Handling Sampling Design in Bayesian Multistate Life Table Estimation
Scott M. Lynch, Princeton University
J. Scott Brown, Miami University
Over the last several years, we have developed two Bayesian approaches to multistate life table estimation that allow the construction of interval estimates of life table quantities. One method requires panel data; the other requires independent cross-sectional data for mortality and health. Both methods involve Gibbs sampling, and both implicitly assume the sample data are from a simple random sample, which is not the case with most panels or cross-sections. Here, we investigate the implications of ignoring sample design. We ask (1) whether sample design influences interval estimates to a substantively meaningful extent, and (2) whether the Bayesian approach can be adapted to handle design. We find that sample design increases interval estimates only slightly in most cases. We also show how the Bayesian approach can be easily adapted to incorporate most design effects via a bootstrap. We describe the procedure and give an applied presentation of its implementation.
Presented in Poster Session 4