Bug with panel data

[PhilW]

When using panel data (fixed effects) we found that in sample forecasted values are inconsistent with fitted values. In sample they should be equal. We believe the forecasted values to be correct. In some cases, for some cross sections, all fitted values are less than the actuals so the errors for the cross section are all of the same sign. This isn’t right. We found that a solution is to use @expand() on the cross section variable rather than structure the data as fixed effects. This works in that the forecasted and fitted values are equal AND equal to the forecasted values when the panel data method is used. However the standard errors are very different between the approaches, indicating that AT LEAST one of them is wrong.

This is a significant bug and needs to be fixed as most of our models use this type of structure. Short term would you please tell me which of the two approaches gives correct standard errors?

[EViews Gareth]

Could you state the version number and build date of your copy of EViews (both available from Help->About EViews)

[PhilW]

Eviews7 Jan 14 2010 build

I sent an email to support on 11/18 at 4:49 but never heard back.

[startz]

You might want to update your copy of EViews and see if the problem is still there.

[EViews Gareth]

Your email was replied to at 13.54 (PST) on November 18th, with essentially the same reply as Startz - your copy of EViews 7 is woefully out of date.

[PhilW]

Updated to Nov 11 2010 build. This is the most current. All problems and bugs persist exactly as previously described. As a paying customer I ask for you to please devote effort towards a resolution.

[EViews Gareth]

Could you provide us with your workfile and instructions on how to replicate the problem?

[PhilW]

I emailed an example data set.

[EViews Glenn]

Thanks. That helps a great deal.

It appears that there is a bug in the exporting of residuals for the AR model with fixed effects. Note that this code is separate from the estimation code so that all statistical results (SSRs, etc.) appear to be correct, but we are outputting the wrong residuals to the workfile (which is why your actual/fitted/resid table/graph looks funny).

I have a fix working in my test code which I'm going to put through some paces to make certain that I haven't caused any other problems. It should show up shortly in a patch.

Our apologies for the inconvenience. Let me know if you have any other questions.

[PhilW]

Thanks. That is one of the two related issues.

If you look at the parameter estimates of the two models you see that they are all exactly the same. In the Dummy_Variables model @expand on the cross section variable acts as fixed effects coef. Despite all the coefficients being identical the standard errors and t-stats are hugely different with much higher confidence in the Dummy_Variables model. This would have a significant impact on model selection. I tend to believe the standard errors in the model with fixed effects but this implies a problem with the other model. Maybe there's something wrong that happens when a dataset is structured as Panel but the model isn't.

[EViews Gareth]

In your panel_options equation you've set the covariance method as "White Cross-section". In the dummy_variables equation, you've set the covariance method as "ordinary". If you switch them to use the same method, you'll find the standard errors match.

[PhilW]

Yes they do. Thanks.

[EViews Gareth]

The latest patch should fix the first issue for you.