This paper complements the results of Toyooka and Kariya ({\em The Annals of Statistics, 1986}) by evaluating an upper bound for the covariance matrix of a typical generalized least squares estimator in a heteroscedastic model.