weights in panel

[startz]

Is there a reason that no way is supplied to do simple weighted least squares in a panel dataset? It's odd that organizing data as a panel should eliminate options.

[EViews Glenn]

It creates computational issues for the usual fixed effects estimators. It could be supported in non-fixed effects cases, but then the estimation interface gets messy, especially if you want the currently supported weighted estimators to apply to the weighted data. We decided that since the demand for this was likely to be rather low and that people could always unstructure if they wanted, we'd leave out support until the inevitable groundswell of public opinion convinced us of the error of our ways.

[startz]

I haven't thought this through, but since the weighted dummies are still orthogonal can't you still use the Frisch-Waugh-Lovell theorem to sweep out the dummies efficiently. Just with some regressions instead of subtracting group means?

[EViews Glenn]

You're right. For one-way effects the orthogonality buys you a lot. I don't even think it's a regression, just be a weighted mean adjustment that's analogous to the mean adjustment. Probably also works for the spectral decomposition used in balanced two-way effects, though the weighted double projection might be a bit messy.

But now let's consider the same thought experiment for unbalanced two-way effects. My instinct is that it'll be much messier to extend the existing algorithm. Suppose we do get that working. Then we ask, do we allow various system-type weighted estimators on top of the weighted data? Probably should, and I don't think it'll be too difficult to do for the group weighted and correlated cases. Now how does the weighting change the various estimators for random effects? This requires some thought since we have several approaches and do a fair amount of bias adjustment in the various calculations...

...etc...

I don't want you to misunderstand my point. What you are asking for can be done. The basics probably aren't that hard. I suspect that much of the analysis would go through. I do think there's probably some internal redesign involved, but none that we couldn't handle. But there are a lot of issues, and all of the detail work and validation that would be required would be very time-consuming. Given finite development resources we decided that the opportunity cost of this feature was high relative to the benefit.

That doesn't mean we won't revisit the issue. Your suggestion is a good one and I will continue to give it some thought. It may even make one of Gareth's famous *things to consider* list. Oh alright, it'll make Gareth's *things to consider* list. But realistically, we are not short on things that we'd like to do and it won't be at the top of the list. But maybe you can convince us to move it up. :)

Besides, now that we have EViews Add-ins, we'd be happy to have you do the simplest version of one of these estimators and submit it ....One of the great advantages of Add-ins is that they can be designed to handle just the simple cases...