Hi there,
I'm doing panel cointegration estimations, comparing group-mean with weighted-pooled method. Something weird is happenning with the fitted value when using group-mean method. It's off-scale compared to the values of the dependant variable. This does not happen with the weighted-pooled method. Also, it doesn't matter which econometric estimator I use (DOLS or FM-OLS).
The off-scale fitted values happens for all 30 countries in the panel. The graph for one country below is representative of the problem.
Blue line: dependant variable, red line: w-p fitted value, green line: g-m fitted value
Thanks for your help.
fitted value off-scale - panel group-mean method
Moderators: EViews Gareth, EViews Jason, EViews Moderator, EViews Matt
-
- Posts: 101
- Joined: Tue Jun 29, 2010 7:09 am
-
- EViews Developer
- Posts: 2671
- Joined: Wed Oct 15, 2008 9:17 am
Re: fitted value off-scale - panel group-mean method
Can you post your data?
-
- Posts: 101
- Joined: Tue Jun 29, 2010 7:09 am
Re: fitted value off-scale - panel group-mean method
Please find it attached. Thank you!
- Attachments
-
- fxfv_panel.wf1
- (1.64 MiB) Downloaded 227 times
-
- EViews Developer
- Posts: 2671
- Joined: Wed Oct 15, 2008 9:17 am
Re: fitted value off-scale - panel group-mean method
One thing to remember is that the grouped estimator estimates completely separate models for each of the cross-sections and then averages them (the only coefficients that are allowed to be individual specific in the final estimates are the deterministics).
The mean coefficients can be quite different than the grouped estimator which can lead to odd residuals patterns (this is true for any between estimator in a panel context). This is even more likely to occur in a case where you have large numbers of coefficients in your cointegrating equation. You have 6 variables in your system and there is a pretty large variance between the individual coefficients.
Note that in contrast, the pooled estimators have constraints on them in that large residuals for an individual will show up in the objective function and that the coefficients for the pooled estimator are quite different from the grouped estimator. The negative R-square in the latter is, I believe, related to this general issue. That is, the fitted values are worse than just using the individual deterministic. Again, this can happen since the grouped estimator does not optimize with respect to the overall residuals.
The mean coefficients can be quite different than the grouped estimator which can lead to odd residuals patterns (this is true for any between estimator in a panel context). This is even more likely to occur in a case where you have large numbers of coefficients in your cointegrating equation. You have 6 variables in your system and there is a pretty large variance between the individual coefficients.
Note that in contrast, the pooled estimators have constraints on them in that large residuals for an individual will show up in the objective function and that the coefficients for the pooled estimator are quite different from the grouped estimator. The negative R-square in the latter is, I believe, related to this general issue. That is, the fitted values are worse than just using the individual deterministic. Again, this can happen since the grouped estimator does not optimize with respect to the overall residuals.
-
- Posts: 101
- Joined: Tue Jun 29, 2010 7:09 am
Re: fitted value off-scale - panel group-mean method
Thank you for your reply. If I understand correctly your explanation, the off-scale is a result of EViews using the mean betas to get the fit. Since a country-specific beta can be quite different from the mean, the resultant country-specific fit can be "off". I manually calculated fitted values using the country-specific betas and constant (obtained with "panelcoefs"), and the fitted value becomes "on-scale". I'll opt for the weighted-pooled method; your exlplanation help to justify my choice. Thanks again.
-
- EViews Developer
- Posts: 2671
- Joined: Wed Oct 15, 2008 9:17 am
Re: fitted value off-scale - panel group-mean method
That sounds right to me.
Who is online
Users browsing this forum: No registered users and 34 guests