Three alternative estimators of a multivariate normal covariance matrix are compared to the usual estimator (the sample covariance matrix) under two loss functions by means of a Monte Carlo experiment. For small and moderate sample sizes, the characteristic roots method of Stein (1975, 1977) and the correlation matrix method of Lin (1977) and Lin and Perlman estimator for a wide range of covariance structures, while the empirical Bayes method (Haff, 1980) offers less improvement.