In the framework of Bayesian approach, though we can select various prior distributions according to the situations, many models which should be evaluated exist. However, we think that the diagnosis methods for these models have proper diagnostic situations. In this paper, we consider two diagnostic methods that focus on prediction: the Bayesian predictive information criterion (BPIC) and the prior and posterior predictive checking approach (PCA) and conduct some simulations for the purpose of clarifying the feature of these methods and suggesting the effective diagnostic situations. As the results, regardless of whether the prior mean was true or not, BPIC showed the low values on the occasion with strong prior information, but PCA showed the high predictive checking probabilities on the occasion with weak prior information. So, it might happen that the model with non-true prior mean was selected by simulation setting. To select a proper model, it is necessary to find the performance of the model diagnoses in the situation before model evaluations.
View full abstract