[Eeglablist] ICA when merging datasets

Ana Navarro Cebrian anavarrocebrian at gmail.com
Wed Dec 4 07:13:46 PST 2019

Hi all,

I'm resending my message to eeglablist at sccn.ucsd.edu (I sent it to
eeglablist-owner at sccn.ucsd.edu by mistake). This is clear for me now, but
I'm also copying below the helpful response of Scott Makeig just in case
this is helpful for other people as well. Thanks!

On Tue, Nov 26, 2019 at 1:15 PM Ana Navarro Cebrian <
anavarrocebrian at gmail.com> wrote:

> Hi all,
> I have two datasets from two tasks collected during one session that I'm
> interested in merging to run a common ICA. Eventually, I'd like to do
> source analysis based on the ICAs to be able to compare those two tasks.
> The problem is that I already run the ICA on each of those datasets
> separately to be able to eliminate the ICA related to blinks.
> I was going to merge the datasets without keeping any of the previous
> ICAs, but I wanted to check with you whether you think this is the best
> procedure.
> I guess my doubt is whether I should merge the original datasets (before I
> removed the component or components for blinks) to make sure that I'm later
> comparing the same sources when I compare those two tasks. Since I've
> already been working on these datasets for some time, it would be easier
> for me to merge the preprocessed datasets (with the components for blinks
> removed), but I don't know if this could significantly change the results.
> It would be helpful to know what your thoughts are.
> Please let me know if that doesn't make sense.
> Thanks!
> Ana

Ana -  It should be best/easiest to concatenate the two sets and run a
single ICA decomposition. Removing separately decomposed eye components,
then concatenating and decomposing would risk dimension and activity
mismatch problems (components might be duplicated, ...).

If the two conditions are quite different, however (for instance ...
reminiscing and video game playing!?) then you might want to see whether
some (brain) components reliably appear in only one of the conditions. To
find such components by their equiv dipole locations, use dipoledensity.
Statistical testing for condition differences could be by creating
dipoledensity maps for surrogate 'condition' data win which the condition
labels are randomly permuted (or drawn at random). Then find whether the
max condition dipoledensity difference between the two actual conditions
exceeds that found in the surrogate dipoledifferences ....

Scott Makeig

More information about the eeglablist mailing list