On arbitrarily underdispersed Conway-Maxwell-Poisson distributions

Abstract: We show that the Conway--Maxwell--Poisson distribution can be arbitrarily underdispersed when parametrized via its mean. More precisely, if the mean $\mu$ is an integer then the limiting distribution is a unit probability mass at $\mu$. If the mean $\mu$ is not an integer then the limiting distribution is a shifted Bernoulli on the two values $\floor{\mu}$ and $\ceil{\mu}$ with probabilities equal to the fractional parts of $\mu$. In either case, the limiting distribution is the most underdispersed discrete distribution possible for any given mean. This is currently the only known generalization of the Poisson distribution exhibiting this property. Four practical implications are discussed, each adding to the claim that the (mean-parametrized) Conway--Maxwell--Poisson distribution should be considered the default model for underdispersed counts. We suggest that all future generalizations of the Poisson distribution be tested against this property.