[Eeglablist] # of ICA data points needed for stable decompositions

[Logan T Trujillo]

Hi all. I have a question concerning the rule of thumb for how many data
points are needed to achieve stable ICA decompositions.  The EEGLAB
tutorial states that as many as 3*N^2 data points per channel may be
necessary to achieve a stable ICA decomposition.  However I have performed
the runica algorithm on 30 channel grand-average ERPS (averaged across
trials and subjects using data gathered from a face recognition task),
with only 640 data points per channel (much less than the suggested
3*(30)^2 = 2700 points per channel). The decompositions that were returned
are relatively stable in the sense that re-performing the ICA (after
clearing/reloading the data and starting with a new inital weight matrix)
yields the same components maps with similar component orderings.  So my
question is this: could it be that these decompositions are stable because
of the consistent features present within the grand-average waveforms, as
compared to what would be found for individual trials?
	 I feel that performing ICA on grand-average ERPs could yield
useful information concerning the source configurations of stimulus-locked
activity. Indeed when using DIPFIT on the resultant components, I find
sources whose locations and timecourses are consistent with what has been
reported before for object and face recognition tasks. However I don't
know if the decompositions count as "good" ones, given that my number of
data points per channel is lower than suggested.

Any advice, comments, ideas?

Thanks.

Logan Trujillo
Department of Psychology
University of Arizona