new problem with ARDL bound (new version) v2.1

[ecofin]

hi for all,
1-the error has been solved for the model with 2 variables but: when i estimate my model with two variables y,k (ADF, 5%, max lag 5, AIC, intercept) by EViews it choose best model ARDL(4,4) but in microfit v4.0 (max lag 5, AIC, intercept) it choose ARDL(5,5).
can you explane me why the difference in the resolt? i must use one of them
2-another error has been appeared with 3 variables (y,k,l, 5%, max lag 5, AIC, intercept): see the picture bellow, for microfit v4.0 i have choose (AIC, max lag 5, intercept).
can you solve this new problem and this error?

Kind Regards.

[EViews Gareth]

You might want to provide the workfile you're using.

[ecofin]

this is my work file.

[Yashar]

There are three points in this subject:
1- The difference between ARDL bound test and ARDL level(conventional)
2- The difference between Eviews and Microfit procedure
3- The number of observation

1- I haven't used Microfit. However, from the results you provided my guess is in Microfit RHS variables are in level. This means Microfit follows an ARDL in the level, whereas in ARDL bound test we have lagged of differeced variables. In other words in ARDL bound test is more like an ECM (look at the equation 4 in Dave Giles explanation http://davegiles.blogspot.com.au/2013/06/ardl-models-part-ii-bounds-tests.html)than ardl as time series structure. Coming from different assumption, they might have different lag structure. Which one you want to choose ?! that's totally depends on what you are doing, ARDL bound test or ARDL ?

2- Again from your result I am guessing that Mircofit starts the lagging process by lagging the main variable(LHS) then goes to other RHS variables. However, in Eviews we start with lagging last variables. The reason for doing so is, in ARDL bound test we are interested in long run co-integration. Or We want to see if we can find any significant changes in our LHS because of changes in trend of RHS. And I personally believe ARDL(2,4,7) is much more meaningful than ARDL(7,4,2). Saying that if you have enough observations and follow the same method (ARDL bound) in both software you should get same lag structure.


3- Personally, I don't feel confident using ARDL bound method with observations less than 35-40, because after lagging the variables you end up with 7 or 8 data points which doesn't seem practical. Therefore I strongly suggest that you should increase n. However, the error that you see is not from add-in it self and it is a standard Eviews error. To check the error you can run the regression manually, something like :
series dy=d(y)
series dl1=d(l(-1))
series dy1=d(y(-1))
series dk1=d(k(-1))
then
ls dy c y(-1) k(-1) l(-1) dy1(-1 to -5) dk1(-1 to -5) dl1(-1 to -5)
and result would be "insufficient number of obs" . Even the result of ls dy c y(-1) k(-1) l(-1) dy1(-1 to -5) dk1(-1 to -4) dl1(-1 to -4) doesn't seem realistic with R2=1 !!

Long story short :
- The difference comes from difference in the methodology ! so which methodology you choose will determine which software you want to use.
- Increase your T, if you can't my suggestion is consider other co-integration methods.

Regards
Yashar

[ecofin]

Got it, thanks for your help and for your reply. ,
i hope that you enhance the Add-in in the future with two methodology, and i hope that the EViews team solve this problem (standard error).
Best Regards.