Bayesian Perturbation Diagnostics and Robustness

A Bayesian analysis may depend critically on the modeling assumptions which include prior, likelihood and loss function. A model that has been judged adequate in previous more or less similar situations may be assumed to be the standard. However one ought to consider the effect of perturbing the standard model in potentially conceivable directions especially if graphical procedures indicate the standard may only be marginally adequate. We discuss a variety of perturbation models and Bayesian diagnostics that can be helpful in a local or a more global analysis of the robustness of the sample.