Abstract
Linear combinations of variance components for which there exist unbiased, non-negative quadratic estimators are characterized. It is shown that the ‘error’ component in ANOVA models is the only single component which can be so estimated.