Multivariate Regression with Monotone Missing Observation of the Dependent Variables

Multivariate regression is discussed, where the observations of the dependent variables are (monotone) missing completely at random; the explanatory variables are assumed to be completely observed.We discuss OLS-, GLS- and a certain form of E(stimated) GLS-estimation.It turns out that (E)GLS-estimation uses the preceding dependent variables in a well-structured way.In case of normality, ML-estimation coincides with (E)GLS-estimation.We include (sets of) MANOVA-tables enabling us to perform exact tests on the coecients based on a (new) generalized Wilks' distribution.Only the very special case of the constant as sole explanatory variable has been treated in the literature so far: our model incorporates this missing data problem.