A Novel Bayesian Method for Modeling Cause of Death by Type and Sub-Type

Alan A. Cohen, Université de Sherbrooke

Cause of death classification is problematic for demographers because it changes over time and is often inaccurate. Broad categories often miss important processes affecting only some sub-types, but narrow categories are hard to define accurately and require enormous sample size, especially for rare causes. Here, I present a general method for modeling cause of death that allows simultaneous estimation for nested categories of causes. The method relies on Bayesian estimation of a series of equations, one for each cause of death by the finest categorization. Equations in the same class or sub-class share a parameter accordingly, and simultaneous estimation of the equations means that the shared parameters have a much larger effective sample size. I present a number of variations and applications, including an example looking at effects of an exposure on cancer mortality generally and by site.