This paper investigates local models; credal sets are associated only to marginal or conditional nodes in a network. A different type of Quasi-Bayesian network can be defined through global perturbations acting on the whole joint distribution. Several classes of distributions can be used to define such global perturbations (Appendix), but some of them are advantageous from an algorithmic point of view. The epsi-contaminated, constant density ratio, constant density bounded and total variation classes lead to robust inferences whose complexity is identical to the complexity of standard Bayesian inferences (the algorithms are presented in ). Future work will reveal whether such global models are useful for practical robustness analysis.
Fri May 30 15:55:18 EDT 1997