Analysis of Mortality Trends Using a Mixture of Reference Priors

• Activity Number: 248
• Type: Contributed
• Date/Time: Monday, August 5, 2013 : 2:00 PM to 3:50 PM
• Sponsor: Section on Bayesian Statistical Science
• Author(s): Michael Sonksen*+
• Companies: The University of New Mexico
• Keywords: noninformative priors ; nonparametric Bayes ; model averaging ; constrained parameters

Abstract

Forecasting mortality rates is an important concern for governments and insurance companies. We propose a Bayesian version of the Lee-Carter model for predicting future mortality trends. This model is capable of incorporating ordering knowledge about the mortality rates with respect to age. Possible knowledge could be the presence of an increase in mortality rate around the ages 15-20 (the so called "accidental hump"). A reference prior is developed for the mortality rates conditional on this ordering information. In the case where knowledge is not known, or is imprecise, we average over all possibilities. The resulting posterior distribution can be used to not only forecast mortality rates but examine if the knowledge about the mortality rates is plausible from the data. We illustrate the proposed model on a dataset of US male mortality rates and discuss extensions to include covariate information.