[Eeglablist] How to recognize successful ICA (number of data points issue)

[arno]

Dear Jim,

good components are components that look like they can be modeled using 
one (or two symmetrical) dipoles. We are actually preparing a report 
comparing how "good" different ICA algorithms are with respect to this 
measure. The minimum amount of data points should be (the number of 
channels)^2 (because that is the number of value in the ICA weight 
matrix and you want at least one data point per value in this matrix). 
If not, you should use PCA to reduce the number of dimensions ('pca' 
option of runica()). You can also compare how "good" your decomposition 
is with or without PCA. From our experience, if you use PCA, you obtain 
less dipolar components (which make sense because you obtain less 
components overall).

Hope this helps,

Arno