatenating all conditions I have a k of almost 50, but with the
restrictions I've been considering k is approaching 10 or 12 and I'm in
danger of dipping down to 5 or 6.  I could always use PCA to try to find
less components and since my research question doesn't really depend on the
exact integrity of the overall EEG dynamics then perhaps it's fine, but I
am not sure what the best approach is to reliably identify and measure the
damned P1.

Perhaps I just shouldn't be using ICA at all?

I apologize for any amateurishness in what are probably pretty basic
research design questions, I have somewhat had to teach myself these
things, especially as they relate to actually carrying them out in EEGLAB.
As always your patience and attention is greatly appreciated.


-- 
_________________________________________________________________
Matthew Stief
Human Development | Sex & Gender Lab | Cornell University
http://www.human.cornell.edu/HD/sexgender


Heterosexuality isn't normal, it's just common.
-Dorothy Parker
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://sccn.ucsd.edu/pipermail/eeglablist/attachments/20120228/5a06d22c/attachment.html>