ABSTRACT
In this chapter, the authors discuss the predictive distributions to make inference on unobserved data and also to illustrate the shape of relationships in statistical models. Prediction enables to characterize how well model represents data-generating mechanisms. In general, studies have shown deviance information criterion to be valid for non-hierarchical and non-mixture models, cases where effective number of parameters is much less than the sample size of the data, and when the posterior mean of the parameters is a good summary of the central tendency of the posterior distribution. Predictive scores like deviance information criterion that depend on estimated corrections for optimism are useful because the authors can calculate them easily using within-sample data.
