A general model for two-level multivariate data, with responses possib
ly missing at random, is described. The model combines regressions on
fixed explanatory variables with structured residual covariance matric
es. The likelihood function is reduced to a form enabling computationa
l methods for estimating the model to be devised.