POOL: unbalanced SUR approximation

[Raffalovich]

Running EV_8. The estimate -->specification window for the POOL procedure has a check-box for unbalanced SUR approximation. I cannot find any mention in help or the documentation files. Pls refer me to technical details and references.
Thanks,
--Larry

[EViews Glenn]

Sorry, it was a feature that we added and the docs weren't updated.

Unbalanced SUR is sort of an odd duck. The basic idea is that you are orthogonalizing the values within a cluster using say the Cholesky factor of an estimate of the inverse of the cluster covariance matrix. In balanced cases, there are no issues with how one performs procedure.

In unbalanced cases, there is a conceptual question of how to orthogonalize within an unbalanced cluster. The way that most people handle this is to simply zero out the "values" for the missing cases and use the overall Cholesky factor applied to the expanded data. Effectively, the missing observations receive no weight within the orthogonalization despite the fact that the Cholesky weights are generally non-zero. An alternative, which is enabled when you select the check-box, is that you recompute the Cholesky factor for the inverse of the submatrix of the cluster covariance corresponding to the observed data. Then the weights are applied directly only to the observations that are present.

EViews originally supported the former method. We offered the latter option to allow people to match results to computations using the different method.

I don't believe you'll find references to any of this since the unbalanced SUR isn't discussed in the literature. Personally, I'm not a fan of doing SUR in an unbalanced setting, preferring instead to simply use robust coefficient covariance methods which don't face this issue.

I hope that this answers your question.