[Gretl-users] interpretation var lag selection

[JOSE FRANCISCO PERLES RIBES]

Dear all.

I'm estimating a Unrestricted VAR in gretl via the GUI, and previosly I
check the best order of the VAR with the option selection menu. I'm a
little surprised to see how AIC BIC and HQC criteria obtained are different
in this option, which then get really when estimating the VAR.

For example, I find that selecting a maximum of 4 lags, the optimal lag
function gives me 1 lag in AIC, BIC and HQC. But when after I estimated the
model with 1 or 2 lags, AIC, BIC and HQC are lower for 2 than for 1 lags, plus
get a better R2 with 2 lags. I have found that the criteria obtained by
estimating the VAR are identical to those of Eviews.


Am I doing something wrong? or is there a problem with the menu
selection optimum
lag for vars? Because the results and interpretations (eg in the
Granger Causality
test) are very different as choose one lag or two lags for the var
estimation.

Thanks in advance and sorry for any inconvenience.

Jos? F. Perles
University of Alicante
Spain
-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://lists.wfu.edu/pipermail/gretl-users/attachments/20121001/5b81de9b/attachment.html

[Henrique Andrade]

Dear Jos? Francisco,

I think when Gretls does the VAR lag selection tests, it
restricts the sample in order to guarantee that all the
estimations could use the same sample.

That said, when you estimate a 1 lagged VAR the criteria
you see on the test results window differs from the criteria
that you see on the estimated VAR window.

Best regards,
Henrique Andrade*
*
-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://lists.wfu.edu/pipermail/gretl-users/attachments/20121001/21fbdb33/attachment.html

[Allin Cottrell]

Post by Henrique Andrade
I think when Gretls does the VAR lag selection tests, it
restricts the sample in order to guarantee that all the
estimations could use the same sample.
That said, when you estimate a 1 lagged VAR the criteria
you see on the test results window differs from the criteria
that you see on the estimated VAR window.

Right: the VAR lag selection command uses a common sample
size, but the sample range used with the regular "var" command
will depend on the number of lags specified.

As a further comment, if the results differ substantially in
the two cases, that suggests that either (a) the sample is too
small for reliable inference or (b) the data are heterogenous
in a way that is not accounted for in the model specification.
Either way it's a warning flag.

Allin Cottrell

[JOSE FRANCISCO PERLES RIBES]

Thanks a lot, Henrique and Allin

Indeed, my sample has both problems, is very small (41 annual observations) and
also has variability. I have reduced the variability taking logarithms. But
the size of the sample can not be increased.... unless the time runs.

Therefore, as you say, I take my inference very very cautiously...

But, If possible I would appreciate your opinion on two questions:

1) Even with all the precautions, in a case like this, it would be better
to choose the model based on the criteria obtained in the model estimation
window, rather than in the selection criteria model window that as you say
restrict the var?

2) Could you recommend any chance to improve the inference with a sample like
this, because I do not know if the Bootstrap technique can be applied in
this context var?

Some reference will be appreciate...

Thanks a lot and sorry for any inconvenience.

Jos? F. Perles
University of Alicante (Spain).
-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://lists.wfu.edu/pipermail/gretl-users/attachments/20121002/badcdde8/attachment.html