Panel cointegration estimations - possible with Eviews 7?

[EViews Glenn]

Even EViews 6 has the built-in panel cointegration tests using FMOLS as proposed by Pedroni. There are some other panel cointegration tests as well, but not ones that use DOLS.

Note that EViews 8 supports estimation of panel cointegration models using various methods including the group mean FMOLS and DOLS estimators.

[stoddj]

Thank you, Glen. I'll have to look at Eviews 8.

[stoddj]

I may be a bit dense here, but it appears you are saying that the Phillips-Perron (PP) statistics, both panel and group, are based on the FMOLS estimate. Is that right? Maybe I missed it, but I didn't see that mentioned in the documentation. (This appears unchanged from EViews 7 to .

[EViews Glenn]

I think I was being unclear.

EViews 6 and 7 allow you to test cointegration in panel workfiles using the Pedroni and the Kao Engle-Granger based-methods, and the Fisher combined Johansen test methods. The Pedroni methods use the same theory as panel FMOLS estimation so in that respect I view them as tests based on FMOLS.

To the best of my knowledge, no one really does the cointegration tests using residuals from first-stage FMOLS or DOLS estimation, though clearly you could since all that is required are consistent long-run equation estimates. Since the first stage long-run regression is super-consistent, the literature doesn't seem to feel that it is necessary. Whether the bias-correction would help or not really is an empirical question for which there is no (to my knowledge) evidence.

EViews 8 supports Pedroni panel-FMOLS estimation and Kao and Chiang, Mark and Sul panel-DOLS estimation. You could always do a panel unit root test on the residuals from these estimation methods, but you'd then have to adjust the critical values yourself.

[stoddj]

Thanks a heap, Glen. I'm frankly surprised that all 11 Pedroni tests can be viewed as tests of cointegration in FMOLS regressions. But looking back at his 1996 and 2000 papers with titles that begin "FMOLS," and especially the 2001 REStat paper, this makes sense. I would never have guessed this from the EViews documentation -- but then I guess I can't expect it to cover every scrap of econometric theory!

[EViews Glenn]

They're not really tests of cointegration using FMOLS as they use standard OLS (with fixed effects or individual trend coefficient) estimates in the first stage, but they do use the same principles and asymptotic results that are used to derive the FMOLS estimates.

[stoddj]

I'm concerned, however, that asymptotic results may not help me much, given my small sample size. I have too look at the small sample size properties of these estimators, which as you know are all over the place. Many of them are quite 'over-sized' at low T (so that rejection of the null of no-cointegration is more likely than the nominal alpha value).

Therefore I'd like to trouble you with the following naive question. Do you have even a rough sense on how much programming would be necessary to convert the EViews panel unit root tests into Pedroni cointegration tests? I already have all the residuals from panel DOLS and FMOLS estimates (which I built up in EViews 7 before knowing that 8 would come out with them!) Thanks again for all your help.

[EViews Glenn]

Do you mean Pedroni tests using the FMOLS and DOLS residuals? That's hard to say. EViews offers all of the long-run covariance calculations you'll need. The tests aren't hard to compute, but neither are they trivial.

[stoddj]

OK, Glen -- that's the sort of basic assurance I was looking for. Guess I'll just have to dig in to the original papers and try to program it. Thanks again.

[May21]

Hi,
I have estimated my model using group-mean FMOLS technique on Eviews student version 8.In order to cross check the results,I also estimated the model using same technique on winRATS version 8.But my results from the two softwares are drastically different for two of the coefficients for one of the countries but they are quite close to each other for other countries.
Could you throw light on this?

Please reply soon!

Thanks.
Regards.

[May21]

Hi,
I have estimated my model using group-mean FMOLS on Eviews version 8. Eviews 8 reports coefficients of both the individual cross-sections as well as panel as a whole but it reports t-statistics for only the panel and not for individuals cross-sections.Is there any way that one can retrieve the t-statistics or p-values of individual coefficients?
Another thing I want your help on is the following: I estimated my model on RATS version8 also because it reports individual cross-sections' t-statistics as well.However,when I compared the results from two softwares,I noticed that most of the times the results were very similar but for one of the members of the group,the coefficients of two variables took reverse signs in the two softwares.I checked my data carefully and re-estimated the model but all in vain.
Can you throw some light on it?
Please reply soon!

Thanks,
Regards.

[EViews Glenn]

t-statistics are not available.

For the pooled estimators, EViews treats the individual coefficients as nuisance parameters so that standard errors are not available.

For the grouped estimator, you could estimate the individual FMOLS equations for each cross-section, by setting a sample to include only that cross-section. That will give you all of the standard errors.

As to the differences. That's entirely possible as FMOLS requires estimation of long-run variance. That computation has a lot of components and differences in settings can lead to substantive differences which, in turn, can affect the coefficient values.

[May21]

Can you please elaborate a little more as to what you mean by setting the sample so as to include only one cross section?Do you mean that I estimate a model for each cross section as a time series analysis?

Thanks.
Regards.

[EViews Glenn]

Yes.

[May21]

Hi,
I am facing a strange problem as follows:
I set my workfile for the time period 1990-2014 for 7 cross sections and did panel FMOLS estimation for the sample size 1992-2014.When I made a new workfile which had data staring from 1992-2014 and did the same estimation,I got different results.In the output window,Eviews reports the sample(adjusted) to be 1992-2014 in case of first estimation and 1993-2014 in case of second .Could you please let me know why this this happening?


Thanks.
Regards.