Why do we get Negative Variance Components in ANOVA

The usefulness of analysis of variance(ANOVA) estimates of variance components is impaired by the frequent occurrence of negative values. The probability of such an occurrence is therefore of interest. In this paper, we investigate a variety of reasons for negative estimates under one way random effects model. It can be shown, through simulation, that this probability increases when the number of treatments is too small for fixed total observations, unbalancedness of data is severe, ratio of variance components is too small, and data may contain many outliers.