[Date Prev][Date Next][Thread Prev][Thread Next][Date index][Thread index]
st: Re: ivreg2 use
On Wednesday, August 6, 2003, at 02:33 AM, Busakom wrote:
Almost correct. You should disabuse yourself of the notion that "this
is an instrument for that"; as frequently mentioned on this list and in
the cited Stata FAQ, instrumental variables estimators do not work that
way. All of the Z variables are instruments for all of the included Y
variables. That said, what about your proposed regression?
I have two simultaneous equations:
(1) y1 = a0*y2 + a1*x1 + a2*x2 + a3*x3 + e
(2) y2 = b0*y1 + b1*z1 + b2*z2 + b3*z3 + u
The variables are first-differences. As I understand,
I can perform 2SLS by using ivreg2:
ivreg2 y1 x1 x2 x3 (y2 = z1 z2 z3)
My problem is slightly more complicated than this.
What if, in each equation, there is a lag dependent
(3) y1 = a0*y2 + a1*x1 + a2*x2 + a3*x3 + a4*Ly1 + e
(4) y2 = b0*y1 + b1*z1 + b2*z2 + b3*z3 + b4*Ly2 + u
Here, Ly1 is correlated with e, and Ly2 is correlated
with u. I plan to instrument Ly1 with L2y1
(two-period lag of y1) and Ly2 with L2y2 (two-period
lag of y2).
I believe I can still use ivreg2 to deal with this
problem. My question is whether the following command
syntax is correct?
ivreg2 y1 x1 x2 x3 (y2 Ly1 Ly2 = z1 z2 z3 L2y1 L2y2)
And is -ivreg2, gmm- is the same as -ivgmm0-?
Your first equation contains regressors y2, x1-x3, and Ly1. It does not
contain Ly2, so Ly2 should not be present in the included endogenous
list to the left of the equals sign, or it will appear as a regressor
in the equation, which presumably it should not. The instrument list is
correct, if your assertion that the second lags are orthogonal to e and
u is warranted. This can be tested with a diff-Sargan test (ivreg2
orthog option). Note that ivreg2 will report that the instruments
include x1, x2, x3 plus those you have specified (and constant) -- and
that is correct. The included exogeneous variables are indeed
instruments as well.
ivreg2, gmm is essentially the same estimator as the earlier ivgmm0. A
version of ivreg2 is currently under development which will provide
'ivgmmN' -- that is, allowing for arbitrary heteroskedasticity and
serial correlation in the error process. ivgmm0 (and its implementation
within ivreg2) only allows for the former.
* For searches and help try: