Research on Mortality Models Incorporating Cohort Effects from a Compositional Data Perspective

Abstract

Against the backdrop of accelerating global population aging, traditional mortality models are often sensitive to outliers and struggle to capture tail characteristics, leading to an underestimation of future trends. It is worth noting that mortality models with cohort effects, by introducing cohort effects on top of the age-period only framework, can capture mortality patterns more comprehensively. To enhance the robustness and accuracy of predictions, this study embeds the Renshaw–Haberman (RH) model with cohort effects into the framework of Compositional Data Analysis (CoDa) to construct a CoDa-RH model. This study uses data from six countries in the Human Mortality Database (HMD), including four developed countries (Australia, Spain, the United States, and the United Kingdom) and two developing countries (Chile and Bulgaria). The centered log-ratio (clr) transformation is applied to the age distribution structure of death counts, and the bootstrap method is used to construct prediction intervals for life expectancy. Empirical results show that, compared with the traditional Lee-Carter (LC) model and the RH model, in terms of indicators based on log mortality rates, the CoDa-based models exhibit lower Aitchison Distance (AD) and Mean Absolute Error (MAE) on the test set. In terms of life expectancy prediction, this study randomly selects three of the six countries (Spain, UK and Chile) and provides their life expectancy forecasts up to 2035.