[Eeglablist] order of analysis

[arno]

Mayuresh Korgaonkar wrote:
> Hello Arno and EEGlab authors,
>  
> Does extracting epochs before running ICA have different results than 
> running the ICA data first on continuous data and then epoching it? I 
> have different trial types plus instructions and for epoching I select 
> only the trial types and leave out the instructions.

If you are short on data (do not have enough data points, for ICA we 
recommend 20*nb_channel^2 and at least more than 3*nb_channel^2), then 
you might want to use high pass filtered continuous data. Also epoching 
may introduce non-linearities (because the baseline is removed 
differently in each data channels so the mixture of brain source at 
scalp sensors is not truly linear anymore; in fact David Groppe 
demonstrated that using a short baseline does impair the reliability of 
the ICA decompositions). However, I do not think we can make a rule out 
of this. Epoching allows you to force ICA to focus on data regions you 
are interested in, and this good too.

What I would recommend, if you have enough data to use data epochs, is 
that you high pass filter the data at 0.5 Hz, extract epochs but do not 
remove the baseline before running ICA. Then, run ICA, and remove 
baseline after this is done.

Best regards,

Arno