We represent random vectors Z that take values in R-n - {0} as Z = RY where
R is a positive random variable and Y takes values in an (n - 1)-dimension
al space Y. By fixing the distribution of either R or Y while imposing inde
pendence between them, different classes of distributions on R-n can be gen
erated. As examples. the spherical. I-q-spherical, b-spherical and anisotro
pic classes can be interpreted in this unifying framework, We present a rob
ust Bayesian analysis on a scale parameter in the purr scale model and in t
he regression model. In particular, we consider robustness of posterior inf
erence on the scale parameter when the sampling distribution ranges over cl
asses related to those mentioned above. Some links between Bayesian and sam
pling-theory results are also highlighted. (C) 2001 Academic Press.