A bayesian spatio-temporal approach to predict mortality rates

Abstract

In this study we extend the Bayesian Lee-Carter approach by incorporating a spatial structure to model log mortality rates at age, time and space and spatial dependencies between neighbouring populations via Conditional Autoregressive (CAR) structure and employing Markov Chain Monte Carlo (MCMC) methods to sample from the posterior distributions. We propose three different approaches to model the correlation between neighbouring areas and fit two datasets: Rio de Janeiro state microregions, available publicly via DataSUS, and Japan prefectures, available publicly via Japanese Mortality Database. This work compares model particularities while comparing with the usual single population Lee-Carter model, exploring advantages and disadvantages, goodness-of-fit metrics and model selection.