I'm estimating a dynamic panel data model via gmm and am trying to test
to
validity of a "subset" of instruments using a C (difference-in-Sargans)
test. The problem of getting negative test statistics when calculating
by
hand has been discussed before on this list. However, Baum, Schaffer &
Stillman (2003; Stata Journal) discuss how using the covariance matrix
of
the moment conditions from the restricted model (i.e., that model
omitting
the subset of instruments) can overcome the negative test statistic
model.
I know that ivreg2's orthog option will conduct this test. However,
when
trying to convert my Arellano-Bond estimation
in an xtivreg2, the results of the C tests produced via the orthog
option
are very differnt from those that I calculated by hand after xtabond2
(the
tests that where insig after xtabond2 become significant via the orthog
option). I assume this means that due to my poor programming skills, I
have
incorrectly used ivreg2 to estimate dynamic models per Arellano-Bond
(perhaps I have not effectively programmed first differences via
ivreg2).
Does anyone have any suggestions about how to produce the "adjusted" C
test
using xtabond2?
Since the C test is just a difference in Sargan, calculate the two
xtabond2 models that you want to compare, save their Sargan stats and
form the appropriate Chi-square. It's just like doing a likelihood
ratio test 'by hand'--you just need the stat for each model and then
take their difference. The number of df of the C stat will be the
difference in the df of the respective Sargans, as you can see from an
example in ivreg2.
I doubt that you can manually create the instruments used by A-B for
use with ivreg2 (or any other IV routine). They are quite complicated
to construct.
