12:04 p.m.
Hello,
I have a minor question concerning how sample size affects ADF unit root
test results in gretl and other econometric software.
Let me give you an example:
I have a series of T=51, for which ADF test results are all the same in
gretl, R (urca) and JMulTi. Yet when I use a subsample of T=20 (last
twenty observations), test statistics obtained with gretl are
significantly different from those of R and JMulTi (both the latter,
however, produce the same output). Of course, I use the same
deterministic term option and lag lenght for those comparisons.
Is there any particular reason for that situation? Or is it me doing
something wrong?
(I attach data file for those, who'd like to replicate my issue)
Best,
Greg
12:28 p.m.
On Mon, 20 Feb 2012, Grzegorz Konat wrote:
I have a minor question concerning how sample size affects
ADF unit root test results in gretl and other econometric
software.
Let me give you an example:
I have a series of T=51, for which ADF test results are all the same in
gretl, R (urca) and JMulTi. Yet when I use a subsample of T=20 (last twenty
observations), test statistics obtained with gretl are significantly
different from those of R and JMulTi (both the latter, however, produce the
same output). Of course, I use the same deterministic term option and lag
lenght for those comparisons.
If you do, for example,
smpl 1991 2010
in gretl, the effect is that gretl will start _estimation_ in
1991. However, if the model contains lags and the lagged terms
are available for dates prior to 1991, the earlier terms will
be used as needed.
If you want the effect of "blanking out" all data prior to
1991 (which means that estimation of a model with lags will
have to start at a later data), you'll have to advance the
start of the sample range further. E.g. your model contains
3 lags -> start the sample in 1994.
Allin Cottrell
12:50 p.m.
Thank you very much, Allin. This has actually solved my issue.
Best,
Greg
W dniu 2012-02-20 18:28, Allin Cottrell pisze:
On Mon, 20 Feb 2012, Grzegorz Konat wrote:
> I have a minor question concerning how sample size affects
> ADF unit root test results in gretl and other econometric
> software.
>
> Let me give you an example:
>
> I have a series of T=51, for which ADF test results are all the same in
> gretl, R (urca) and JMulTi. Yet when I use a subsample of T=20 (last twenty
> observations), test statistics obtained with gretl are significantly
> different from those of R and JMulTi (both the latter, however, produce the
> same output). Of course, I use the same deterministic term option and lag
> lenght for those comparisons.
If you do, for example,
smpl 1991 2010
in gretl, the effect is that gretl will start _estimation_ in
1991. However, if the model contains lags and the lagged terms
are available for dates prior to 1991, the earlier terms will
be used as needed.
If you want the effect of "blanking out" all data prior to
1991 (which means that estimation of a model with lags will
have to start at a later data), you'll have to advance the
start of the sample range further. E.g. your model contains
3 lags -> start the sample in 1994.
Allin Cottrell
_______________________________________________
Gretl-users mailing list
Gretl-users(a)lists.wfu.edu
http://lists.wfu.edu/mailman/listinfo/gretl-users
--
Pozdrawiam,
Grzegorz Konat
7:29 a.m.
New subject: logit threshold
Hi everybody,
In the logit binary model output table, there is an "Actual - Predicted"
table which is obtained by taking 0.5 threshold probability. I can do
the following manually, but I think that it would be nice to have it in
the "specify model" window.
Specifically,
1) sometimes it would be desirable to choose another probability instead
of 0.5, so a box for a threshold value. Also,
2) some people advice to take a threshold such that the fraction of
yhat=1 in the sample is the same as sample mean of y (one more box to
check).
3) It would also be convenient to have in the output table z such that
logistic(z)=threshold
Not quite sure that these proposals are worth attention, just my 0.02.
Best regards,
Rem
4587
days inactive
4611
days old
3 comments
3 participants
participants (3)
-
Allin Cottrell -
Grzegorz Konat -
Remigijus Lapinskas