I am not sure the case is really so complicated. Do you talk about statistically dependent/independent variables? I attached some links for you to get clear about definitions.

Actually, if variables are statistically dependent, it is mathematically NOT a good idea to treat them as independent. You will be "punished" and accused of cheating. For example, in Markov chains variables are considered to be independent (link). Otherwise, the corresponding mathematical laws don't hold (learning etc). If they are known to be dependent, you have to treat them accordingly and can't use in a HMM classifier.

However, variables you mention (an EEG-wave, being amplitude variability, peak latency and peak amplitude) don't seem for me to be statistically dependent since there could be a simpler dependence between them, for instance, a mathematical formula. For example, the peak variable could be considered as the maximum of the peak amplitude.

If they are really dependent, you might be interested in the project (EBMC) in WEKA that can be used for Bayesian multivariate classification. WEKA is an open source project of the Wakato University. Using WEKA you can experiment with your data in a GUI or perform statistical experiments programmatically.

https://en.wikipedia.org/wiki/Hidden_Markov_model

https://en.wikipedia.org/wiki/Dependent_and_independent_variables#Statistics

http://weka.sourceforge.net/packageMetaData/

Dear Hugo, I think you mean as a group of three dependent variables rather than as independent variables. Yes, I think this is a good idea provided that there is no clear dependency between these variables and that the distribution of your data satisfies the assumptions of MANOVA. I'd recommend getting hold of Andy Field's book and looking at the example in his chapter on MANOVA.

Thank you all for your awnsers. I will see If I can get the Andy field book in my library. I had it for a course, but left it in a train sadly.