Title: A Bayesian Approach to Multistate Life Tables for Use in Social Epidemiology
Speaker: Scott Lynch, Princeton University
Date: March 27, 2006 12:00 - 1:30 pm
Location: DIMACS Center, CoRE Bldg, Room 431, Rutgers University, Busch Campus, Piscataway, NJ
Abstract:
Healthy Life Expectancy (HLE)?the length of remaining life one can expect to live healthy?is a summary measure that combines mortality and health information to produce an overall picture of quality of life. HLE is potentially an important measure for social epidemiology, the key concern of which is to examine social factors that produce health inequalities. However, to date, HLE's usefulness and use in social epidemiology has been limited. The estimation of HLE typically relies on multistate demographic methods, which are limited in (1) their ability to include covariates, and (2) their ability to capture uncertainty when sample data, rather than population data, are used. Over the past several years, I have developed a method that overcomes these limitations when panel data on transitions between states are available. In a nutshell, the method involves (1) estimating a multivariate hazard model using Gibbs sampling to produce samples from the posterior distribution of the model parameters; (2) combining these parameter samples with specific covariate values to produce distributions of transition probability matrices; (3) generating distributions of multistate life tables from these matrices; and (4) summarizing these life tables and performing statistical comparisons between groups using these distributions of life table quantities (like HLE).
This method has proven to work very well when panel data are available. However, the most commonly available data on health are cross-sectional. In this presentation, I show how this panel method can be extended to handle cross-sectional data. Such data have their own limitations, including (1) that mortality and health data must generally come from different sources, (2) that covariate specificity in mortality data is often much coarser than that in health data, and (3) that no transitions are actually observed, yielding an ecological inference problem that adds additional uncertainty to our estimates of HLE. In this presentation, I show how these issues can be addressed, and I compare results obtained using this new method with results obtained using Sullivan's method, the most commonly used method to estimate HLE when cross-sectional data are all that are available.
see: DIMACS Computational and Mathematical Epidemiology Seminar Series 2005 - 2006