coefficient covariance matrix panel least squares

[harry_tuttle]

Hi everybody

I'm using Eviews 7.2. I'm more of a practical econometrician and not that proficient in the technical background.
My sample: N (202) > T (19).
Notation: i = 1,...,N (cross section), t=1,...,T (time)

I'm doing a panel least squares estimation and I'm using the White period coefficient covariance matrix estimator for robust inference. Is the White period estimator implemented in Eviews the same as the one proposed in Arellano 1987 ("Computing Robust Standard Errors for Within-groups Estimators", attached)? The equation in User Guide II (35.33) looks pretty much the same as the equation in the paper (4), except for:
- the degress of freedom adjustment in Eviews
- the two additional sigma signs in the terms sum(Xi'Xi) in Eviews -> Is sum(Xi'Xi) (in Eviews) equal to X'X (in the paper)?
May I quote Arellano (1987) as the source for the White period estimator implemented in Eviews?

Did I get this right:
- The White period estimator is robust against heteroskedasticity along the cross section dimension for a fixed t: V[uit|X=xit] not equal for different i for a fixed t.
- The White period estimator is robust against autocorrelation along the time dimension for a fixed i.
- For N > T it is common practice to use the White period estimator in favor of the White cross section or White diagonal estimator.

Thanks a lot for advice.

Cheers, Harry

[EViews Glenn]

Correct on all. The asymptotics in the Arellano/White-period require N -> infinity.

(P.S. - You sound like an econometrician to me ).

[harry_tuttle]

Hi Glenn

Thanks a lot for the quick answer! Well, my matrix algebra could be better...

[fanis]

and what if T=N?

thanks in advance

[EViews Glenn]

Then this particular estimator is problematic since the asymptotics don't work. You can probably get away with T->infty at a slower rate than N->infty (Wooldridge discusses just this issue), but T=N clearly violates the conditions.

[fanis]

thanks!

[EconometricAarhus]

I know that it is an old post but I have a question:

Harry_tuttle says: "For N > T it is common practice to use the White period estimator". Why is this?
And do we have a source/academic article on this?

[EViews Glenn]

The asymptotics in the Arellano/White-period require N -> infinity.