We introduce a class of multivariate dispersion models suitable as error di
stributions for generalized linear models with multivariate non-normal resp
onses. The models preserve some of the main properties of the multivariate
normal distribution, and include the elliptically contoured distributions a
nd certain other known distributions as special cases. We give explicit met
hods for constructing multivariate proper dispersion models. This is exempl
ified by constructing multivariate gamma, Laplace, hyperbola, and von Mises
distributions. (C) 2000 Academic Press AMS subject classifications: 62H05,
62H11.