I am not sure if my proceedings are correct, but I have been doing the follow. I estimate confidence intervals (CI) for mean and variance, from dependent variable. If these intervals have some values that are equals, so I say that variable do not violate Poisson presuppose: mi = variance = lambda. If variance's CI is higher comparing to any of mean's IC value, so we have a super-variability and negative binomial regression is used.
I developed this method from my self, as I have been understanding Poisson models. I am not sure it is correct, and I apologize for mistakes. It would be interesting listen to someone who could criticize this method.