Subnational Mortality Modelling: A Bayesian Hierarchical Model with Common Factors

Abstract

We propose a new model in a Bayesian hierarchical framework to project mortality at both national and subnational levels based on sparse or missing data. The new model, which has a country–region–province structure, uses common factors to pool information at the national level and within regions consisting of several provinces or states. We illustrate the model’s use by drawing on a new database containing provincial-level mortality data for China from four censuses conducted during the period 1982–2010. The new model provides good estimates and reasonable forecasts at both the country and provincial levels. The model’s forecast intervals reflect provincial- and regional-level uncertainty. Using subnational data for the period 1999–2018 from the Centers for Disease Control and Prevention (CDC), we also apply the model to the United States. We use mortality forecasts to compute and compare national and subnational life expectancies for China and the United States. The model predicts that, in 2030, China will have a similar national life expectancy at age 60 and a similar heterogeneity in subnational life expectancy as the United States.

Keywords

• mortality modelling
• Bayesian framework
• subnational populations
• life expectancy