New Approaches to Model Selection in Bayesian Mixed Modeling
- Author(s): Christensen, Fletcher George William
- Advisor(s): Johnson, Wesley O
- et al.
Because the marginal densities corresponding to data modeled with generalized linear mixed models (GLMMs) usually lack closed-form expressions, model selection via existing tools like the deviance information criterion (DIC) can yield inconsistent results. We discuss why marginalization is preferable for the evaluation of competing mixed models, provide a new method for fast and accurate approximation of the marginal DIC for GLMMs, and demonstrate through simulation how numerical approximation of the DIC relative to our marginalization scheme gives more accurate model selection results than other numerical approximation methods for DIC. We also discuss some issues related to model selection in an analysis of longitudinal data collected to assess the effect of polycyclic aromatic hydrocarbons on hormone functioning in women who were attempting to get pregnant.