#97 cross-sectional dependence test

[KTTK]

It seems, gretl misses tests for cross-sectional dependence, e.g. Breusch/Pagan LM Test (also scaled) and Pesaran's CD Test (see Pesaran (2004)). The only hit for a search for gretl and cross-sectional dependence I found it this (unanswered) mail from gretl-users mailing list:
http://lists.wfu.edu/pipermail/gretl-users/2010-December/005630.html

Also, there is no function package for this type of test. Would be good to have these tests available in gretl.

[Sven Schreiber]

Am 04.08.2016 um 19:07 schrieb KTTK:

It seems, gretl misses tests for cross-sectional dependence, e.g.
Breusch/Pagan LM Test (also scaled) and Pesaran's CD Test (see
Pesaran (2004)).

Yes, cross-sectional dep. (CSD) is a topic which I also find important.
In principle these formulas are pretty easy, the most challenging part
is to handle the data in the unbalanced panel case. I have a first
prototype function for the two tests that you mention (only the scaled
version of the LM test). This could evolve into a function package
fairly quickly.

Are you on one of the gretl mailing lists? I'm not sure the discussion
with lengthy code examples should be carried out here on the tracker.
Although we could try, why not -- maybe there's also an attachment
feature somewhere (right now I'm not using the web interface but doing
this via email reply).

cheers,
sven

[KTTK]

I am not on the list (anymore). Unfortunately, I cannot really be of help here because I do have neither hansl nor C skills. My post was really more of a hint/feature request.

As a benchmark, you can look at R's plm (make sure to use the development version), EViews 9.5 (also make sure to not use an earlier version) or Stata (at least for some tests).

[Sven Schreiber]

Am 05.08.2016 um 17:47 schrieb KTTK:

I am not on the list (anymore). Also, I cannot really help there
because I do not have neither hansl nor C skills. My post was really
a like a feature request.

OK, let's keep it on the tracker then. And I didn't mean that you have
to develop the package, I'll do it. I just thought that you could do
some testing.

As a benchmark, you can look at R's plm (make sure to use the
development version), EViews 9.5 (also make sure to not use an
earlier version) or Stata (at least for some tests).

I do have Eviews 9.5 and will cross-check with that. But can you tell me
why you're saying no earlier version? Were the tests wrong or simply not
there?

thanks,
sven

[KTTK]

About EViews: In some former version they had the CD tests only as a (possibly) user-supplied add-in and that gave wrong test statistics for some time (also they spelled Pesaran as Pearson there...), see http://forums.eviews.com/viewtopic.php?t=8109

The add-in had more/other tests compared to the now official incorporated tests in EViews, e. g. Frees' test (but that one is in Stata xtcsd).

A hint would be to use the first 100 lines of the Grunfeld data for comparision (most p-values are != 0 for that data, allowing for better comparison). And then also make an unbalanced version of that data.

[Sven Schreiber]

Am 05.08.2016 um 18:17 schrieb KTTK:

About EViews: In some former version they had the CD tests only as a
(possibly) user-supplied add-in and that gave wrong test statistics
for some time (also they spelled Pesaran as Pearson there...), see
http://forums.eviews.com/viewtopic.php?t=8109

Ok, thanks. I've also noticed now that eq 44.29 in the Eviews 9.5 help
contains a typo. (missing root over the T_ij)
So I've already got Pesaran's statistic alright in my code. The others
still need some checks. Due to the holiday season a (preliminary)
function package is probably going to take another two weeks.

cheers,
sven

[Sven Schreiber]

This is now implemented in core gretl by Allin (development version, to become version 2016d) as modtest ... --xdepend. Thanks for the suggestion.

[KTTK]

This is great news, thank you guys!

How do you think about implementing the LM [1] and scaled LM [2] test for cross sectional dependence as well? Should be pretty easy to compute those once the components for Pesaran's CD statistic are there. However, I am not sure if those are used widly in applied econometrics...

[1] Breusch, T.S. and A.R. Pagan (1980), The Lagrange multiplier test and its applications to model specification in econometrics, Review of Economic Studies, 47(1), pp. 239–253.

[2] Pesaran, H. (2004), General Diagnostic Tests for Cross Section Dependence in Panels, CESifo Working Paper 1229.

[Sven Schreiber]

Am 02.11.2016 um 10:21 schrieb KTTK:

How do you think about implementing the LM [1] and scaled LM [2] test
for cross sectional dependence as well? Should be pretty easy to
compute those once the components for Pesaran's CD statistic are
there. However, I am not sure if those are used widly in applied
econometrics...

I think there should be a good econometric reason to offer other tests.
The standard way in gretl is to have one "modtest" variant per topic, as
in "modtest --autocorr" for example. (But I admit there are some
exceptions.)
I am not up to date on the literature in that area, is there a consensus
that the other tests are better (at least in some relevant situations)?

Apparently I wrote that I had already started programming (in hansl)
other test statistics, but to be honest right now I cannot remember
anything about that. (Although it's probably still lying around on my
hard disk.)

thanks,
sven

[KTTK]

As I wrote, not sure about what is used in the wild and if the statistics (their properties) differ substantially. The user-written Stata module xtcsd offers Friedman's test and Frees' test besids Pesaran's CD test:
http://fmwww.bc.edu/RePEc/bocode/x/xtcsd.html