11:47 a.m.
Am 20.02.2014 12:09, schrieb smyks:
I have checked with denmark.gdt dataset:
coint 10 1 2 --verbose --test-down <= Gretl reports sample size of 44
coint 10 1 2 --verbose --test-down --skip-df <= Gretl reports sample
size of 53
Ok I now see what you mean, effectively different samples are used in
what should be the same approach, like you said. I don't see a
justification for that (i.e. I tend to agree it's a bug).
To my understanding both command lines should employ the same sample
size of 53, which is in line with Hayashi (2000).
Actually, when looking at the different samples that are used in the
various steps of the procedure I think they're debatable in general. If
you pick a max lag of 10 I would actually tend to favor the shortest
sample (T=44) instead of the longest (T=53), because only the shortest
is the common denominator for everything. Of course, if results differ
wildly, it's a sign of problems anyway.
cheers,
sven
3:34 p.m.
New subject: Difference in EG tests
On Thu, 20 Feb 2014, Sven Schreiber wrote:
Am 20.02.2014 12:09, schrieb smyks:
> I have checked with denmark.gdt dataset:
> coint 10 1 2 --verbose --test-down <= Gretl reports sample size of 44
> coint 10 1 2 --verbose --test-down --skip-df <= Gretl reports sample
> size of 53
Ok I now see what you mean, effectively different samples are used in
what should be the same approach, like you said. I don't see a
justification for that (i.e. I tend to agree it's a bug).
> To my understanding both command lines should employ the same sample
> size of 53, which is in line with Hayashi (2000).
Actually, when looking at the different samples that are used in the
various steps of the procedure I think they're debatable in general. If
you pick a max lag of 10 I would actually tend to favor the shortest
sample (T=44) instead of the longest (T=53), because only the shortest
is the common denominator for everything. Of course, if results differ
wildly, it's a sign of problems anyway.
I've now fixed the breakage in the --test-down option for "coint", so that
it works in the same way as for "adf". (When I updated adf a while back,
I'm afraid I just forgot about coint.)
On this sample-size issue, I see what's happening. When the --skip-df
option is not given we run initial (A)DF tests on two or more variables,
and we make an effort to impose a uniform sample range, which is then also
applied to the final test of the residuals from the cointegrating
regression. Under --skip-df, when only one (A)DF test is run, we just let
the sample default to the longest available.
I agree that the resulting difference in the Engle-Granger test is hard to
justify, but like Sven I'm not sure what the "right" fix is. Obviously
testing for the minimized information criterion must be done on a uniform
sample range that reserves the maximum number of lags. If we then --
having found the optimal lag, k* -- switch to the longest sample
consistent with that lag, we're arguably being inconsistent since we might
have got a different k* if we tested on a longer sample in the first
place.
Any other thoughts on this?
Allin Cottrell
4 p.m.
New subject: Difference in EG tests
On Thu, 20 Feb 2014, Allin Cottrell wrote:
On this sample-size issue, I see what's happening. When the
--skip-df option
is not given we run initial (A)DF tests on two or more variables, and we make
an effort to impose a uniform sample range, which is then also applied to the
final test of the residuals from the cointegrating regression. Under
--skip-df, when only one (A)DF test is run, we just let the sample default to
the longest available.
I agree that the resulting difference in the Engle-Granger test is hard to
justify, but like Sven I'm not sure what the "right" fix is.
Oh, what the heck, it's all asymptotic anyway: if adding/dropping a few
initial observations makes a big difference the test is not going to be
very meaningful. So I've provisionally (in CVS) made the final DF test in
Engle-Granger use the maximal sample (regardless of whether --skip-df is
given). If anyone has a strenuous objection, please say.
Allin Cottrell
4:22 p.m.
New subject: Difference in EG tests
Am 20.02.2014 17:00, schrieb Allin Cottrell:
On Thu, 20 Feb 2014, Allin Cottrell wrote:
> On this sample-size issue, I see what's happening. When the --skip-df option
> is not given we run initial (A)DF tests on two or more variables, and we make
> an effort to impose a uniform sample range, which is then also applied to the
> final test of the residuals from the cointegrating regression. Under
> --skip-df, when only one (A)DF test is run, we just let the sample default to
> the longest available.
>
> I agree that the resulting difference in the Engle-Granger test is hard to
> justify, but like Sven I'm not sure what the "right" fix is.
Oh, what the heck, it's all asymptotic anyway: if adding/dropping a few
initial observations makes a big difference the test is not going to be
very meaningful. So I've provisionally (in CVS) made the final DF test in
Engle-Granger use the maximal sample (regardless of whether --skip-df is
given). If anyone has a strenuous objection, please say.
I haven't thought this through if the problem actually exist, but I
would want to have gretl's behavior be invariant to whether or not the
dataset is subsampled. That is, the test result should only be a
function of the currently selected sample and the chosen max lag, not be
dependent on whether or not some pre-sample values exist.
Not sure if that's actually against the spirit in the rest of gretl.
-sven
3989
days inactive
3989
days old
3 comments
2 participants
participants (2)
-
Allin Cottrell -
Sven Schreiber