"I asked a question previously regarding the speed of "xtabond2"
command in
Stata (please see below), since it's unbearably slow to play with large
datasets. However, I got the reply from its author saying there's
little to
be done about it given the way this programme was written. In
particular,
"xtabond2" was not written in compiled C++ as opposed to "xtabond". I
am no
good in programming and am wondering whether anybody has written
his/her own
version in C++. Or could anyone kindly tell me whether there's any
other way
to speed up the computation or to obtain system GMM estimators in
Stata? "
The assertion (be it his or David Roodman's) that xtabond is written in
C/C++ is incorrect:
. which xtabond
/Applications/Stata/ado/updates/x/xtabond.ado
*! version 3.5.9 20sep2004
It is an ado-file, like almost all of the 'xt' routines (and over 80%
of the commands in Stata 8).
The major slowdown in xtabond2 is the "Windmeijer correction" that
allows consistent standard errors to be computed for two-step
estimates. If you don't need to compute them while analyzing a number
of specifications, don't, and the routine will run much faster. The
other thing that may be increasing runtime dramatically is letting the
A-B estimator come up with hundreds of instruments (if T is
nontrivial). You can avoid this (which does little good anyway from an
econometric standpoint) by using the options that limit how far back
the lagged instruments go.
It is possible that the routine could be coded more efficiently; it is
a volunteer effort by a smart guy, but he (like me, or any member of
the Stata user community) doesn't work at StataCorp LP and doesn't do
this for a living. It stands to reason that those who do are more
skilled at squeezing the inefficiencies out of code.
Personally, I have lobbied for a long time for StataCorp to take a hard
look at xtabond2, which IMHO has a more flexible and usable syntax than
official xtabond, and implements the system GMM estimator and the
Windmeijer correction (for instance, Greene's econometrics text has an
example of A-B which is impossible to estimate via xtabond, but can be
done with xtabond2. Otherwise official Stata compares poorly with
GAUSS, Ox, and other packages in which the full set of A-B tools are
available.
