[Eeglablist] ERSP comparisons

[Arnaud Delorme]

Dear Stefania,

> i need an help!when i compute ERSP comparisons between two conditions, i 
> obtain 3 plots: ERSP on condition 1, ERSP on condition 2 and ERSP 
> difference.
> sometime, it seems that, for a specific time point and estimate of 
> linear-spaced frequency, there are marked differences; indeed, that 
> differences are not correctly detected.
>   

Just so that other people reading this message know, you are using the 
newtimef() function from the command line to compare conditions (as 
describe in the function header) between sets of trials for the same 
subject.
If you do not see anything on the difference plot, it might be that 
these difference are not significant (the difference might only be due 
to a few trial outliers).
Significance is computed by repetitively recomputing the same two images 
(then subtracting them) but shuffling trials between conditions (we will 
soon have a bootstrap version of that too). For each time-frequency 
point, you thus obtain a distribution of value, and you may use it to 
determine if the same time-frequency point on the actual (non-shuffled) 
difference falls within the tails of that distribution. It is not (yet) 
corrected for multiple comparisons so you would need to use your own 
correction method.

> i don't understand very well the bootstrap method for significance, and 
> i think that some bias can derive from "common baseline computing".
>   

The common baseline computing (which is displayed on screen) is 
necessary. Otherwise, you could observe differences that could be due to 
differences in the baseline. Let me explain. If the baseline are 
different and the activity is the same at say 100 ms after the stimulus 
onset, if you subtract the baseline independently for each condition, 
then the baseline difference will not be significant BUT the 100 ms 
activity will be (which does not make sense). By subtracting the common 
baseline, you ensure that you will be able to see BOTH significant 
differences during the baseline and after.

It is true thought that in some cases (pharmaceutical agent modifying 
dramatically the EEG) you would want to subtract the baseline condition 
by condition to see the specific interaction effect between the stimulus 
and the agent (and ignore the baseline change induced by the agent). 
Then you would have to accept the risk describe above. This is not a 
common set up though.

Best regards,

Arno