Methods are proposed for the analysis of data from a multivariate normal distribution when observations are missing completely at random on some of the variates. Park (1) proved the equivalence of the solutions given by maximum likelihood and generalized estimating equations when data are complete and an unstructured covariance is assumed. He suggested that generalized estimating equations may be used if sample sizes are large relative to the amount of missing data and the estimated covariance matrix is positive definite. We give several examples indicating that the estimating equations give results similar to those of maximum likelihood when smoothing of the covariance matrix to eliminate nonpositive definiteness is not encountered as, for example, under an assumption of exchangeable correlation. Generalized linear models are formulated that are appropriate for a wide class of experimental plans. Contrasts among the means are investigated in terms of robust estimators of the contrasts and their covariance matrices. Large sample chi-square statistics for testing relevant hypotheses are discussed.
展开▼
