Prediction intervals for Lee-Carter-based mortality forecasts

Nico Keilman, University of Oslo

Mortality predictions based on the model proposed by Lee and Carter are uncertain for three reasons. The predictions are based on 1. observed death counts and empirical death rates. These are subjected to Poisson variability; 2. estimates of the parameters of the LC-model; 3. extrapolated values of the model's time index. Authors who have calculated prediction intervals for future death rates and life expectancies have usually taken the third source into account, but many have neglected the first or the second source, or both. The extent to which each source contributes to overall uncertainty is an empirical issue. We use simulation to construct prediction intervals around future death rates and life expectancies computed by means of the LC-model, considering all three sources. We determine the relative contribution of the second and the third source, and analyse the impact of the first source on prediction intervals. The method is applied to data from Norway for the period 1900-2004.