On the validity of inferences drawn from the fitting of Poisson and negative binomial distributions to observed accident data