Re: [Gretl-devel] modtest option --xdepend
On Fri, 12 Aug 2016, Sven Schreiber wrote:
Am 12.08.2016 um 17:30 schrieb Allin Cottrell:
> Following up on
> http://lists.wfu.edu/pipermail/gretl-devel/2016-August/006890.html
>
> The new --xdepend option to the modtest command -- which computes
> Pesaran's "CD" test for cross-sectional dependence in panel-data
models
> -- is now documented and available via the GUI (in git and snapshots).
>
That's great of course; in relation to the feature request
(https://sourceforge.net/p/gretl/feature-requests/97/) I'm not sure whether
Pesaran's statistic should remain the only one available, though.
I have to admit I'm not up to date on this particular topic, i.e. whether
there's a consensus view or diffferent preferences. Does anybody know more
about these tests?
So far I've just worked from Pesaran's 2004 paper and the Stata
Journal article on the "xtcsd" command from 2006. Stata supports a
couple of other tests; they seem a bit fiddly to me and I haven't
yet investigated how well motivated they are, but it may be worth
looking into them.
The new "modtest" option is open to extension, via the syntax
--xdepend=method, if we reckon that's worth doing. One potential
drawback with Pesaran's test is that since it's based on a sum of
correlation coefficients, it could be "fooled" into non-rejection in
case there are significant correlations both positive and negative
that more or less cancel out. To guard against this, to some degree,
the modtest output prints the average absolute correlation by
default.
Allin