© 1994 by Biometrika Trust
MISCELLANEA |
Priors and likelihoods that are arbitrarily sensitive to outlying observations
RAND P. O. Box 2138, Santa Monica, California 904072138, U. S. A.
For location parameters the posterior can be extremely sensitive to the functional forms of the prior and likelihood when there is an extreme outlying observation. Conditions on the prior, likelihood and set are provided such that as one observation goes to infinity the posterior probability that a location parameter is contained in the set goes to zero. This is of particular interest when the set ranges over intervals with a finite upper bound. The conditions for this extreme reaction to an outlying observation cover all Box-Tiao densities with a shape parameter greater than one.
Key Words: Bayesian inference Outlier Outlier-resistance Robustness