Statalist The Stata Listserver


[Date Prev][Date Next][Thread Prev][Thread Next][Date index][Thread index]

st: choice of alpha in qvf family(nbinomial alpha)


From   <durham@gmx.de>
To   <statalist@hsphsun2.harvard.edu>
Subject   st: choice of alpha in qvf family(nbinomial alpha)
Date   Wed, 17 May 2006 00:14:06 -0400

Thanks, Mark, qvf works very well. However, I have one problem: If I
estimate models with both nbreg and qvf (without endogenous variables), the
results differ quite a bit. From what I've read about glm, qvf, and nbreg,
the reason is that nbreg estimates alpha (the overdispersion factor), while
alpha has to be specified in glm (and qvf) (Stata manual glm p. 398). The
logical choice seems to be to first estimate alpha_hat using nbreg and to
then specify qvf..., family(nbreg alpha_hat).
I tried that and estimated quite large alphas (4-6, depending on model
specification); if I input these in qvf, some models did not converge "after
101 iterations". If I lower the alpha parameter in qvf, the model converges,
but the results start to differ from those produced by nbreg :(

So here are my questions:

1. is it a good idea to estimate alpha using nbreg and then specify that
alpha in qvf (where I also add my instrument)?
2. what explains the lack of convergence in qvf with an alpha estimated from
nbreg - and what is the best way to address that problem?

Thanks so much for your help,
Henry

-----Original Message-----
From: owner-statalist@hsphsun2.harvard.edu
[mailto:owner-statalist@hsphsun2.harvard.edu] On Behalf Of Schaffer, Mark E
Sent: Tuesday, May 02, 2006 6:28 AM
To: statalist@hsphsun2.harvard.edu
Subject: st: RE: correct SEs with 2SLS and count models

Henry, 

> -----Original Message-----
> From: owner-statalist@hsphsun2.harvard.edu
> [mailto:owner-statalist@hsphsun2.harvard.edu] On Behalf Of 
> durham@gmx.de
> Sent: 02 May 2006 05:12
> To: statalist@hsphsun2.harvard.edu
> Subject: st: correct SEs with 2SLS and count models
> 
> Hi @ll,
> 
> I have an endogeneity problem (continuous variable X1) in a regression 
> model with count outcomes (50% zeros); ideally, I would like to use 
> NBREG or ZINB in stage 2. I could simply use IVREG to estimate both 
> equations simultaneously, but that doesn't work given my DV. 
> Alternatively, I could do a first-stage regression and then add the 
> predicted values X1_hat in my stage 2. However, doing this results in 
> incorrect SEs in stage 2 (they are conditional on the b_hat from stage 
> 1). So here are my questions:
> 
> 1. Is there a stata command or .ado out there that does ML estimation 
> of instrumental variables models with count outcomes (similar to IVREG 
> for continuous DVs)?

-qvf- will fit generalized linear models and supports IV as well as Poisson,
negative binomial, etc.  -findit qvf- will take you there.

Cheers,
Mark
 
> or
> 
> 2. Is there a convenient way to compute the correct standard errors in 
> a traditional 2SLS estimation? I am thinking of bootstrapping of both 
> equations to get the empirical distribution of SEs, but I don't know 
> how to implement that in STATA :(.
> 
> Thanks so much!
> 
> Henry
> 
> *
> *   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/



© Copyright 1996–2014 StataCorp LP   |   Terms of use   |   Privacy   |   Contact us   |   What's new   |   Site index