[Eeglablist] SIFT model order, how much data?

[Chadwick Boulay]

Dear list,
CC Tim Mullen.

In the SIFT tutorial, I noticed that optimization of the information 
criteria identifies different optimal model orders for the RespWrong and 
RespCorr conditions but a single (more conservative) model order is used 
for both conditions. Was this just a coincidence that the conservative 
order was the same for both conditions? Or, if it is better to use a 
single model order, shouldn't we merge the two conditions when 
calculating the optimal model order? However, my impression from the 
provided example scripts 
(eeglab\plugins\SIFT\scripts\ScriptingExample_1_0b.m) is that they use a 
different model order for each condition.

So, of the following 3 options, which is least wrong?
(1) Combine all (within subject*session) data when calculating the 
optimal model order and use that single order for all subsequent model 
fitting (per condition or per trial).
(2, tutorial method) Calculate the optimal model order independently for 
each data unit (condition or single trial) that will be fit, but somehow 
extract from this array of model orders a single model order to be used 
for all subsequent model fitting.
(3, example script method) Calculate the optimal model order 
independently for each unit and use each unit's optimal model order for 
subsequent model fitting.

Thank you for any advice,
Chad

-- 
Chadwick Boulay
JSPS Postdoctoral Fellow
Keio University
Yokohama Japan
boulay at brain.bio.keio.ac.jp