Alejandro Delafuente --
I cannot speak to your data, but your manual two-step is invalid (see
Wooldridge http://www.stata.com/bookstore/cspd.html for more detail),
as you cannot include the fitted Xb as a regressor and get an IV
estimate. You can include the residual from the first step as an
additional regressor in many nonlinear models to get an IV estimate,
and the test that the coefficient on the residual is zero is a test of
endogeneity. See below code for a comparison of methods. How are you
calculating marginal effects? Carefully, I hope.
sysuse auto, clear
ivprobit for mpg (pr=wei), nolog
est sto ivprobit
ivprobit for mpg (pr=wei), twostep
est sto ivp2step
reg pr mpg wei
predict vhat, resid
predict xb
probit for mpg pr vhat, nolog
est sto man2step
ren pr waspr
ren xb price
probit for mpg pr, nolog
est stor wrong2step
est table *, eq(1)
On 12/6/06, Alejandro Delafuente
<[email protected]> wrote:
Hello statalisters, am trying to estimate an ivprobit through maximum
likelihood: ivprobit dependent_var a b c (endogenous_var= d e d*b e*c)
I have a group of instruments two of which stand alone and another two have
been interacted with some variables from my group of exogenous regressors.
Some of my marginal effects have an implausible magnitude, in particular the
instrumented variable (ie, 4.8). Could someone explain to me the reason for
this inflated effects?
I also tried to reestimate the above model in a two-step fashion:
ivprobit dependent_var a b c (endogenous_var= d e d*b e*c), twostep
Then again I obtained high magnitudes in the marginal effects of my
instrumented regressor.
I did the two-step estimation manually, first, regressing my endogenous var on
my set of instruments and the other exogenous regressors, then I predict the
values and fit them into the probit. This time the coefficient of my regressors
gave more plausible marginal effects. For instance, the marginal effect on my
instrumented variable is 0.44). Obviously the manual procedure and Stata are
not doing the same thing, but not sure why. I can only think on the lack of
adjusment of the variance-covariance matrix in the manual procedure, but that
should only affect the standard errors, isn't it?
P.S.: A last bit of information is that my endogenous variable was generated
through a regression with some of the coefficients that are also included in
the probit model. Could this explain part of the inflated marginal effects in
both of my ivprobit estimations.
Many thanks,
Alejandro
--
Alejandro de la Fuente
Department of International Development/QEH
University of Oxford, Mansfield Road, Oxford OX1 3TB
Tel: 01865 281836
*
* 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/
*
* 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/
