Re: [Gretl-devel] exog. lags in VECM
Allin Cottrell schrieb:
On Thu, 29 Jan 2009, Sven Schreiber wrote:
> I encountered the following problem when specifying a VECM with
> exogenous variables:
>
> I wanted to specify x(-1) as a restricted exogenous variable. It
> turned out to be not so easy to get the lag in there instead of
> just the contemporaneous value...
>
> Given that the VAR specification dialog has a great interface
> for choosing the lags of exogenous variables, wouldn't it be
> natural to have the same thing in the VECM window?
Yes, it would, but it's very awkward combining the two dimensions
of restricted/unrestricted and lag order via the GUI pulleys and
wheels. However, I've now had a go at this in CVS and Windows
snapshot -- so please test!
Thanks, I will, today (hopefully...)
> And while we're (or I'm) at it, couldn't this flexible approach for
> choosing lags also be extended to endogenous variables?
Sorry, not sure what you mean: do you want "gappy" VARs/VECMs
with holes in the lag order?
Yes. The state of the "competition" is as follows AFAIK:
Eviews can have holes in the lag order, but only one lag specification
for all variables.
With PcGive/Oxmetrics you can choose the lags separately for each
variable; I don't remember right now if that requires the specification
of a "model" (in their parlance) instead of a VAR, but I don't think so,
as long as all equations have the same regressor set. I'm aware that
this is also possible in gretl by appropriately specifying a system, but
that is not as user-friendly.
In Jmulti you can easily impose zero restrictions on individual
coefficients after the first standard VAR-estimation and then
re-estimate the model.
AFAICT extending the current state of affairs with exogenous variables
to the endogenous variables would bring gretl up to par with PcGive,
leaving Eviews a little behind. OLS would still be efficient, in
contrast to the general approach of Jmulti, where SUR would need to be done.
With VECMs it's a little more complicated (especially with the flexible
Jmulti approach, but as long as all chosen lags/regressors appear in all
equations it should be fine.
thanks,
sven