Granger causality using panel data

[evan]

Dear sir/madam,

i am trying to perform a Granger causality test using panel data with 7 different cross identifiers. some of the people who participate in this forum say that this can not be done in Eviews 7. they say that this is possible only in Eviews 8. could please somebody confirm this for me. also is there some way i can get an electronic copy of Eviews 8 to see exactly how this is done. thank you very much. Best, Evan

[EViews Gareth]

EViews 8 added the Dumitrescu-Hurlin Granger Causality Test for panel data.

EViews 8 is available for electronic purchase, if that's what you're asking.

[evan]

Dear Gareth,

thank you very much for your reply. actually i was wondering if i could have an electronic copy of the Eviews 8 manual to make sure that this is exactly what i want to do. Do you know where i could find an electronic copy? i came across something which is called Eviews 8 illustrated but it does not describe the Dumitriescu-Hurlin Granger caulality for panel data anywhere. then i could place an order for the new software. currently I have Eviews 7. thank you.

[EViews Gareth]

The PDF versions of the EViews 8 Documentation are available from the Downloads page:
http://www.eviews.com/download/download.html

[Kavorka]

Hi!
I would draw different conclusions if there is granger causality between X1 and X2, in the sense that increases in X1 Granger causes increases in X2, or if increases in X1 would Granger cause decreases in X2.

When running Granger Causality Tests in a panel (or in a simple VAR system) - how do we retreive all the estimated coefficients (in version 8.)?


I understand that there are other measures such as impulse response functions for general VAR-models, but it is interesting to see what models one really are estimating. Otherwise it will be a black box, and we cannot interpret how credible the results are - if we only look at a neutral F-statistic.

It is very relevant if there is a positive or a negative Granger causality relationship between the studied variables. Even if there are many lags it is relevant to what sign most coefficients show.

If it is not possible to find this, I suggest that you include this feature since it is absolutely crucial for the conclusion (but also for the intuitive understanding of the estimated relationship). I cannot really tel, but in my mind it should be really easy for you to include this feature. It could be something similar as to how one is retreiving the fixed effects in a panel data model in EViews

[EViews Gareth]

The problem is that there are, potentially, thousands of coefficients. They're pretty meaningless. If a user really wants to see them, given that Granger Causality works by computing simple least squares regressions, they can always perform the regressions themselves.

[Kavorka]

I do not agree that it is not interesting to see if we have a positive or a negative Granger causality.

I understand that it is sometimes problematic to interpret if there are many coefficients... but the researcher can make that judgement by themself. If 50% are positive and 50% negative coefficients... and there are many coefficient... then maybe retrieving the coefficients is useless. However, if most coefficients are positive one may draw different conclusions compared with when most are negative.

If we ONLY get an F test I would say that the entire GC test often is useless. I am not interested to see if there is a relationship (if I have not idea if this is a positive or negative relationship). It only shows IF there is a relationship between X1 and X2, not if this is a positive or a negative relationship (which usually is what we want to discuss in our conclusions... despite that the test itself is formulated like a neutral F-test).

Moreover, for a bivariate VAR(1) model (for Granger causality) I would say that it is easy to see if it is positive or negative. If there are many coefficients I would look at the magnitudes and see if most of the coefficients with high magnitudes are positive or negative. If we do not have an opinion if there is positive or negative Granger causality I would not use it to draw conclusions in a study.

Since this would be easy to implement in EViews I would suggest that you should include it (even if the coefficients definitely should be interpreted with caution). More information is better than no information - regarding if it is positive or negative. I would say that 99% of those who are running a granger causality test are really interested if it is a positive or a negative relationship. If one simply assumes a positive, but if there is a negative, relationship the conclusions in the paper would be directly misleading.

If one has to calculate the models by themselves, then I do not see the point of including canned routines (with only partial information about what one is really doing).

[EViews Gareth]

I don't understand your point for the VAR example. The coefficients are shown in the VAR output...

[Kavorka]

Well, I mean if you have a normal time-series data set. Highlight two variables. Open As group. View>Granger Causality. Lags to include: 1

Pairwise Granger Causality Tests
Sample: 1985M01 2011M04
Lags: 1

Null Hypothesis: Obs F-Statistic Prob.
Y_FINLAND does not Granger Cause y_SPAIN 315 13.7157 0.0003
Y_SPAIN does not Granger Cause Y_FINLAND 0.51798 0.4722

In this arbitrary silly example. I can of course run a couple of completely new OLS regressions to get the coefficients, but then I could calculate the F-statistic in the same way by the unrestricted and restricted regression. If you have a canned routine, it should be able to provide all necessary information. E.g. it would be nice with an option where one tells EViews to present the models too.

I want to know if there is a positive or a negative relationship regarding how Finland Granger causes Spain. If I do not check this up, the F test is useless since it can both mean that there is positive Granger causality between Finland to Spain, or that it is negative Granger causality between Finland to Spain.

I understand that it is harder to interpret large systems, but I do not see the harm in including it (except that it increases your workload).

But, anyway I will not tell you what is most important to include in EViews, but I still think that it is relevant if a relationship is mainly positive or negative.

In my mind, more information gives the researcher/student a greater intuitive understanding of the overall relationship. I always want to understand what I am doing and analyze how the data behaves. Only a Granger causalily F test is black box in my mind, that should be combined with the coefficient estimates in order to analyze whether the model makes sense. But you are of course entitled to whatever opinion you want. If you for whatever reason do not want to include it, then you do not want to include it... and then I have to accept that

[EViews Gareth]

Ah, I thought you were talking about from the VAR object.