Dear all,
since I am also trying to do a two-step estimation using a multinomial logit in the first step, I had a look at the paper suggested by Rafa. Please correct me if I am wrong but what I understood was that what should be avoided in such a case is the approach of Lee(1983) except if we have small samples.
Instead the suggest the Dubin and Mc Fadden (1984) model or even beter their modification to it that does not use the strong assumptions made by Lee. In this case, the authors suggest that the multinomial logit performs very well even when the IIA assumption is violated.
They also have a stata command that can do such an estimation (I haven't tried it yet), the selmlog.
I would like to address a more difficult question to the list: How can I apply correction for endogeneity (with the endogenous variable having more than 2 categories) in a panel context? So, would it make sense to run an mlogit as the first step and then an xtreg fixed effects as the second step?
Best,
Dimitris Pavlopoulos
-----Original Message-----
From: "R.E. De Hoyos" <redeho@hotmail.com>
To: <statalist@hsphsun2.harvard.edu>
Date: Mon, 10 Apr 2006 17:55:33 +0100
Subject: Re: st: RE: Econometrically sound to use Mills ratio after mprobit?
Stephen,
You are assuming that the selection problem in a multinomial context can be
accounted for by the same technique as in the binary problem (Heckman 1979).
Using the multinomial logit as the first-step estimation, Bourguignon et al.
(2004) have shown that this is not the right way to approach the problem.
Generally speaking, the selection problem in a multinomial context can be
define as:
y1 = xb + u1
y^*_m = zl + u_m, m = 1...M (outcomes)
Where y^* is a latent function and E(u1 | x,z)=0. Define p1...pM as the
conditional probabilities of observing each of the M outcomes. Then, the
selectivity-adjusted y1 can be estimated as:
y1 = xb + mu(p1...pM) + e1
You would need to take into account the conditional probabilities of
observing NOT ONLY outcome (1) but all other outcomes as well. The problem
is how to parameterize function mu(.)?
Rafa
PS. I have a working paper version of the reference, if you are interested I
can send it to you off-list.
Reference:
F. Bourguignon, M. Fournier & M. Gurgand, «Selection bias corrections based
on the multinomial logit model : Monte-Carlo comparisons», Journal of
Economic Surveys, forthcoming
----- Original Message -----
From: "Stephen Johnston" <sjohns21@umbc.edu>
To: <statalist@hsphsun2.harvard.edu>
Sent: Monday, April 10, 2006 5:01 PM
Subject: Re: st: RE: Econometrically sound to use Mills ratio after mprobit?
> Hello Miet,
>
> Thanks for your advice. Here is how I understand the procedure to work,
> I read this on an archived statalist post and I have tested it.
>
> mprobit y x1 x2 x3
>
> predict phat if e(sample), xb outcome (1)
>
> capture drop phat
> capture drop mills
>
> gen mills = normden(phat)/norm(phat)
>
> reg y2 x1 x2 mills
>
> For the "outcome" command in the predict line, you have to specify which
> of the choice outcomes (in your dependent variable) you are predicting a
> probability for. In this case the probability predicted for the choice
> that is set equal to 1. You can then generate an IMR for each choice in
> y.
>
> You can check that this works by generating the inverse mills ratio from
> a probit and then including it in an OLS equation - then use the twostep
> command for the heckman procedure to make sure the results match. For
> this you will not need to specify an outcome since it will be generated
> from a probit. I hope this helps. Let me know if you have any trouble.
>
> Thanks,
> Stephen
>
>
> On Apr 10, 2006, at 4:35 AM, Maertens, Miet wrote:
>
>> Dear Stephen,
>>
>> Maybe the following two articles by Wooldridge and by Lechner can help
>> you further:
>> http://www.msu.edu/~ec/faculty/wooldridge/current%20research/ ape1r5.pdf
>> http://ideas.repec.org/p/iza/izadps/dp91.html
>>
>> I'm also trying to perform a similar estimation with stata but I'm
>> struggling with calculating the Inverse Mills ratio's. Could you let me
>> know how exactly you are implementing the procedure?
>>
>> Thanks,
>> Miet
>>
>> -----Original Message-----
>> From: owner-statalist@hsphsun2.harvard.edu
>> [mailto:owner-statalist@hsphsun2.harvard.edu] On Behalf Of Stephen
>> Johnston
>> Sent: 07 April 2006 18:35
>> To: statalist@hsphsun2.harvard.edu
>> Subject: st: Econometrically sound to use Mills ratio after mprobit?
>>
>> Hello,
>>
>> I am estimating a multinomial probit for a selection equation with 3
>> choices and I am interested in using the inverse mills ratio
>> generated from the MNP in a second step equation. I know how to
>> implement this procedure, however, I have not been able to find any
>> literature that proves that the Heckman two-step estimation procedure
>> can appropriately and directly extend from a probit selection
>> equation to a multinomial probit selection equation. Does anyone
>> know of any papers that address this issue?
>>
>> Thanks,
>> Stephen
>> *
>> * 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/
>>
>> Disclaimer: http://www.kuleuven.be/cwis/email_disclaimer.htm
>>
>>
>> *
>> * 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/
>
*
* 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/
****************************************
Dimitris Pavlopoulos
PhD student
Tilburg University
Faculty of Social Sciences
Warandelaan 2, Postbus 90153
5037 AB, TILBURG
THE NETHERLANDS
Tel. ++31 13 466 3001/ Room S-173
email: D.Pavlopoulos@uvt.nl
***************************************
*
* 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/