“Title: The challenges of estimating mortality in small areas — using German counties as a case study”
We develop and analyze Bayesian models that produce good estimates of complete mortality schedules for small areas, even when the expected number of deaths is very small. The models also provide estimates of uncertainty about local mortality schedules. The TOPALS relational model is the primary building block, used to model age-specific mortality rates within each small area. TOPALS models produce estimates for single-year ages from a small number of local parameters. We experiment with Bayesian models for smoothing and ‘borrowing’ mortality information across space, using two alternative specifications. First we test a Bayesian model with conditional autoregressive (CAR) priors for TOPALS parameters. CAR priors assign higher probability to parameters that are similar across adjacent areas, thus emphasizing spatial smoothness in estimated rates. Second, we test a hierarchical Bayesian model, which assigns higher probability to parameters that are similar for locations that are close in terms of political geography.