st: Re: RE: Nonlinear panel data models and their efficient estimation

["Rodrigo A. Alfaro" <raalfaroa@gmail.com>]

The question is valid for a linear model, in that case it is fine to say: 
autocorrelation or heteroskedasticity with endogenous regressors. The 
algorithm used in -ivreg2- for dealing with the unspherical errors only 
affects the standard errors. In the case of the model that you are asking it 
is hard to think that this is a feasible solution. Adding sample selection 
will make the things even more complicated, are you thinking in a two-step 
model or maximum-likelihood?

My suggestion will be similar to Maarten's proposal. If you are doing this 
for a empirical paper, it is better to use -xtlogit- or -xtprobit- (both are 
random effects)... you will spend a lot of time checking that the model is 
right under different quadratures. Moreover, for a reasonable number of 
cross sectional units and a powerful computer you can estimate the logit and 
probit model adding dummies for the cross sectional units. In the case of 
logit, -clogit- could help you in this matter. There are corrections for the 
probit+fe model, but W. Greene considers that the raw-estimation is fine.

Now, if your live depends on this estimation and you have the time to spend 
on it you can write down the loglikelihood assciated with your problem. In 
that case, you define the endogeneity (linear or even not linear) of your 
covariate, the behaviro of the unobserved component (fixed or random 
effects) and also the sample selection rule. Finally, you could impose a 
reasonable structure for the "error" such as AR(1) with heteroskedasticity. 
All of these can be estimated by -ml- routines that you have to code on your 
own. With a reasonable load of 10-hours per day, 6 days a week you will not 
spend less than 6 months. For the case of fixed effects you need to improve 
the algorithm to make feasible the estimation, Greene proposes an algorithm 
to improve the estimationof MLE with fixed effects avoiding to add the 
dummies. Moreover, you need to fix your code to allow unbalanced panels (if 
you need it). You need to prove things like consistency and asymptotic 
normality to believe in your estimator... and probably those are different 
for large-n fixed-T and large-n large-T asymptotics. In conclusion, it is 
not simple... and this could be the main reason why there is not such a 
package!! A lazy and risky alternative is to search on the web if that 
medicine was written for someone else, I doubt that there is something even 
close that you are looking for. Maybe something smaller in features and 
available for Matlab, Gauss of Fortran.

Rodrigo.



----- Original Message ----- 
From: "Maarten Buis" <M.Buis@fsw.vu.nl>
To: <statalist@hsphsun2.harvard.edu>
Sent: Monday, June 11, 2007 9:45 AM
Subject: st: RE: Nonlinear panel data models and their efficient estimation



> So you want to estimate a fixed/random effects model with
> heteroscedasticity, serial correlation, and sample selection
> correction and endogenous variables? As you can imagine this
> is not going to be easy. My advise would be to first think
> very very very hard about whether you really need it. So
> think about the question you are trying to answer. Do this
> with somebody else who hasn't worked on your problem before.
> When somebody comes into my office with the question of how
> to ``bootstrap a multilevel sample selection MCMC propensity
> weighted semi parametric model'' the outcome has always been
> that there is a much simpler way of answering the research
> question (sometimes as easy as comparing 4 means, which
> economist tend to call a difference in difference estimator
> but I think terminology like that just hides the simplicity
> of procedures like that (but maybe that is intentional)).
>
> Hope this helps (and isn't too pretentious),
> Maarten
>
> -----------------------------------------
> Maarten L. Buis
> Department of Social Research Methodology
> Vrije Universiteit Amsterdam
> Boelelaan 1081
> 1081 HV Amsterdam
> The Netherlands
>
> visiting address:
> Buitenveldertselaan 3 (Metropolitan), room Z434
>
> +31 20 5986715
>
> http://home.fsw.vu.nl/m.buis/
> -----------------------------------------
>
> -----Original Message-----
> From: owner-statalist@hsphsun2.harvard.edu 
> [mailto:owner-statalist@hsphsun2.harvard.edu]On Behalf Of Annika Meng
> Sent: maandag 11 juni 2007 14:51
> To: statalist@hsphsun2.harvard.edu
> Subject: st: Nonlinear panel data models and their efficient estimation
>
> Dear Statalist users,
>
> I would like to estimate a binary panel data model with fixed and random 
> effects with "xtlogit" / "xtprobit". However, I found no information on 
> how to correct my estimated coefficients for heteroskedasticity and serial 
> correlation. In addition, I have to deal with endogenous regressors and 
> sample selection.
> The Stata command "ivreg2" is the command for these problems concerning 
> linear endogenous variables.
>
> Does anyone know if there is a nonlinear equivalent or another way to 
> correct the estimated standard errors?
>
> Thank you for your replies.
>
> Annika Meng
> *
> *   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/