The RATS Software Forum

Dear Tom,
I have two questions to be asked as follows:
Does estimation of cointegrating vector by using Pedroni's group-mean FMOLS tell us anything about direction of causality between the variables in the model?

If not,then having done the tests for cointegration by Pedroni and estimations of the cointegrating vector by using group-mean FMOLS, can we also estimate a panel error-correction model in RATS so as to see the direction of causality?

Please reply soon!

Thanks.
Regards.

sanjeev wrote:Dear Tom,
I have two questions to be asked as follows:
Does estimation of cointegrating vector by using Pedroni's group-mean FMOLS tell us anything about direction of causality between the variables in the model?

No.

sanjeev wrote: If not,then having done the tests for cointegration by Pedroni and estimations of the cointegrating vector by using group-mean FMOLS, can we also estimate a panel error-correction model in RATS so as to see the direction of causality?

Yes. You would have to do joint tests on the lagged difference and on the lagged error correction term to test a null of no causality.

Dear Tom,
Thanks for the reply!
So,is there a program in RATS which enables us to estimate the VECM and do the subsequent tests?


Please help me!

Thanks.

You would base that on https://estima.com/forum/viewtopic.php?f=31&t=1480. Convert into a VECM by using the differences and including the lag of the FM estimated residuals and test the lagged other variables + the lagged residual to get a causality test.

Dear Tom,Thanks for your reply!
Could you please tell me about tests for heteroscedasticity and individual heterogeneity in RATS?

Regards.

Dear Tom,
I have seen certain papers reporting diagnostic tests for FMOLS estimations like White's heteroscedasticity test,Adjusted-R-squared,Misspecification test and LM test for serial correlation. Please find one on the link: http://faculty.smu.edu/millimet/classes ... 202010.pdf
Could you please tell me if we could do such tests in RATS after group-mean FMOLS estimation?

Please reply!

Thanks.
Regards.

Someone already asked about that. https://estima.com/forum/viewtopic.php? ... 1289#p9895. From what I can tell, those are econometric gibberish. FMOLS isn't least squares, so there isn't a meaningful R^2. (Perhaps they used EViews---I saw an example of panel FMOLS where it kicked out an R^2 of -4000, which shows how silly the whole concept is). Not only is it not least squares, but group mean FMOLS is an average of estimators which themselves aren't least squares. RESET is a specification test for....least squares, and I have no idea how you would even define that for cointegrating vector estimates.

Dear Tom,
I am doing Group Mean Panel FM estimations.I have a quwstion.I put an inetraction dummy for the crisis in my model.Is it right to have such a variable in the model?
Secondly,when I also put the original variable along with the inetraction dummy variable,it makes the matrix of regressors singular and hence nothing can be estimated.

Cuuld you please help?
Please reply soon! Its urgent.


Regards.

You already asked effectively the same question. Interaction with what?

TomDoan wrote:You already asked effectively the same question. Interaction with what?

I have a dummy variable which takes value 1 for the period 1997-1999 and 0 otherwise for two countries while it takes value 1 only for rest of the countries in the sample.I interact this dummy with one of the regressors in the model.When I run the regression with this interaction dummy along with other variables in the model,I get the results.But when I also add the original variable(the regressor with which I interact the dummy) along with the interaction dummy variable,the estimation doesn't go through.It reports singularity of the matrix as the problem.
Could you please help? Please reply soon!


Thanks.
Regards.

You probably have the time dummies on, which would be collinear with a time-based 1-0 dummy. At any rate, there is no theory of which I'm aware for how FMOLS works with what you're doing.

Dear Tom,
I am facing following problem while doing group-mean FMOLS estimation:
When I don't include the common time dummies,I get expected signs and significance of my variables while when I include the time dummies,the signs and significance of some variables get worse.
Also, the signs of some of the variables are not as expected for some cross-sections while I compare the individual cross-section results with the aggregated ones.Does it imply that the results are not robust?
Please help.

sanjeev wrote:Dear Tom,
I am facing following problem while doing group-mean FMOLS estimation:
When I don't include the common time dummies,I get expected signs and significance of my variables while when I include the time dummies,the signs and significance of some variables get worse.

It sounds like some of your variables are too similar across individuals. Time dummies remove a common component from each time period across individuals. If x(i,t) is approximately x(t)+small noise (i,t), then, no matter what the time behavior of x(t) (it could be trending, or could be almost randomly varying), the time dummies will take out the x(t) from all of them, leaving only a limited amount of information.

sanjeev wrote:Also, the signs of some of the variables are not as expected for some cross-sections while I compare the individual cross-section results with the aggregated ones.Does it imply that the results are not robust?
Please help.

That's not that unexpected---it depends on how big your T dimension is.

TomDoan wrote:

sanjeev wrote:Dear Tom,
I am facing following problem while doing group-mean FMOLS estimation:
When I don't include the common time dummies,I get expected signs and significance of my variables while when I include the time dummies,the signs and significance of some variables get worse.

It sounds like some of your variables are too similar across individuals. Time dummies remove a common component from each time period across individuals. If x(i,t) is approximately x(t)+small noise (i,t), then, no matter what the time behavior of x(t) (it could be trending, or could be almost randomly varying), the time dummies will take out the x(t) from all of them, leaving only a limited amount of information.

sanjeev wrote:Also, the signs of some of the variables are not as expected for some cross-sections while I compare the individual cross-section results with the aggregated ones.Does it imply that the results are not robust?
Please help.

That's not that unexpected---it depends on how big your T dimension is.

Dear Tom,
Thanks for replying!
Regarding the first response that you gave,does it imply that I should reconsider the specification of model and hence variables in my model?
Secondly,I have 35 annual observations and 8 countries in my sample.

Dear Tom,In continuation with my previous mail,does it imply that these variables are cross-sectionally dependent?

Thanks.
Regards.