I have 20 observations, time series data, lag order selection criteria (FPE AIC HQIC SBIC statistics). These show that I can choose 4 lags, co-integrated equations based on max trace statistics, showing that there are 4 cointegrated equations in the system. Taking into account that the sample size is small, what should be the number of lags and cointegrated equations(rank) for running VECM?
Significance of the coefficients does not change much while I am trying different number of lags and co-integrated equations.
How many right hand side variables do you have in your equation? Even if it's just one, then four lags, four lagged dependent variables, and a constant term would eat up 10 degrees of freedom, leaving you with just 10. I am not sure how meaningful any results would be.
I have 5 right hand side/independent variables. I check with 5 then 3 then with just one
5 RHS variables excluding lags? I don't see how you can do very much that's meaningful at all.
John do you know how to run ECM for panel data what will significantly increase number of observations.
Thank You
I am using STATA. it seems that Stata is not the most appropriate software for Panel VECM or is it possible to do it with Stata?
Dear Balázs,
Could you please suggest a software that can deal with panel ECM and commands required to run panel ECM at that software?
Thank you very much
Even with installing required additional files I could only check for cointegration through xtwest command. Could you suggest how to get these do ado files and run ECM for panel using STATA? Thanks
Hi Davron,
Eviews should be able to do it easily. After running panel cointegration, from proc, you may select unrestricted VECM model and you should be able to find error correction component. Good luck.
Murshed thanks. Does it require installation of extra components?or is it a standard command in Eviews 8?
Dear Murshed I just installed Eviews8 and tried to do ECM it does not require any additional installations as you said it is very simple thx.
Dear Murshed,
Thank you for your useful answer to Davron's question which was about lag selection issue.
Do you think that is that possible to estimate Panel VECM in EVews 8? I mean, technically we can do it, by following the same procedure that we have in time series framework. However, my concern is about, how to treat issues such as fixed effect, random effect as well as inclusion of instrumental variables, choosing different estimator rather than OLS for Panel VAR in Eviews 8?
I really appreciate your answer and thank you for that in advance.
You have data problem .Increase the sample size as much as possible.
I like to complement the last comment. What is the reason for the small sample? It is always better to have a longer period for time series analysis to improve the degree of freedom. It will also help you to do more robust analysis like checking for structural break etc.
The lag length should be determined on grounds of the corresponding VAR in levels. It should give meaningful results if the series are cointegrated. The SBC is the better choice: It has the stronger penalty term, and you are interested in smaller models.
You have a vvery limited sample size. Be sure that you have the annual frequency, at least. If you have monthly data, for example, a cointegration analysis would not make huge sense, as you examine a long run phenomenon.
You should also test for unit roots in the variables. Otherwise you run the risk of obtaining trivial cointegrating vectors. They will include only a single variable, i.e. the stationary one
If you are using yearly data then lags can be decreased.
On the other hand there are some methods through which u can increase your data set
it is better to employ ARDL instead of VECM when the number of observation is small
i have a question . i have an ECM estimation output based on an ARDL model but am confused on which one to choose as the short run coefficient:
D(Debt) -0.006(0.019) and
D(Debt(-1)) 0.055(0.019)*** where values in parenthesis are std errors
which one is the real short run coefficient ?
For lag selection with an Error Correction Model (ECM), use information criteria like AIC or BIC to determine the optimal lag length for the integrated variables in the model.
20 observations is a very small sample, particularly if the data is quarterly (monthly is even worse). You do not mention the number of variables but 4 cointegration relations seem too many for the usual macroeconomic setting. Possibly only one or two may be economically interpreted.
4 lags appears to be balanced for quarterly data and for such a small sample but a data dependent method is better. You may use AIC or GTS (general-to-specific). BIC is great when the purpose is forecasting; otherwise it usualy leads to dynamically incomplete models (and serially correlated errors).
- Badges
- Science topic
- Similar topics
- Books