How much does output coding affect performance in neural networks?

I googled around a bit and couldn't find very much on this subject. Let's say I'm trying to train a binary classifier; I could set up an output layer with one neuron and threshold its output to derive an answer, or I could set up an output layer with two neurons and derive an answer by determining which one was most activated. With more than two classes, there are yet more choices (coding output as a binary number, etc).

Does anyone know of a paper on the subject, or even a rule-of-thumb out there putting one approach above the others?