Validation and Mortality Forecasting: When to Choose Lee-Carter over More Complex Methods?

Ricarda Duerst , Max Planck Institute for Demographic Research
Christina Bohk-Ewald, Max Planck Institute for Demographic Research

The Lee-Carter model is widely used to forecast mortality by age over time. Since its introduction in 1992, a variety of methods with increasing methodological complexity have been proposed to fit and predict particular mortality developments more closely. Considering that so many simple and also complex methods are available our study has two main goals. First, to establish validation as test prior to mortality forecasting to confirm that a selected method is suitable. Second, to show that the simple Lee-Carter model is broadly applicable, particularly in mortality contexts without substantial trend changes and decisive shifts in the age at death distribution. We adopt a detailed and objective validation design to assess the accuracy and bias of Lee-Carter forecasts 20 years ahead. For our analysis, we use age-specific mortality rates of all available calendar years for 24 highly-developed countries from the Human Mortality Database. We quantify Lee-Carter's forecast performance with the percentage error for life expectancy and life span disparity at birth in three analytical settings. Most strikingly, we find that the simple Lee-Carter model is indeed suitable to forecast mortality for many highly-developed countries in the most recent years. More complex methods, however, would have been more suitable than Lee-Carter's model between 1960 and 2000, when discontinuous mortality developments have been prevalent in many countries. Finally, we show that validation serves as a meaningful first test to decide whether a method is likely to be appropriate to forecast mortality in a country of interest.

See paper

 Presented in Session 31. Mortality Predictions