Disaggregating Death Rates of Age-Groups Using Deep Learning Algorithms

Abstract

Reliable estimates of age-specific vital rates are crucial in demographic studies, while ages are, in most cases, commonly grouped in bins of five years. Indeed, public health and national systems require single age-specific data to achieve accurate social planning. This paper introduces a deep learning approach for splitting the abridged death rates, providing a more comprehensive perspective on the indirect age-specific vital rates estimation from grouped data. Additionally, we contribute to the existing literature by introducing a multi-population (countries and genders) approach, providing reliable estimates considering the heterogeneity of longevity dynamics over age, years, and across populations. We also contribute to the state of the art in indirect estimation by introducing, for the first time, a multi-population indirect estimation leveraging subnational data. Our model accurately captures mortality dynamics by age over time and among different populations. We prove the model’s ability to estimate reliable predictions of age-specific mortality rates by also studying how the hyperparameters’ choice affects the model reliability and analyzing the age-specific relative differences between the real and the estimated mortality rates.