Fast Bayesian Estimation of Disaggregated Demographic Rates by Borrowing Strength from International Databases

John Bryant, Bayesian Demography Limited
Junni Zhang, Peking University

Contemporary demographers work with datasets that are disaggregated not only by age, sex, and time, but also by variables such as region, ethnicity, and educational attainment. As the detail and size of datasets grows, statistical models can become slow and unreliable. One way to address these problems is to take advantage of regularities in the age-sex patterns of demographic rates. Adding information on demographic regularities to a statistical model can make the model faster and more robust. We describe new methods that use the Singular Value Decomposition to extract information on age-sex patterns from demographic databases such as the Human Mortality Database or the Human Fertility Database, and then incorporate this information into a family of general-purpose Bayesian statistical models. We illustrate our methods using regional mortality rates in Sweden. Fitting a model to a dataset with 141,400 combinations of age, sex, region, and time, and with a median death count per cell of 1, takes around 50 seconds. The methods are implemented in the R package bage, available on CRAN.

Keywords: Small area estimation, Digital and computational demography, Multi-level modeling , Data and Methods

See extended abstract.