On the existence of the maximum likelihood estimates for Poisson regression

We note that the existence of the maximum likelihood estimates for Poisson regression depends on the data configuration. Because standard software does not check for this problem, the practitioner may be surprised to find that in some applications estimation of the Poisson regression is unusually difficult or even impossible. More seriously, the estimation algorithm may lead to spurious maximum likelihood estimates. We identify the signs of the non-existence of the maximum likelihood estimates and propose a simple empirical strategy to single out the regressors causing this type of identification failure.