ABSTRACT
In the specific cases, one may suspect a correlation between errors but have no structure to suggest a parameterized form such as serial correlation or compound symmetry. The problem is that there are too many pairwise correlations to be estimated and not enough data to do it. The error will have a variance in proportion to the number of voters because it will be like a variance of a sum rather than a mean. By using as many parameters as data points, the readers can fit the data exactly. Very little is achieved by doing this since they learn nothing beyond the data and any predictions made using such a model will tend to have high variance.
