Thanks to Kit Baum, the new package -ivpois- is available from SSC;
type -ssc install ivpois- to get it.
ivpois implements a Generalized Method of Moments (GMM) estimator
of Poisson regression and allows endogenous variables to be
instrumented by excluded instruments, hence the acronym for
Instrumental Variables (IV) in its name. Standard errors are
estimated by bootstrapping. Poisson regression assumes
E[y|X]=exp(Xb) to get a consistent estimate of b, so it is
appropriate for a wide variety of models where the dependent
variable is nonnegative (zero or positive), not just where the
dependent variable measures counts of events. The central code
was promulgated by Bill Gould at a Seminar in DC on November 2,
2007.
The code was written at some earlier date and demonstrated at the
November 2 seminar by David Drukker, so the package is just a
repackaging of others' work, but I thought it was worth advertising
what I had not known until that November 2 seminar, namely that a GMM
estimator is nearly trivial to write using Mata's optimize command.
I had not used Mata at all before the November 2 seminar, but I am now
a convert. The guts of the command are just to specify a moment
condition:
m=((1/rows(Z)):*Z'((y:*exp(-X*b') :- 1)))'
and a criterion function to be minimized, which just weights the cross
product of that moment condition to get an efficient estimator:
crit=(m*W*m')
where W=rows(Z)*cholinv(Z'Z). These are wrapped in a Mata function:
mata:
void iv_pois(todo,b,crit,g,H)
{
external y,X,Z,W
m=((1/rows(Z)):*Z'((y:*exp(-X*b') :- 1)))'
crit=(m*W*m')
}
and then minimizing the criterion function is as simple as specifying
the variables in y,X,Z and:
init=J(1,cols(X),0)
S=optimize_init()
optimize_init_evaluator(S, &iv_pois())
optimize_init_which(S,"min")
optimize_init_evaluatortype(S,"d0")
optimize_init_params(S,init)
p=optimize(S)
p
where the first line defines an initial parameter vector of zeros, the
second a name for our optimization problem, and the third the function
that will be optimized in that problem. The fourth line specifies
whether we seek a min or max, the sixth specifies our initial
parameter vector, and the seventh conducts the optimization and
assigns the resulting vector to p. The fifth line is the interesting
one, and there is some information at
type Description
"d0" function() returns scalar value
"d1" same as "d0" and returns gradient rowvector
"d2" same as "d1" and returns Hessian matrix
"v0" function() returns colvector value
"v1" same as "v0" and returns score matrix
"v2" same as "v1" and returns Hessian matrix
In short, it is ridiculously easy to program a new GMM estimator in
Stata now (Stata 10 is required for the -optimize- command in Mata).
If you haven't upgraded to Stata 10 yet, now is the time.
I fully expect a host of GMM estimators to flower on SSC in the near
future. Presumably, a *generic* GMM estimator is just a matter of
specifying a convenient syntax by which users may supply moment
conditions on the fly.
*
* For searches and help try:
* http://www.stata.com/support/faqs/res/findit.html
* http://www.stata.com/support/statalist/faq
* http://www.ats.ucla.edu/stat/stata/
