When you do a two-stage least squares by hand, estimating first one equation, and then using the predicted dependent variables as predictors in the second equation, you are going to get inflated t tests. The reason is that when you inserted the predicted values in the second equation, the computer program (or formula if you are doing the computation by hand) doesn't know that these are estimates, not observed values. It will compute the standard errors and t-tests in a way that does not take into account the existence of error in those estimates, and so it will exaggerate the precision with which the parameters are being estimated in the second stage. David Greenberg, Sociology Department, New York University
----- Original Message -----
From: daniel cassar <[email protected]>
Date: Thursday, April 5, 2007 2:24 pm
Subject: Re: st: 2sls standard errors.
To: [email protected]
> Dear Austin,
>
> you're right. The numbers that were completely
> inflated were t-statistics, not SEs. Really sorry
> about that.
> Using the data set of your example, what I was trying
> to "by hand", was:
> reg3 (iq knn tenure s med)(med lw s expr), 2sls
>
> Best regards & thanks,
> DC
> --- Austin Nichols <[email protected]> escribi�:
>
> > daniel cassar --
> > I would prefer -ivreg2- (from SSC) for such a model.
> > It is impossible to tell from your post whether you
> > have specified
> > your model correctly, and it cannot be the case that
> > "the standard
> > errors were completely inflated (they grew very
> > noticeably) so that
> > almost all of the explanatory variables were
> > significant" since bigger
> > SEs imply less significance. Perhaps you can
> > specify your question in
> > terms of a dataset that ships with Stata, or one
> > available to everyone
> > via the web, e.g.
> >
> > ssc inst ivreg2, replace
> > ssc inst estout, replace
> > use
> >
> http://fmwww.bc.edu/ec-p/data/hayashi/griliches76.dta,
> > clear
> > qui reg3 (lw=s expr iq) (iq=s expr med kww age)
> > est store reg3
> > qui ivreg2 lw s expr (iq=med kww age)
> > est store ivreg
> > qui reg iq med kww age s expr
> > ren iq was_iq
> > predict iq
> > qui reg lw s expr iq
> > est store byhand
> > esta reg3 ivreg byhand, eq(1) se mti(reg3 ivreg
> > byhand)
> >
> > On 4/5/07, daniel cassar <[email protected]>
> > wrote:
> > > Hi,
> > > I have a two-equation system that I solved with
> > > -reg3-. However, when I did the same system "by
> > hand",
> > > the standard errors were completely inflated (they
> > > grew very noticeably) so that almost all of the
> > > explanatory variables were significant. the point
> > > estimates were the same, though.
> > >
> > > Any ideas as to why this may happen? Which of the
> > two
> > > ways has the most reliable standard errors?
> > >
> > > Note: my system has the following form:
> > > x1=a1 + by1 + by2 + bx2 + e
> > > x2=a2 + bz1 + bz2 + bz3 + e
> > >
> > > And I got the weird results only in the second
> > > equation 2. (Note that the dependent variable of
> > the
> > > second equation enters the first equation, but the
> > > dependent variable of the first eq. does not enter
> > the
> > > second eq.)
> > >
> >
>
>
> __________________________________________________
> Correo Yahoo!
> Espacio para todos tus mensajes, antivirus y antispam �gratis!
> Reg�strate ya - http://correo.espanol.yahoo.com/
> *
> * 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/
