What is the standard way to analyze EEG data in a mismatch negativity paradigm?

Question

I'm running an EEG experiment using a modified auditory mismatch negativity (MMN) design, and I'm wondering if anyone can tell me the best method for data analysis (and recommend any stats programs/packages for this purpose as well).

The experiment will only have a few (6-8) subjects with a high number of trials per subject (~700). The matched condition makes up 75% of the trials. The stimuli are selected from a closed set of 10 items, with an equal frequency of items in the matched/mismatched conditions.

One of the differences between this study and normal auditory MMN tasks is that the expected/unexpected stimuli vary between trials with the expectation set up on a trial-by-trial basis. Therefore, it's possible that the difference in event-related potentials (ERPs) to expected/unexpected stimuli will be averaged out due to small variations in response to these changing stimuli.

Does someone out there know the best EEG data processing and analysis techniques to detect the small ERP in response to auditory MMN stimuli? Can anyone recommend techniques to avoid averaging out small ERPs whose latencies might vary?

Answer 1

Programs/packages for EEG analysis
There are decent MatLab toolboxes with good tutorials for for the analysis of EEG data. The EEGLAB toolbox (tutorial) can be operated by both GUI and command-line (and script). The fieldtrip toolbox (tutorial) is mainly operated by command line / script.

Of course there are also (commercial) software packages for EEG analysis that do not require Matlab (e.g., BESA, Brain Vision Analyser, eeprobe), but I am not familiar with them.

Are there techniques to avoid averaging out small ERPs whose latencies might vary?
No, probably not. ERPs are calculated by averaging across multiple trials because this averaging cancels out all components that jitter in time (that is, mostly noise), i.e. the thing you want to avoid (canceling out of jittering components) is actually the purpose of calculating ERPs.

However, to be sure you should have a look at reviews or tutorials for the measurement of the mismatch negativity. See for instance

Duncan CC, Barry RJ, Connolly JF, Fischer C, Michie PT, Näätänen R, Polich J, Reinvang I, Van Petten C. (2009) Event-related potentials in clinical research: guidelines for eliciting, recording, and quantifying mismatch negativity, P300, and N400. Clin Neurophysiol. 120(11):1883-908. (pubmed)

Answer 2

I use the FieldTrip toolbox in Matlab to analyze my own modified auditory MMN experiment :) But I use MEG, so I don't have that many software options. The toolbox is very powerful but it has a steep learning curve and I would recommend it only if you already have both Matlab and EEG data analysis experience.

I don't analyze my data in the classical MMN way so I can't comment on that. I do hear that the MMN crowd can sometimes be a bit dogmatic in terms of how to approach data analysis, so it's a good idea to get precise information before you sit down to do it. Perhaps contacting someone who has a published MMN study would do the trick.

One of the differences between this study and normal auditory MMN tasks is that the expected/unexpected stimuli vary between trials with the expectation set up on a trial-by-trial basis. Therefore, it's possible that the difference in event-related potentials (ERPs) to expected/unexpected stimuli will be averaged out due to small variations in response to these changing stimuli.

Since the frequencies of the types of physically different stimuli are identical in the two conditions, this isn't an item of concern. However, there will be more noise in the unexpected condition relative to expected, because of the smaller total number of trials. I'm not sure in what way this can affect the stats (I believe it wouldn't bias them in any systematic way, but I haven't asked an expert - please let me know if you find out!), but generally as a precaution I take a random sample of expected trials, equal to the number of unexpected trials per participant. This way, there is an equal amount of noise in both conditions because you have an equal-sized sample of physically identical stimuli, and the only difference is in stimulus expectation.