Re: [Gretl-devel] coint2 with exogenous variables
On Sun, 6 Nov 2011, Andreas Noack Jensen wrote:
> On Sat, 5 Nov 2011, Sven Schreiber wrote:
>
>> On 11/05/2011 09:35 PM, Allin Cottrell wrote:
>>> On Sat, 5 Nov 2011, Andreas Noack Jensen wrote:
>>>
>>> Yes, I came to the same conclusion, and I've just committed a change
>>> to gretl CVS that renders the results the same (from gretl and Ox)
>>> for your original script.
I have just installed and tested it. Great job. Thank you.
Glad you like it.
>> I'm not sure but could it be that we had discussed the
question of
>> whether the exogenous restricted variables should be lagged or not?
>> Ultimately it's a matter of convention, of course, but what I may have
>> argued back then is this: since the error correction term also (usually)
>> appears with a t-1 index, I'd say it's not wrong to include x_{t-1}
>> instead of x_t.
That is right and this is also the notation of Harbo et al. but I
think it is confusing if the restricted and unrestricted exogenous
variables are treated differently with respect to lags. After
Allins change exogenous variables are taken literally and you are
free to include the lagged variables as restricted if you want.
Allin here: I guess I'd say that one way or another there will be
something at least slightly confusing here (I mean, either the
restricted Xs will be treated differently from the unrestricted Xs,
or they will be treated differently from the restricted Ys).
> I would guess that in the vecm field the "industry
standard" is
> pcgive/ox. If we're unsure on a convention, I'd rather follow
> Doornik.
Here I agree with Jack (pragmatism talking).
As far as I can see, OxMetrics has not really a convention on
this.
Well, it seems that there _is_ a convention de facto. That's what
gave rise to your original suggestion that there was something amiss
in gretl's version of the analysis.
>>> You're suggesting that we shouldn't allow
including unrestricted
>>> deterministic terms in cointegration testing (though we should allow
>>> them in VECM specifications, having determined cointegrating rank)?
Yes something like that. I cannot see why you would want to
determine rank in the models with unrestricted terms but I do see
that people do that. I tend to believe it is because people are
unaware of the similarity problems. If you don't want the trend in
the cointegration relation it is easily excluded afterwards with
linear restrictions on the cointegration vectors and hence the
unrestricted models are not needed at all. At least if you could
get the full output after imposing restrictions, so that would be
another suggestion from my side.
>> I think you must take into account that he has been trained in
>> Copenhagen…
I guess that fact cannot disqualify me when the topic is
cointegration i VECMs
Allin: No, far from it. (At least as far as I am concerned.)
>> They put a lot of emphasis on separating the cointegration
>> rank determination from the specification of the deterministics.
Is this controversial?
>> We can discuss this further if you like, but I am certainly
>> strongly against disabling some of the possible test setups.
>> Note that the most widespread case of an unrestricted constant
>> also belongs to that category.
But as mentioned above. You do not loose that model. It is just a
restriction. However, my main point is that it should't be the
default.
Hmm, the "default" is in a sense up to the user. It may be that it's
unwise to test for cointegration while including unrestricted
exogenous terms. But as you've illustrated, Ox is "quite happy" to
do that if it's what the user requests. It would be be awkward
(though of course not impossible) for gretl to say "Sorry, you can't
do that!". Some (presumably rather extended) explanation would be
required for why this was prohibited.
Allin Cottrell