Here's another take on this: the probability, as determined by logistic regression of our latest simulation run, of the candidate winning the electoral vote based on his share of the popular vote.
So a 1-point popular vote win translates to about a 75 percent chance of winning the Electoral College, a 2-point win about a 90 percent chance, a 3-point win a 97 percent chance, and a 4-point win a 99 percent chance.
This is, of course, assuming that my simulation model is getting this approximately right. Note that I've lumped together McCain and Obama's numbers here, making the distribution exactly symmetrical. I don't know to what extent these numbers would hold with two different candidates and with a different set of states in play, but I'd bet it's pretty close to the mark.