I see. Please take what I say with a pinch of salt, because I am far from
being an expert in this area.
Dear statatisters, feel free to correct me.
The first syntax assumes that the errors are jointly normally distributed.
You probably have in mind the picture of a bivariate normal; it looks pretty
much like a sand dune. It is the most intuitive way of thinking of a normal
in 3D all nicely symmertic. Just that in -treatreg- it is not nicely
symmertric because we are allowing for correlation.
If the assumption is true (and all the exclusion restrictions are true) then
this estimator is consistent, efficient, and asymptotically normal. Sounds
attractive, doesn't it? Unfortunately, if even one of the restrictions fails
then the properties crumble and you loose consistency for all parameters.
The second syntax is much more cautious in what it assumes. It does not
assume joint normality; it assumes normality of the error in the treatment
equation and a few other things. Which means that the two-step estimator is
consistent much more often than the first syntax (called full information ML
or FIML) one. The two-step is although inefficient if the errors are indeed
jointly normal.
What does it mean in practice? My understanding is that microeconometrics
has been moving away from the first syntax for a long time now, and that you
need really a good story to justify the assumption of joint normality if you
want people to buy the results from the first syntax.
According to your level of expertease and the importance of the project you
might want to read
Maddala (1983) Limited-Dependent and Qualitative Variables in Econometrics
and/or Jeffrey M. Wooldridge (2001) Econometric Analysis of Cross Section
and Panel Data
I much prefer Wooldriged, although Maddala is for various reasons the usual
reference.
Renzo
-----Original Message-----
*From: CHO,HYE-JEE [mailto:[email protected]]
Sent: Friday, November 28, 2003 2:15 PM
To: [email protected]; Renzo Comolli
Cc: [email protected]
Subject: Re: st: 'treatreg' questions
Hello again,
Thank you for your response, Renzo. Here is what I am trying to do: I am
trying to estimate two different types of treatment effects model. One is a
maximum likelihood model and the other is a two-step model.
The STATA command for the maximum likelihood model is:
.treatreg depvar1 [varlist1], treat [depvar2=varlist2]
(for treatreg, MLE is the default)
The command for the two-step model is:
.treatreg depvar1 [varlist1], treat [depvar2=varlist2] twostep
Thank you,
HyeJee
-----Original Message-----
> *From "CHO,HYE-JEE" <[email protected]>
> To [email protected]
> Subject st: 'treatreg' questions
> Date Fri, 28 Nov 2003 02:21:27 -0800
>
> Hello,
>
> I am using treatment effects model to assess the impact of IMF programs
> on
> fiscal balance. I use the command "treatreg" in STATA, and when I use a
> two-step estimator I get totally different results from a MLE estimator
> (i.e. the coefficient for the IMF programs is significant at 1% level
> when
> I use MLE, but is not significant in the two-step). Why do I get such
> different results and what would be the implication, if any, of such
> difference in light of endogeneity and selection bias?
>
> Also , I do not get a Wald chi-sqaure value when I use MLE model. What
> would be the problem?
>
> I would be very grateful to receive your comments.
>
> Thank you very much,
> HyeJee
*
* 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/
