3:25 p.m.
Hi,
I've noticed a strange behaviour of GRETL with Rank Correlation. I
tried to get the value of the Spearman \rho with two series of about
15000 obs and I got \rho=-2038.292267 (???).
If I do the same with R(even launched inside GRETL) or Stata I get a
more reasonable 0.5715.
Am I missing something? How is the Spearman's \rho defined in Gretl?
Thanks,
Angelo
--
========================================================
Angelo Secchi PGP Key ID:EA280337
========================================================
Research Fellow Scuola Superiore S.Anna
Piazza Martiri della Liberta' 33, Pisa, 56127 Italy
ph.: +39050883365 fax: +39050883344
email: secchi(a)sssup.it www.sssup.it/~secchi/
========================================================
7:41 p.m.
On Wed, 25 Jul 2007, Angelo Secchi wrote:
I've noticed a strange behaviour of GRETL with Rank Correlation.
I
tried to get the value of the Spearman \rho with two series of about
15000 obs and I got \rho=-2038.292267 (???).
Oops, that doesn't sound good. I'll take a look at the issue and
report back.
--
Allin Cottrell
Department of Economics
Wake Forest University, NC
5:13 a.m.
On Wed, 25 Jul 2007, Angelo Secchi wrote:
I've noticed a strange behaviour of GRETL with Rank
Correlation...
I hope this is fixed now, in CVS and the current Windows snapshot
at
http://ricardo.ecn.wfu.edu/pub/gretl/gretl_install.exe
I think the problem was incorrect treatment of the case where
there are ties in the rank calculation. This area of gretl hasn't
been examined in quite a long time. I would be glad to receive
reports of good or bad behaviour with the modified gretl,
Allin Cottrell
10:37 a.m.
On Thu, 26 Jul 2007 01:13:00 -0400 (EDT)
Allin Cottrell <cottrell(a)wfu.edu> wrote:
On Wed, 25 Jul 2007, Angelo Secchi wrote:
> I've noticed a strange behaviour of GRETL with Rank
> Correlation...
I hope this is fixed now, in CVS and the current Windows snapshot
at
http://ricardo.ecn.wfu.edu/pub/gretl/gretl_install.exe
I think the problem was incorrect treatment of the case where
there are ties in the rank calculation. This area of gretl hasn't
been examined in quite a long time. I would be glad to receive
reports of good or bad behaviour with the modified gretl,
Now it seems ok, thanks. I checked also the test
and I have a couple of doubts:
1. There something I do not understand in p-values. I believe you used
the Normal approximation (instead of Student-T) because of ties. Am I
right? Still the value obtained are significantly different from the
ones obtained in R and Stata and I do not believe it is an
approximation issue. If you provide me some details (or the reference)
on the formula used for the p-value of this test I can compare results
with different software and give you more detailed feed-back.
2. If I may, I would suggest to implement as default a two tailed
alternative. In general this test is considered to verify the presence of
correlation between two samples independently from its sign.
Thanks again,
Angelo
Allin Cottrell
_______________________________________________
Gretl-users mailing list
Gretl-users(a)lists.wfu.edu
http://lists.wfu.edu/mailman/listinfo/gretl-users
--
========================================================
Angelo Secchi PGP Key ID:EA280337
========================================================
Research Fellow Scuola Superiore S.Anna
Piazza Martiri della Liberta' 33, Pisa, 56127 Italy
ph.: +39050883365 fax: +39050883344
email: secchi(a)sssup.it www.sssup.it/~secchi/
========================================================
8:26 p.m.
On Thu, 26 Jul 2007, Angelo Secchi wrote:
Now it seems ok, thanks. I checked also the test and I have a
couple of doubts:
1. There something I do not understand in p-values. I believe
you used the Normal approximation (instead of Student-T) because
of ties. Am I right? Still the value obtained are significantly
different from the ones obtained in R and Stata and I do not
believe it is an approximation issue...
Yes, I was using (possibly mis-using) the normal approximation. I
have now changed this to use the t-approximation when there are
ties.
t(n - 2) = rho / sqrt((1 - rho) * (1 - rho) / (n - 2))
See what you think.
2. If I may, I would suggest to implement as default a two
tailed alternative. In general this test is considered to verify
the presence of correlation between two samples independently
from its sign.
Yes, that's a good idea.
Allin.
6117
days inactive
6118
days old
4 comments
2 participants
participants (2)
-
Allin Cottrell -
Angelo Secchi