Do cancer predictions work?

Two different types of simple extrapolation models were investigated as tools for cancer incidence prediction. The number of incident cancer cases by sex and site in Finland was predicted using a prediction interval for each year from 1967 to 2003 on the basis of historical cancer incidence data obtained 5 to 15 years earlier. Cancer sites where major human-made changes in aetiology and diagnostics had occurred were analysed separately. Assuming that such changes had not occurred, the 95% prediction intervals based on normal errors of the age-standardised rate and on Poisson models included the observed number in 65-100% of the years. The Poisson models produced, on average, shorter intervals and were more capable of indicating a site where the model assumptions did not hold true. Simple extrapolation models may be used with some caution on coverage when there are no known factors that might make abrupt changes in the temporal development of cancer incidence. On the other hand, they may be used for detecting the effects of such factors.