A preliminary-test estimator for the error variance in the one-way random model is considered. The optimum levels of significance for the preliminary test are obtained based on a regret function. A pooling procedure for estimating the error variance, based on weighting functions, is also considered. A comparison of these estimators is made.