Empirical evaluation of prediction intervals for cancer incidence

BACKGROUND: Prediction intervals can be calculated for predicting cancer incidence on the basis of a statistical model. These intervals include the uncertainty of the parameter estimates and variations in future rates but do not include the uncertainty of assumptions, such as continuation of current trends. In this study we evaluated whether prediction intervals are useful in practice. METHODS: Rates for the period 1993–97 were predicted from cancer incidence rates in the five Nordic countries for the period 1958–87. In a Poisson regression model, 95% prediction intervals were constructed for 200 combinations of 20 cancer types for males and females in the five countries. The coverage level was calculated as the proportion of the prediction intervals that covered the observed number of cases in 1993–97. RESULTS: Overall, 52% (104/200) of the prediction intervals covered the observed numbers. When the prediction intervals were divided into quartiles according to the number of cases in the last observed period, the coverage level was inversely proportional to the frequency (84%, 52%, 46% and 26%). The coverage level varied widely among the five countries, but the difference declined after adjustment for the number of cases in each country. CONCLUSION: The coverage level of prediction intervals strongly depended on the number of cases on which the predictions were based. As the sample size increased, uncertainty about the adequacy of the model dominated, and the coverage level fell far below 95%. Prediction intervals for cancer incidence must therefore be interpreted with caution.