Mortality forecasts by age and cause of death: How to forecast both dimensions?

Mortality forecasts by age and cause of death are important for more efficient spending on, for example, health care and medical technology. However, there is a reluctance in including the cause of death dimension to the forecast, as forecasts by cause are confronted with many methodological problems. While some of these problems have been addressed in the last two decades, an important remaining issue with forecasts by cause is their inconsistence with all- causes forecasts. This problem relates to how changes in mortality by age and cause interact. So how can we forecast this relation in a coherent manner? To address this problem, we use a model framework based on a Compositional Data Analysis (CoDA) approach which models 1) age and cause simultaneously; 2) cause-of-death distribution within each age group and 3) age-at-death distribution within each cause. We specify multiple models within each of the three frameworks to obtain a better understanding of the age and cause interactions. The results show that forecasting cause-of-death distribution within each age group generally provides the most accurate forecasts and allows for the forecast by cause and for all-cause to be consistent with one another.