Suppose the Johansen Cointegration test suggests (and supports) one cointegrating vector among the variables while the lag length is 2 but no cointegrating vector at lag length 4. But the unrestricted VAR process says that the max. lag length is 4. Should I follow the Johansen test results for further analysis, or, I follow the unrestricted VAR process’s lag length?
You might also consider using an information criterion, such as AIC, BIC or HQC, to determine the lag length.
Further to the previous posting, an ARDL or NARDL model might be a useful alternative approach to Johansen's cointegration procedure.
There are several problems: the test assumption is very much important in cointegration test, especially the deterministic term. For instance, if the theory suggests to include an intercept in VAR process, then if I think that there will be no intercept in CE , then I may get my desired result for the lags less than the suggested lags. (by AIC, BIC etc). But, If I think that there will be an intercept in CE, probably I may not get any cointegrating vector for any lag. This is the problem that I encountered now. Interestingly, most of the articles do not show the test assumptions, they only show trace test/eigen value test results or the number of cointegrating vector.
I assume that there must be some articles regarding this issue. Is there any paper?
N.B: ARDL assumes the mixture of I(1) and I(0) variables. In my exercise all variables are I(1). Can I do the same thing by ARDL or NARDL?
If you are interested in Johansen and how you can test your first paragraph then I shall suggest Enders Applied Econometrics book but the use of Pantula principle suggested by Johansen can also help. In that case, see a small book by Richard Harris.
Even though all are I(1), yes you can use ARDL.
To the best of my knowledge, I didn't find any significant differences in results when I used theses criteria to determine the most suitable lag.
Then, no problem Ibrahim. The key point is not having serial correlation and your RESET looks good. If it is okay with you send me ur output.
I also face the same problem in choosing the lag length, when I use unrestricted VAR, all the criteria suggested lag 1 but when I use this lag in running Johannsen cointegration, I fail to reject the null hypothesis of no cointegration, though it works well when I use lag 2 which is not suggested by any of the criteria; which lag is more suitable and why is this happening?
Please assist
You do not blindly use those rules. If using lag 1 results in cointegration but lag 2 suggests cointegration, you simply report the 2 results but now focusing on lag 2 in your discussion .Longer lags like 2 is more meaningful here
@Chuck A Arize thank you very much, should I state any reason for me to select lag 2 instead of the one selected by the criteria?
Suleyman Abas
No you do not really need to but if you want " Further experimentation suggests th at lag 2 yields results that are economically more meaningful than those of lag one since longer lags are preferable than shorter lags, see Shin, Harris and Inder and see International Review of Economics and Finance - I have over 10 papers there and it is rated A+ journal by Thomsom. Most importantly, see 2008 paper I wrote with Osang and Slottje (p38-39).. See also, Angelos Kanas and Georgios P. Kouretas (2005) in the same journal, Also see Johansen and Juselius 1992page 215) I have given you lots of my time today!!
@ Chuck A Arize
Really you did Sir, and thank you very much Sir, I really appreciate your answers
Be blessed abundantly
Regards
Suleyman