8:54 p.m.
>For a bivariate case, if the trace test rejects c=0 and does not
>reject c=1, I report c=1. If it is the other way around, then I report
>c=0 as the test result.
>As you know, sometimes the results can be contradictory so that
c=0
>and c=1 are rejected (or not rejected) simultaneously.
>My question is that would it be OK to report
"inconclusive" in those
>cases? Or am I expected to follow another further procedure?
Three things to think about (might make you recognize that your case
is more common than many realize):
1.You are searching for results using different lag numbers and the
null hypothesis probabilities are based on knowing the right lag
number beforehand. Of course no one knows the right lag number in a
real study (only known in constructed data cases), but once you
perform a search you should be willing to be skeptical of the test
statistic probabilities.
2. Not rejecting both c=0 and =1 is not an anomaly if you understand
you are only computing the odds each hypothesis is incorrect. You are
not computing the odds of 'correctness' given the results from the
other test. Failing to reject at the 5% level is just that-- failure
to say an hypothesis is blatantly wrong (+ recognizing point 1 that
the 5% number may be misleading).
3. Deciding c=1 vs c=0, i.e. testing whether two time series need to
be differenced or do not need to be differenced to create a stationary
cointegrating relationship is often not as interesting or
controversial as researchers believe it is. Assuming you are only
constructing the co-integrating vector for modeling purposes and this
is a first step, you might find similar results either way. Or once
one understands the data and its source you might conclude c=0 or c=1
is implausible. For example, one might reject c=0 and accept c=1 (or
vice-versa) when testing whether the UK and the US$ price levels are
cointegrated, when the true answer depends on how accurately the price
levels are computed. (I.e. I'd argue they must be differenced at
least once to control for measurement differences before seeing any
Johansen tests.)
(this post might be considered a test of whether an econometric
methodology list connected to GRETL would be worthwhile or fills a
need).
9:07 p.m.
New subject: Johansen question
Thanks Micheal,
In addition to the concern over what optimal lags to choose. When observations
are less than 60, the AIC and FPE best to optimise lags that fit the model to the data
structure. However, when observations are above 60, the HQC is best. The SIC or BIC
underestimate lags with increase in degrees of freedom( See: Khim(2004):
http://ideas.repec.org/a/ebl/ecbull/v3y2004i33p1-9.html)
--- On Sun, 5/1/11, MICHAEL BOLDIN <mboldin(a)temple.edu> wrote:
From: MICHAEL BOLDIN <mboldin(a)temple.edu>
Subject: Re: [Gretl-users] Johansen question
To: gretl-users(a)lists.wfu.edu
Date: Sunday, May 1, 2011, 1:54 PM
>For a bivariate case, if the trace test rejects c=0 and does not
>reject c=1, I report c=1. If it is the other way around, then I report
>c=0 as the test result.
>As you know, sometimes the results can be contradictory so that
c=0
>and c=1 are rejected (or not rejected) simultaneously.
>My question is that would it be OK to report
"inconclusive" in those
>cases? Or am I expected to follow another further procedure?
Three things to think about (might make you recognize that your case
is more common than many realize):
1.You are searching for results using different lag numbers and the
null hypothesis probabilities are based on knowing the right lag
number beforehand. Of course no one knows the right lag number in a
real study (only known in constructed data cases), but once you
perform a search you should be willing to be skeptical of the test
statistic probabilities.
2. Not rejecting both c=0 and =1 is not an anomaly if you understand
you are only computing the odds each hypothesis is incorrect. You are
not computing the odds of 'correctness' given the results from the
other test. Failing to reject at the 5% level is just that-- failure
to say an hypothesis is blatantly wrong (+ recognizing point 1 that
the 5% number may be misleading).
3. Deciding c=1 vs c=0, i.e. testing whether two time series need to
be differenced or do not need to be differenced to create a stationary
cointegrating relationship is often not as interesting or
controversial as researchers believe it is. Assuming you are only
constructing the co-integrating vector for modeling purposes and this
is a first step, you might find similar results either way. Or once
one understands the data and its source you might conclude c=0 or c=1
is implausible. For example, one might reject c=0 and accept c=1 (or
vice-versa) when testing whether the UK and the US$ price levels are
cointegrated, when the true answer depends on how accurately the price
levels are computed. (I.e. I'd argue they must be differenced at
least once to control for measurement differences before seeing any
Johansen tests.)
(this post might be considered a test of whether an econometric
methodology list connected to GRETL would be worthwhile or fills a
need).
_______________________________________________
Gretl-users mailing list
Gretl-users(a)lists.wfu.edu
http://lists.wfu.edu/mailman/listinfo/gretl-users
10:45 a.m.
New subject: Johansen question
El 01/05/11 23:07, Anutechia Asongu escribió:
Thanks Micheal,
In addition to the concern over what optimal lags to choose. When
observations are less than 60, the AIC and FPE best to optimise lags
that fit the model to the data structure. However, when observations are
above 60, the HQC is best. The SIC or BIC underestimate lags with
increase in degrees of freedom( See: Khim(2004):
http://ideas.repec.org/a/ebl/ecbull/v3y2004i33p1-9.html)
I am not sure this last sentence is correct (although I see it is
correctly interpreted from the work of Khim). In fact HQC and BIC are
criteria that obtain a consistent estimator for p (the order of the
VAR). So, when T (or T-k) grows the probability of underestimate p
should decrease.
With respect to this question it is very detailed the explanation given
in chapter 4 of "New Introduction to Multiple time Series" by Helmut
Lütkepohl (Springer, 2005). Let me copy here some conclusions that may
be appropiate given the debate that has emerged in this list: (page 151)
"... It is worth emphasizing that the foregoing results do nor
necessarily mean that AIC and FPE are inferior to HQ and SC [BIC]. Only
if consistency is the yardstick for evaluating the criteria, the latter
two are superior under conditions of the previous section. So far we
have not considered the small sample properties of the estimators. In
small samples, AIC and FPE may have better properties (chose the correct
order more often) than HQ and SC. Also, the former two criteria are
designed for minimizing the forecast error variance. Thus, in small as
well as large samples, models based on AIC and FPE may produce superior
forecasts although they may not estimate the orders correctly. In fact,
Shibata(1980) derived asymptotic optimality properties of AIC and FPE
for univariate processes. He showed that, under suitable conditions,
they indeed minimize the 1-step ahead forecast MSE asymptotically".
--
Ignacio Diaz-Emparanza
DEPARTAMENTO DE ECONOMÍA APLICADA III (ECONOMETRÍA Y ESTADÍSTICA)
UPV/EHU Avda. Lehendakari Aguirre, 83 | 48015 BILBAO
T.: +34 946013732 | F.: +34 946013754
www.ea3.ehu.es
11:11 a.m.
New subject: Johansen question
Ignacio Diaz-Emparanza wrote:
"...........Thus, in small as
well as large samples, models based on AIC and FPE may produce superior
forecasts although they may not estimate the orders correctly.......".
Indeed, Khim's work(2004) doesn't take into account
"underestimation of lags"(which could be an inconvinience of the AIC
and FPE approaches: as lag overestimation decreases the power of the
autoregressive model ). The rule for selecting which Criterion is best
is simply based on the Criterion's ability to overestimate optimal
lags.
Ignacio's explanation is therefore compatible with Khim(2004)
when no priority is given lag-overestimation over lag-underestimation.
--- On Mon, 5/2/11, Ignacio Diaz-Emparanza <ignacio.diaz-emparanza(a)ehu.es> wrote:
From: Ignacio Diaz-Emparanza <ignacio.diaz-emparanza(a)ehu.es>
Subject: Re: [Gretl-users] Johansen question
To: "Gretl list" <gretl-users(a)lists.wfu.edu>
Date: Monday, May 2, 2011, 3:45 AM
El 01/05/11 23:07, Anutechia Asongu escribió:
Thanks Micheal,
In addition to the concern over what optimal lags to choose. When
observations are less than 60, the AIC and FPE best to optimise lags
that fit the model to the data structure. However, when observations are
above 60, the HQC is best. The SIC or BIC underestimate lags with
increase in degrees of freedom( See: Khim(2004):
http://ideas.repec.org/a/ebl/ecbull/v3y2004i33p1-9.html)
I am not sure this last sentence is correct (although I see it is
correctly interpreted from the work of Khim). In fact HQC and BIC are
criteria that obtain a consistent estimator for p (the order of the
VAR). So, when T (or T-k) grows the probability of underestimate p
should decrease.
With respect to this question it is very detailed the explanation given
in chapter 4 of "New Introduction to Multiple time Series" by Helmut
Lütkepohl (Springer, 2005). Let me copy here some conclusions that may
be appropiate given the debate that has emerged in this list: (page 151)
"... It is worth emphasizing that the foregoing results do nor
necessarily mean that AIC and FPE are inferior to HQ and SC [BIC]. Only
if consistency is the yardstick for evaluating the criteria, the latter
two are superior under conditions of the previous section. So far we
have not considered the small sample properties of the estimators. In
small samples, AIC and FPE may have better properties (chose the correct
order more often) than HQ and SC. Also, the former two criteria are
designed for minimizing the forecast error variance. Thus, in small as
well as large samples, models based on AIC and FPE may produce superior
forecasts although they may not estimate the orders correctly. In fact,
Shibata(1980) derived asymptotic optimality properties of AIC and FPE
for univariate processes. He showed that, under suitable conditions,
they indeed minimize the 1-step ahead forecast MSE asymptotically".
--
Ignacio Diaz-Emparanza
DEPARTAMENTO DE ECONOMÍA APLICADA III (ECONOMETRÍA Y ESTADÍSTICA)
UPV/EHU Avda. Lehendakari Aguirre, 83 | 48015 BILBAO
T.: +34 946013732 | F.: +34 946013754
www.ea3.ehu.es
_______________________________________________
Gretl-users mailing list
Gretl-users(a)lists.wfu.edu
http://lists.wfu.edu/mailman/listinfo/gretl-users
1:03 p.m.
New subject: Johansen question
Hi gretl-users,
just an off-topic question to Allin: You have started to write this nice
paper on the critical values for the Johansen cointegration test given some
weakly exogenous I(1) variables. Is there already a more recent version
available?
I also support the idea of a list for discussing econometric issues.
Best,
Artur
2011/5/2 Anutechia Asongu <simplice_peace(a)yahoo.com>
Ignacio Diaz-Emparanza wrote:
"...........Thus, in small as
well as large samples, models based on AIC and FPE may produce superior
forecasts although they may not estimate the orders correctly.......".
Indeed, Khim's work(2004) doesn't take into account
"underestimation of lags"(which could be an inconvinience of the AIC and FPE
approaches: as lag overestimation decreases the power of the autoregressive
model ). The rule for selecting which Criterion is best is simply based on
the Criterion's ability to overestimate optimal lags.
Ignacio's explanation is therefore compatible with Khim(2004) when no
priority is given lag-overestimation over lag-underestimation.
--- On *Mon, 5/2/11, Ignacio Diaz-Emparanza <ignacio.diaz-emparanza(a)ehu.es
>* wrote:
From: Ignacio Diaz-Emparanza <ignacio.diaz-emparanza(a)ehu.es>
Subject: Re: [Gretl-users] Johansen question
To: "Gretl list" <gretl-users(a)lists.wfu.edu>
Date: Monday, May 2, 2011, 3:45 AM
El 01/05/11 23:07, Anutechia Asongu escribió:
> Thanks Micheal,
> In addition to the concern over what optimal lags to choose. When
> observations are less than 60, the AIC and FPE best to optimise lags
> that fit the model to the data structure. However, when observations are
> above 60, the HQC is best. The SIC or BIC underestimate lags with
> increase in degrees of freedom( See: Khim(2004):
> http://ideas.repec.org/a/ebl/ecbull/v3y2004i33p1-9.html)
I am not sure this last sentence is correct (although I see it is
correctly interpreted from the work of Khim). In fact HQC and BIC are
criteria that obtain a consistent estimator for p (the order of the
VAR). So, when T (or T-k) grows the probability of underestimate p
should decrease.
With respect to this question it is very detailed the explanation given
in chapter 4 of "New Introduction to Multiple time Series" by Helmut
Lütkepohl (Springer, 2005). Let me copy here some conclusions that may
be appropiate given the debate that has emerged in this list: (page 151)
"... It is worth emphasizing that the foregoing results do nor
necessarily mean that AIC and FPE are inferior to HQ and SC [BIC]. Only
if consistency is the yardstick for evaluating the criteria, the latter
two are superior under conditions of the previous section. So far we
have not considered the small sample properties of the estimators. In
small samples, AIC and FPE may have better properties (chose the correct
order more often) than HQ and SC. Also, the former two criteria are
designed for minimizing the forecast error variance. Thus, in small as
well as large samples, models based on AIC and FPE may produce superior
forecasts although they may not estimate the orders correctly. In fact,
Shibata(1980) derived asymptotic optimality properties of AIC and FPE
for univariate processes. He showed that, under suitable conditions,
they indeed minimize the 1-step ahead forecast MSE asymptotically".
--
Ignacio Diaz-Emparanza
DEPARTAMENTO DE ECONOMÍA APLICADA III (ECONOMETRÍA Y ESTADÍSTICA)
UPV/EHU Avda. Lehendakari Aguirre, 83 | 48015 BILBAO
T.: +34 946013732 | F.: +34 946013754
www.ea3.ehu.es
_______________________________________________
Gretl-users mailing list
Gretl-users(a)lists.wfu.edu <http://mc/compose?to=Gretl-users@lists.wfu.edu>
http://lists.wfu.edu/mailman/listinfo/gretl-users
_______________________________________________
Gretl-users mailing list
Gretl-users(a)lists.wfu.edu
http://lists.wfu.edu/mailman/listinfo/gretl-users
2:04 p.m.
New subject: Johansen question
In my opinion, the additional discusions about econometric methods is tied
to the methods available in GRETL, so, I think is unnecesary to create
another list. In addition, on average, the amount of emails is low in
comparison to other lists.
Unmistakenly, when you use a a method in a software, is going to be totally
related to the underlying econometric theory. Since the existence of the
former is conditional on the later. Therefore, the discussion of both
topics go hand in hand and deserve to be discussed simultaneously.
Cheers,
Charles
--
Charles Koss
http://charlesonnet.blogspot.com
On Mon, May 2, 2011 at 8:03 AM, artur tarassow <
artur.tarassow(a)googlemail.com> wrote:
Hi gretl-users,
just an off-topic question to Allin: You have started to write this nice
paper on the critical values for the Johansen cointegration test given some
weakly exogenous I(1) variables. Is there already a more recent version
available?
I also support the idea of a list for discussing econometric issues.
Best,
Artur
2011/5/2 Anutechia Asongu <simplice_peace(a)yahoo.com>
>
> Ignacio Diaz-Emparanza wrote:
> "...........Thus, in small as
> well as large samples, models based on AIC and FPE may produce superior
> forecasts although they may not estimate the orders correctly.......".
> Indeed, Khim's work(2004) doesn't take into account
> "underestimation of lags"(which could be an inconvinience of the AIC and
FPE
> approaches: as lag overestimation decreases the power of the autoregressive
> model ). The rule for selecting which Criterion is best is simply based on
> the Criterion's ability to overestimate optimal lags.
> Ignacio's explanation is therefore compatible with Khim(2004) when
> no priority is given lag-overestimation over lag-underestimation.
>
>
> --- On *Mon, 5/2/11, Ignacio Diaz-Emparanza <
> ignacio.diaz-emparanza(a)ehu.es>* wrote:
>
>
> From: Ignacio Diaz-Emparanza <ignacio.diaz-emparanza(a)ehu.es>
> Subject: Re: [Gretl-users] Johansen question
> To: "Gretl list" <gretl-users(a)lists.wfu.edu>
> Date: Monday, May 2, 2011, 3:45 AM
>
>
> El 01/05/11 23:07, Anutechia Asongu escribió:
> > Thanks Micheal,
> > In addition to the concern over what optimal lags to choose. When
> > observations are less than 60, the AIC and FPE best to optimise lags
> > that fit the model to the data structure. However, when observations are
> > above 60, the HQC is best. The SIC or BIC underestimate lags with
> > increase in degrees of freedom( See: Khim(2004):
> > http://ideas.repec.org/a/ebl/ecbull/v3y2004i33p1-9.html)
>
> I am not sure this last sentence is correct (although I see it is
> correctly interpreted from the work of Khim). In fact HQC and BIC are
> criteria that obtain a consistent estimator for p (the order of the
> VAR). So, when T (or T-k) grows the probability of underestimate p
> should decrease.
>
> With respect to this question it is very detailed the explanation given
> in chapter 4 of "New Introduction to Multiple time Series" by Helmut
> Lütkepohl (Springer, 2005). Let me copy here some conclusions that may
> be appropiate given the debate that has emerged in this list: (page 151)
>
> "... It is worth emphasizing that the foregoing results do nor
> necessarily mean that AIC and FPE are inferior to HQ and SC [BIC]. Only
> if consistency is the yardstick for evaluating the criteria, the latter
> two are superior under conditions of the previous section. So far we
> have not considered the small sample properties of the estimators. In
> small samples, AIC and FPE may have better properties (chose the correct
> order more often) than HQ and SC. Also, the former two criteria are
> designed for minimizing the forecast error variance. Thus, in small as
> well as large samples, models based on AIC and FPE may produce superior
> forecasts although they may not estimate the orders correctly. In fact,
> Shibata(1980) derived asymptotic optimality properties of AIC and FPE
> for univariate processes. He showed that, under suitable conditions,
> they indeed minimize the 1-step ahead forecast MSE asymptotically".
>
>
> --
> Ignacio Diaz-Emparanza
> DEPARTAMENTO DE ECONOMÍA APLICADA III (ECONOMETRÍA Y ESTADÍSTICA)
> UPV/EHU Avda. Lehendakari Aguirre, 83 | 48015 BILBAO
> T.: +34 946013732 | F.: +34 946013754
> www.ea3.ehu.es
>
>
>
>
> _______________________________________________
> Gretl-users mailing list
> Gretl-users@lists.wfu.edu<http://mc/compose?to=Gretl-users@lists.wfu.edu>
> http://lists.wfu.edu/mailman/listinfo/gretl-users
>
>
> _______________________________________________
> Gretl-users mailing list
> Gretl-users(a)lists.wfu.edu
> http://lists.wfu.edu/mailman/listinfo/gretl-users
>
_______________________________________________
Gretl-users mailing list
Gretl-users(a)lists.wfu.edu
http://lists.wfu.edu/mailman/listinfo/gretl-users
4:22 p.m.
New subject: Johansen question
Am 02.05.2011 16:04, schrieb Charles Koss:
In my opinion, the additional discusions about econometric methods
is
tied to the methods available in GRETL, so, I think is unnecesary to
create another list. In addition, on average, the amount of emails is
low in comparison to other lists.
Unmistakenly, when you use a a method in a software, is going to be
totally related to the underlying econometric theory. Since the
existence of the former is conditional on the later. Therefore, the
discussion of both topics go hand in hand and deserve to be discussed
simultaneously.
Cheers,
Charles
Two points related to this:
Before creating a new list for discussion of econometrics, maybe it
would be worthwhile to check if another list already serves the same
purpose (see for example
http://www.feweb.vu.nl/econometriclinks/mailing/). Even if we come to
the conclusion that there is a niche to be filled, maybe it would make
sense to at least coordinate with the R people?
OTOH, I'm not overly enthusiastic to have general econometric
discussions on gretl-users. It's true that the traffic here is not high,
and this is also due to the fact that in the past off-topic (=not
gretl-specific) questions have been discouraged. But this low traffic
makes it possible that a relevant number of subscribers are actually
following most (if not all) threads. I fear that if general econometric
questions are discussed (which I believe will be quite basic most of the
time) then some people might pay less attention to the list in general,
with detrimental effects on the gretl-specific issues.
But maybe I'm paranoid.
cheers,
sven
6:18 p.m.
New subject: Johansen question
What about encouraging in this list discussions tied to the
econometric theory as it relates to the methods implemented in GRETL?
So that, we discourage, unrelated econometric theory to the
implemented methods in GRETL.
As the old old old saying goes "We can not have bread without the
dough as we can not have dough without the wheat"
Sincerely,
Charles
On Mon, May 2, 2011 at 11:22 AM, Sven Schreiber <svetosch(a)gmx.net> wrote:
Am 02.05.2011 16:04, schrieb Charles Koss:
> In my opinion, the additional discusions about econometric methods is
> tied to the methods available in GRETL, so, I think is unnecesary to
> create another list. In addition, on average, the amount of emails is
> low in comparison to other lists.
>
> Unmistakenly, when you use a a method in a software, is going to be
> totally related to the underlying econometric theory. Since the
> existence of the former is conditional on the later. Therefore, the
> discussion of both topics go hand in hand and deserve to be discussed
> simultaneously.
>
> Cheers,
>
> Charles
>
Two points related to this:
Before creating a new list for discussion of econometrics, maybe it
would be worthwhile to check if another list already serves the same
purpose (see for example
http://www.feweb.vu.nl/econometriclinks/mailing/). Even if we come to
the conclusion that there is a niche to be filled, maybe it would make
sense to at least coordinate with the R people?
OTOH, I'm not overly enthusiastic to have general econometric
discussions on gretl-users. It's true that the traffic here is not high,
and this is also due to the fact that in the past off-topic (=not
gretl-specific) questions have been discouraged. But this low traffic
makes it possible that a relevant number of subscribers are actually
following most (if not all) threads. I fear that if general econometric
questions are discussed (which I believe will be quite basic most of the
time) then some people might pay less attention to the list in general,
with detrimental effects on the gretl-specific issues.
But maybe I'm paranoid.
cheers,
sven
11:01 p.m.
New subject: Johansen question
Michael,
Your second point sounds like you're giving a Bayesian posterior probability
interpretation to a frequentist hypothesis test, and I'm not sure that's warranted
(it may be -- this is off the top of my head). Cointegration is one of the areas that I
think makes a lot more sense from a Bayesian point of view, but I'll have to do some
remedial reading before I can say any more about this particular issue.
I guess this also relates to your parenthetical 'point'.
PS
________________________________________
From: gretl-users-bounces(a)lists.wfu.edu [gretl-users-bounces(a)lists.wfu.edu] On Behalf Of
MICHAEL BOLDIN [mboldin(a)temple.edu]
Sent: Sunday, May 01, 2011 3:54 PM
To: gretl-users(a)lists.wfu.edu
Subject: Re: [Gretl-users] Johansen question
>For a bivariate case, if the trace test rejects c=0 and does not
>reject c=1, I report c=1. If it is the other way around, then I report
>c=0 as the test result.
>As you know, sometimes the results can be contradictory so that
c=0
>and c=1 are rejected (or not rejected) simultaneously.
>My question is that would it be OK to report
"inconclusive" in those
>cases? Or am I expected to follow another further procedure?
Three things to think about (might make you recognize that your case
is more common than many realize):
1.You are searching for results using different lag numbers and the
null hypothesis probabilities are based on knowing the right lag
number beforehand. Of course no one knows the right lag number in a
real study (only known in constructed data cases), but once you
perform a search you should be willing to be skeptical of the test
statistic probabilities.
2. Not rejecting both c=0 and =1 is not an anomaly if you understand
you are only computing the odds each hypothesis is incorrect. You are
not computing the odds of 'correctness' given the results from the
other test. Failing to reject at the 5% level is just that-- failure
to say an hypothesis is blatantly wrong (+ recognizing point 1 that
the 5% number may be misleading).
3. Deciding c=1 vs c=0, i.e. testing whether two time series need to
be differenced or do not need to be differenced to create a stationary
cointegrating relationship is often not as interesting or
controversial as researchers believe it is. Assuming you are only
constructing the co-integrating vector for modeling purposes and this
is a first step, you might find similar results either way. Or once
one understands the data and its source you might conclude c=0 or c=1
is implausible. For example, one might reject c=0 and accept c=1 (or
vice-versa) when testing whether the UK and the US$ price levels are
cointegrated, when the true answer depends on how accurately the price
levels are computed. (I.e. I'd argue they must be differenced at
least once to control for measurement differences before seeing any
Johansen tests.)
(this post might be considered a test of whether an econometric
methodology list connected to GRETL would be worthwhile or fills a
need).
_______________________________________________
Gretl-users mailing list
Gretl-users(a)lists.wfu.edu
http://lists.wfu.edu/mailman/listinfo/gretl-users
5002
days inactive
5003
days old
8 comments
7 participants
participants (7)
-
Anutechia Asongu -
artur tarassow -
Charles Koss -
Ignacio Diaz-Emparanza -
MICHAEL BOLDIN -
Summers, Peter -
Sven Schreiber