FORECASTING LIFE EXPECTANCY: A COMPARISON OF TIME SERIES MODELS FOR A PARSIMONIOUS CH SURVIVAL FUNCTION

Abstract

We extend the CH survival function by Wong and Tsui (2015) to a more parsimonious model by reducing the number of parameters from the original 6 to 5. The modification gives a modified CH function which preserves the key features of “youth-to-adulthood” and “old-to- oldest-old” components of the original CH function. It is able to fit US population data from 1950 to 2016 well, and the series of parameter estimates are subsequently modelled via 3 different time series structures, namely: univariate, vector autoregressive, and multivariate random walk with drift model. We compare the 3 time series structures to determine which is the best model for forecasting life expectancy based on the resultant forecast errors, and 95% confidence intervals obtained via residual-based bootstrapping. The vector autoregressive model is identified as our optimal model out of the 3 time series structures. Our simulated forecast intervals for a modified CH function with vector autoregressive model are also consistently narrower than the corresponding intervals reported in Wong and Tsui (2015) under the original CH function.