Back to the main page.

Bug 2161 - make FAQ "why demean before ICA"

Reported 2013-05-08 15:50:00 +0200
Modified 2013-09-23 15:05:48 +0200
Product: FieldTrip
Component: documentation
Version: unspecified
Hardware: PC
Operating System: Windows
Importance: P3 normal
Assigned to: Jim Herring
Depends on:
See also:

Eelke Spaak - 2013-05-08 15:50:12 +0200

see bug 962 I don't immediately have the answer to the question, so someone with more ICA expertise can feel free to do this :)

Johanna - 2013-05-08 16:34:53 +0200

For the FAQ: 1) Why in general ICA requires demeaning. 2) That cfg.demean in ft_componentanalysis is performing a per-trial demeaning. However, and maybe this does require a code modification, different methods have different defaults or assumptions within them. I haven't looked at all, but the 2 most common: 'runica' states that, if data comes from dis-continuous trials, then the mean from each trial should be removed first, prior to input to runica. In other words, cfg.demean='yes' should be mandatory in combination with 'runica'. 'fastica' removes the mean as a first step. However, this is performed on the trial-concatenated data. If cfg.demean=yes, then this step does nothing. If cfg.demean=no, then could end up with some outlier trials nowhere near the mean. 'pca' it will matter signficantly whether per-trial demean or not.

Jim Herring - 2013-08-31 16:05:18 +0200

I've added myself to the CC as this bug is relevant to me for a part of the TMS-EEG tutorial. For the FAQ, according to Hyvärinen and Oja ( demeaning on the entire dataset is done to simplify the ICA. According to them if the mixed signal has zero-mean, you can assume the source has zero-mean as well. This per-channel demeaning is done (in 'runica' and 'fastica') regardless of what the setting for cfg.demean is. Per-trial demeaning is done to avoid non-physiological components reflecting the mean of the trial. However, this is problematic in the case of trials where your event of interest causes an offset. According to Makeig: "When the channel means of the input data differ significantly from the baseline means, making the data mean-zero prior to ICA training introduces an 'active' DC (square-wave) component without physiological significance into the ICA decomposition. This problem may may be minimized by appending response epochs without baseline offsets to the input data (example: responses to standard stimuli appended to responses to target stimuli containing large monophasic late waves)." -