Precision of incidence predictions based on Poisson distributed observations

Disease incidence predictions are useful for a number of administrative and scientific purposes. The simplest ones are made using trend extrapolation, on either an arithmetic or a logarithmic scale. This paper shows how approximate confidence prediction intervals can be calculated for such predictions, both for the total number of cases and for the age-adjusted incidence rates, by assuming Poisson distribution of the age and period specific numbers of incident cases. Generalizations for prediction models, for example, using power families and extra-Poisson variation, are also presented. Cancer incidence predictions for the Stockholm-Gotland Oncological Region in Sweden are used as an example.