A0675
Title: Using Bayesian methods and the singular value decomposition for fast, scalable demographic estimation and forecasting
Authors: Junni Zhang - National School of Development, Peking University (China) [presenting]
Abstract: Statistical demographers estimate and forecast detailed age-sex profiles for quantities such as mortality, fertility, migration, health expenditure, or labor force participation. Increasingly, profiles are required for many combinations of geography, ethnicity, education status, or other stratifying variables. The dimensions of the resulting models can easily become large. However, demographic processes often have highly regular age-sex patterns. The number of parameters required to accurately represent age-sex patterns is typically much smaller than the number of age-sex categories. Statistical methods are developed that take advantage of these regularities. Singular value decompositions are applied to high-quality data from international databases, and the results are used to formulate informative prior distributions for age-sex profiles. These prior distributions are then embedded into a larger hierarchical model. Inference is done using a Laplace approximation, as implemented in R package TMB, and is extremely fast, even with thousands of parameters. It is illustrated using examples from analyses of mortality rates and labor force participation. The methods are implemented in an open-source R package.
