Extending Likelihood-based Algorithms and Inference to Nonidentified Bias Models

• Activity Number: 373
• Type: Invited
• Date/Time: Wednesday, August 10, 2005 : 10:30 AM to 12:20 PM
• Sponsor: Section on Statistics in Epidemiology
• Author(s): Sander Greenland*+
• Companies: University of California, Los Angeles
• Keywords: Bias ; Epidemiology ; Observational studies ; Risk Assessment ; Sensitivity Analysis ; Bayesian methods

Abstract

Approaches to nonidentified bias models include sensitivity analysis, in which fixed values for bias parameters are assumed and estimates are plotted against those values; Monte-Carlo sensitivity analysis, which average the sensitivity results over prior distributions using prior sampling; and Bayesian analysis. Although posterior sampling is a widely preferred approach for Bayesian methods, some samplers can be difficult to use with nonidentified models. An alternative that is well known in conventional settings (with identified models) is to treat the posterior as if it were a likelihood from an identified model and employ standard approximations to obtain percentiles and areas. This approach is highly intuitive when used with generalized conjugate priors, and can be easily implemented with common commercial software. In this talk, this approach is illustrated with data from a study of sudden infant death syndrome in which a standard EM penalized-likelihood analysis yields posterior percentiles and probabilities from a nonidentified misclassification model.