Modeling and Forecasting Subnational Mortality in the Presence of Aggregated Data

This study proposes a new approach to modeling subnational mortality that relies on individual features (e.g., sex, geographical region, socioeconomic status) instead of dealing directly with subpopulations. Our strategy leads to more parsimonious models because fewer parameters are needed to explain mortality. Also, data providers might aggregate data over privacy concerns, and our framework allows for the use of such data, unlike the common subnational mortality modeling approach. A general one-step Bayesian estimation methodology that works well with most age–period–cohort mortality models proposed thus far in the literature is presented; it uses Markov chain Monte Carlo techniques by combining deterministic filtering with adaptive Metropolis steps and is well-suited for high-dimensional cases like the one investigated in this article. In a case study using real data, the framework is applied to Canadian mortality data from three datasets that encompass three features: sex, geographic region, and socioeconomic status. We show that the proposed approach combined with a reasonable mortality model provides realistic, coherent, and plausible mortality projections and that it fits the data reasonably.