Bookmark and Share

Notice: On April 23, 2014, Statalist moved from an email list to a forum, based at

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: st: Understanding how to use biprobit to control for endogenous dummy

From   Paula Herrera <[email protected]>
To   [email protected]
Subject   Re: st: Understanding how to use biprobit to control for endogenous dummy
Date   Fri, 30 Mar 2012 16:24:23 +0200

I have also estimated  a model with a binary outcome and a binary
endogenous variable, with a biprobit and with linear instrumental
variables  (using ivreg2). From the output of the ivreg2 I can check
that my instrument works pretty well, it pass all the tests.

To compare the two models ones needs to calculates the marginal
effects or the average marginal effects, since the coefficient of the
biprobit is only informative about the sign.To obtain the average
treatment effects I use the function: margins, dydx predict (pmarg1)
force, after I run the biprobit.

However, the biprobit average marginal effect differ significantly
from the beta estimated with the linear model, which I understand is
the local average treatment effect. Is this a normal result? or is an
indicative that something may be wrong?



El día 30 de marzo de 2012 15:51, Austin Nichols
<[email protected]> escribió:
> Kevin Infante <[email protected]>:
> See
> and always estimate the linear model too,
> with e.g. -ivreg2- on SSC,
> to get a sense of instrument strength etc.
> The linear model is much easier to interpret, most of the time.
> On Fri, Mar 30, 2012 at 12:13 AM, Kevin Infante <[email protected]> wrote:
>> Hello statalist,
>> I'm an undergraduate senior Economics major. I have a basic
>> understanding of econometrics
>> and a basic understanding of Stata. Unfortunately, the models I
>> apparently need to run demand a more in-depth understanding of both.
>> Here is one of my regressions:
>> y1=a*y2+b*w+u1   (1)
>> y2=c*z+d*w+u2     (2)
>> y1=dummy representing whether a respondent voted
>> y2= endogenous dummy representing whether a respondent lists Internet news as
>> their main news source
>> w=collection of demographic controls, as well as dummies for the year
>> and state of residence of the respondent
>> z= a collection of instrument variables
>> u1=error term of equation (1)
>> u2= error term of equation (2)
>>  From all I have read on this list, -biprobit- is the stata command
>> that is best suited to deal with an endogenous dummy variable in a
>> discrete choice model.
>>  I've run my models using biprobit, and successfully gotten output.
>> The first problem I have is I don't know how to interpret the output.
>> I gather that rho is important, but I don't know what rho is or what
>> it measures. The coefficient on rho is -.55, and the SE is .055. I'm
>> assuming the likelihood-ratio test of rho=0 is also important, which
>> indicates p>chi2=0.0000. I again have no idea what that means. Lastly,
>> I'm not sure what coefficients on the variables I should be looking
>> at.
>> From reading different message threads from the archives, I gather I
>> may need to impose additional restrictions, specifically a coherency
>> restriction and an exclusions restrictions. Unfortunately, I again
>> have no idea what those are. I have checked out both Greene(2003) and
>> Maddala(1983), and read the sections that relate to this analysis,
>> butI understand very little of it, as the econometrics is much more
>> advanced than what I've been exposed to.
>> If anyone could help explain to me relatively simply how to interpret
>> the biprobit output, whether I need additional restrictions and why,
>> and just some of the basic theory behind what I am trying to do, I
>> would much appreciate it. Also if anyone has any references or sources
>> that would help me learn about what I'm doing, that would be very
>> useful. My faculty adviser and the resident Stata expert were both not
>> able to help me much as both were unfamiliar with the theory and stata
>> command. Thank you for all your help!
>> Best,
>> Kevin Infante
> *
> *   For searches and help try:
> *
> *
> *

*   For searches and help try:

© Copyright 1996–2018 StataCorp LLC   |   Terms of use   |   Privacy   |   Contact us   |   Site index