ABSTRACT
In this chapter, we model mortality rates for Portugal by age, sex, and time, and estimate and forecast life expectancies. We first examine the data on mortality rates and life expectancy. There appear to be interactions between age, sex, and time. Variables interact when the nature of the relationship between a variable and the outcome of interest depends on the level of one or more other variables. We develop a decomposition technique that allows us examine interactions in the Portugese mortality data.
We then specify two Bayesian models. We pay particular attention to modelling how age patterns and sex differences change over time. We use heldback data to choose a model to use for forecasting Portuguese life expectancy up to 2035. The choice is based on three performance measures: (i) absolute error, (ii) interval score, and (iii) continuous ranked probability score. We show that the candidate model with more interaction terms avoids jump-off errors (big errors at the very start of the forecasting period). We then retrain the chosen model using all observed data, and forecast life expectancy up to 2035.
